initial upload

ImageBind/CODE_OF_CONDUCT.md

@@ -0,0 +1,80 @@
+# Code of Conduct
+
+## Our Pledge
+
+In the interest of fostering an open and welcoming environment, we as
+contributors and maintainers pledge to make participation in our project and
+our community a harassment-free experience for everyone, regardless of age, body
+size, disability, ethnicity, sex characteristics, gender identity and expression,
+level of experience, education, socio-economic status, nationality, personal
+appearance, race, religion, or sexual identity and orientation.
+
+## Our Standards
+
+Examples of behavior that contributes to creating a positive environment
+include:
+
+* Using welcoming and inclusive language
+* Being respectful of differing viewpoints and experiences
+* Gracefully accepting constructive criticism
+* Focusing on what is best for the community
+* Showing empathy towards other community members
+
+Examples of unacceptable behavior by participants include:
+
+* The use of sexualized language or imagery and unwelcome sexual attention or
+advances
+* Trolling, insulting/derogatory comments, and personal or political attacks
+* Public or private harassment
+* Publishing others' private information, such as a physical or electronic
+address, without explicit permission
+* Other conduct which could reasonably be considered inappropriate in a
+professional setting
+
+## Our Responsibilities
+
+Project maintainers are responsible for clarifying the standards of acceptable
+behavior and are expected to take appropriate and fair corrective action in
+response to any instances of unacceptable behavior.
+
+Project maintainers have the right and responsibility to remove, edit, or
+reject comments, commits, code, wiki edits, issues, and other contributions
+that are not aligned to this Code of Conduct, or to ban temporarily or
+permanently any contributor for other behaviors that they deem inappropriate,
+threatening, offensive, or harmful.
+
+## Scope
+
+This Code of Conduct applies within all project spaces, and it also applies when
+an individual is representing the project or its community in public spaces.
+Examples of representing a project or community include using an official
+project e-mail address, posting via an official social media account, or acting
+as an appointed representative at an online or offline event. Representation of
+a project may be further defined and clarified by project maintainers.
+
+This Code of Conduct also applies outside the project spaces when there is a
+reasonable belief that an individual's behavior may have a negative impact on
+the project or its community.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported by contacting the project team at <opensource-conduct@fb.com>. All
+complaints will be reviewed and investigated and will result in a response that
+is deemed necessary and appropriate to the circumstances. The project team is
+obligated to maintain confidentiality with regard to the reporter of an incident.
+Further details of specific enforcement policies may be posted separately.
+
+Project maintainers who do not follow or enforce the Code of Conduct in good
+faith may face temporary or permanent repercussions as determined by other
+members of the project's leadership.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage], version 1.4,
+available at https://www.contributor-covenant.org/version/1/4/code-of-conduct.html
+
+[homepage]: https://www.contributor-covenant.org
+
+For answers to common questions about this code of conduct, see
+https://www.contributor-covenant.org/faq

ImageBind/CONTRIBUTING.md

@@ -0,0 +1,31 @@
+# Contributing to ImageBind
+We want to make contributing to this project as easy and transparent as
+possible.
+
+## Pull Requests
+We actively welcome your pull requests.
+
+1. Fork the repo and create your branch from `main`.
+2. If you've added code that should be tested, add tests.
+3. If you've changed APIs, update the documentation.
+4. Ensure the test suite passes.
+5. Make sure your code lints.
+6. If you haven't already, complete the Contributor License Agreement ("CLA").
+
+## Contributor License Agreement ("CLA")
+In order to accept your pull request, we need you to submit a CLA. You only need
+to do this once to work on any of Meta's open source projects.
+
+Complete your CLA here: <https://code.facebook.com/cla>
+
+## Issues
+We use GitHub issues to track public bugs. Please ensure your description is
+clear and has sufficient instructions to be able to reproduce the issue.
+
+Meta has a [bounty program](https://www.facebook.com/whitehat/) for the safe
+disclosure of security bugs. In those cases, please go through the process
+outlined on that page and do not file a public issue.
+
+## License
+By contributing to Omnivore, you agree that your contributions will be licensed
+under the [LICENSE](LICENSE) file in the root directory of this source tree.

ImageBind/LICENSE

@@ -0,0 +1,437 @@
+Attribution-NonCommercial-ShareAlike 4.0 International
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+     Considerations for licensors: Our public licenses are
+     intended for use by those authorized to give the public
+     permission to use material in ways otherwise restricted by
+     copyright and certain other rights. Our licenses are
+     irrevocable. Licensors should read and understand the terms
+     and conditions of the license they choose before applying it.
+     Licensors should also secure all rights necessary before
+     applying our licenses so that the public can reuse the
+     material as expected. Licensors should clearly mark any
+     material not subject to the license. This includes other CC-
+     licensed material, or material used under an exception or
+     limitation to copyright. More considerations for licensors:
+    wiki.creativecommons.org/Considerations_for_licensors
+
+     Considerations for the public: By using one of our public
+     licenses, a licensor grants the public permission to use the
+     licensed material under specified terms and conditions. If
+     the licensor's permission is not necessary for any reason--for
+     example, because of any applicable exception or limitation to
+     copyright--then that use is not regulated by the license. Our
+     licenses grant only permissions under copyright and certain
+     other rights that a licensor has authority to grant. Use of
+     the licensed material may still be restricted for other
+     reasons, including because others have copyright or other
+     rights in the material. A licensor may make special requests,
+     such as asking that all changes be marked or described.
+     Although not required by our licenses, you are encouraged to
+     respect those requests where reasonable. More considerations
+     for the public:
+    wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International
+Public License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution-NonCommercial-ShareAlike 4.0 International Public License
+("Public License"). To the extent this Public License may be
+interpreted as a contract, You are granted the Licensed Rights in
+consideration of Your acceptance of these terms and conditions, and the
+Licensor grants You such rights in consideration of benefits the
+Licensor receives from making the Licensed Material available under
+these terms and conditions.
+
+
+Section 1 -- Definitions.
+
+  a. Adapted Material means material subject to Copyright and Similar
+     Rights that is derived from or based upon the Licensed Material
+     and in which the Licensed Material is translated, altered,
+     arranged, transformed, or otherwise modified in a manner requiring
+     permission under the Copyright and Similar Rights held by the
+     Licensor. For purposes of this Public License, where the Licensed
+     Material is a musical work, performance, or sound recording,
+     Adapted Material is always produced where the Licensed Material is
+     synched in timed relation with a moving image.
+
+  b. Adapter's License means the license You apply to Your Copyright
+     and Similar Rights in Your contributions to Adapted Material in
+     accordance with the terms and conditions of this Public License.
+
+  c. BY-NC-SA Compatible License means a license listed at
+     creativecommons.org/compatiblelicenses, approved by Creative
+     Commons as essentially the equivalent of this Public License.
+
+  d. Copyright and Similar Rights means copyright and/or similar rights
+     closely related to copyright including, without limitation,
+     performance, broadcast, sound recording, and Sui Generis Database
+     Rights, without regard to how the rights are labeled or
+     categorized. For purposes of this Public License, the rights
+     specified in Section 2(b)(1)-(2) are not Copyright and Similar
+     Rights.
+
+  e. Effective Technological Measures means those measures that, in the
+     absence of proper authority, may not be circumvented under laws
+     fulfilling obligations under Article 11 of the WIPO Copyright
+     Treaty adopted on December 20, 1996, and/or similar international
+     agreements.
+
+  f. Exceptions and Limitations means fair use, fair dealing, and/or
+     any other exception or limitation to Copyright and Similar Rights
+     that applies to Your use of the Licensed Material.
+
+  g. License Elements means the license attributes listed in the name
+     of a Creative Commons Public License. The License Elements of this
+     Public License are Attribution, NonCommercial, and ShareAlike.
+
+  h. Licensed Material means the artistic or literary work, database,
+     or other material to which the Licensor applied this Public
+     License.
+
+  i. Licensed Rights means the rights granted to You subject to the
+     terms and conditions of this Public License, which are limited to
+     all Copyright and Similar Rights that apply to Your use of the
+     Licensed Material and that the Licensor has authority to license.
+
+  j. Licensor means the individual(s) or entity(ies) granting rights
+     under this Public License.
+
+  k. NonCommercial means not primarily intended for or directed towards
+     commercial advantage or monetary compensation. For purposes of
+     this Public License, the exchange of the Licensed Material for
+     other material subject to Copyright and Similar Rights by digital
+     file-sharing or similar means is NonCommercial provided there is
+     no payment of monetary compensation in connection with the
+     exchange.
+
+  l. Share means to provide material to the public by any means or
+     process that requires permission under the Licensed Rights, such
+     as reproduction, public display, public performance, distribution,
+     dissemination, communication, or importation, and to make material
+     available to the public including in ways that members of the
+     public may access the material from a place and at a time
+     individually chosen by them.
+
+  m. Sui Generis Database Rights means rights other than copyright
+     resulting from Directive 96/9/EC of the European Parliament and of
+     the Council of 11 March 1996 on the legal protection of databases,
+     as amended and/or succeeded, as well as other essentially
+     equivalent rights anywhere in the world.
+
+  n. You means the individual or entity exercising the Licensed Rights
+     under this Public License. Your has a corresponding meaning.
+
+
+Section 2 -- Scope.
+
+  a. License grant.
+
+       1. Subject to the terms and conditions of this Public License,
+          the Licensor hereby grants You a worldwide, royalty-free,
+          non-sublicensable, non-exclusive, irrevocable license to
+          exercise the Licensed Rights in the Licensed Material to:
+
+            a. reproduce and Share the Licensed Material, in whole or
+               in part, for NonCommercial purposes only; and
+
+            b. produce, reproduce, and Share Adapted Material for
+               NonCommercial purposes only.
+
+       2. Exceptions and Limitations. For the avoidance of doubt, where
+          Exceptions and Limitations apply to Your use, this Public
+          License does not apply, and You do not need to comply with
+          its terms and conditions.
+
+       3. Term. The term of this Public License is specified in Section
+          6(a).
+
+       4. Media and formats; technical modifications allowed. The
+          Licensor authorizes You to exercise the Licensed Rights in
+          all media and formats whether now known or hereafter created,
+          and to make technical modifications necessary to do so. The
+          Licensor waives and/or agrees not to assert any right or
+          authority to forbid You from making technical modifications
+          necessary to exercise the Licensed Rights, including
+          technical modifications necessary to circumvent Effective
+          Technological Measures. For purposes of this Public License,
+          simply making modifications authorized by this Section 2(a)
+          (4) never produces Adapted Material.
+
+       5. Downstream recipients.
+
+            a. Offer from the Licensor -- Licensed Material. Every
+               recipient of the Licensed Material automatically
+               receives an offer from the Licensor to exercise the
+               Licensed Rights under the terms and conditions of this
+               Public License.
+
+            b. Additional offer from the Licensor -- Adapted Material.
+               Every recipient of Adapted Material from You
+               automatically receives an offer from the Licensor to
+               exercise the Licensed Rights in the Adapted Material
+               under the conditions of the Adapter's License You apply.
+
+            c. No downstream restrictions. You may not offer or impose
+               any additional or different terms or conditions on, or
+               apply any Effective Technological Measures to, the
+               Licensed Material if doing so restricts exercise of the
+               Licensed Rights by any recipient of the Licensed
+               Material.
+
+       6. No endorsement. Nothing in this Public License constitutes or
+          may be construed as permission to assert or imply that You
+          are, or that Your use of the Licensed Material is, connected
+          with, or sponsored, endorsed, or granted official status by,
+          the Licensor or others designated to receive attribution as
+          provided in Section 3(a)(1)(A)(i).
+
+  b. Other rights.
+
+       1. Moral rights, such as the right of integrity, are not
+          licensed under this Public License, nor are publicity,
+          privacy, and/or other similar personality rights; however, to
+          the extent possible, the Licensor waives and/or agrees not to
+          assert any such rights held by the Licensor to the limited
+          extent necessary to allow You to exercise the Licensed
+          Rights, but not otherwise.
+
+       2. Patent and trademark rights are not licensed under this
+          Public License.
+
+       3. To the extent possible, the Licensor waives any right to
+          collect royalties from You for the exercise of the Licensed
+          Rights, whether directly or through a collecting society
+          under any voluntary or waivable statutory or compulsory
+          licensing scheme. In all other cases the Licensor expressly
+          reserves any right to collect such royalties, including when
+          the Licensed Material is used other than for NonCommercial
+          purposes.
+
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+  a. Attribution.
+
+       1. If You Share the Licensed Material (including in modified
+          form), You must:
+
+            a. retain the following if it is supplied by the Licensor
+               with the Licensed Material:
+
+                 i. identification of the creator(s) of the Licensed
+                    Material and any others designated to receive
+                    attribution, in any reasonable manner requested by
+                    the Licensor (including by pseudonym if
+                    designated);
+
+                ii. a copyright notice;
+
+               iii. a notice that refers to this Public License;
+
+                iv. a notice that refers to the disclaimer of
+                    warranties;
+
+                 v. a URI or hyperlink to the Licensed Material to the
+                    extent reasonably practicable;
+
+            b. indicate if You modified the Licensed Material and
+               retain an indication of any previous modifications; and
+
+            c. indicate the Licensed Material is licensed under this
+               Public License, and include the text of, or the URI or
+               hyperlink to, this Public License.
+
+       2. You may satisfy the conditions in Section 3(a)(1) in any
+          reasonable manner based on the medium, means, and context in
+          which You Share the Licensed Material. For example, it may be
+          reasonable to satisfy the conditions by providing a URI or
+          hyperlink to a resource that includes the required
+          information.
+       3. If requested by the Licensor, You must remove any of the
+          information required by Section 3(a)(1)(A) to the extent
+          reasonably practicable.
+
+  b. ShareAlike.
+
+     In addition to the conditions in Section 3(a), if You Share
+     Adapted Material You produce, the following conditions also apply.
+
+       1. The Adapter's License You apply must be a Creative Commons
+          license with the same License Elements, this version or
+          later, or a BY-NC-SA Compatible License.
+
+       2. You must include the text of, or the URI or hyperlink to, the
+          Adapter's License You apply. You may satisfy this condition
+          in any reasonable manner based on the medium, means, and
+          context in which You Share Adapted Material.
+
+       3. You may not offer or impose any additional or different terms
+          or conditions on, or apply any Effective Technological
+          Measures to, Adapted Material that restrict exercise of the
+          rights granted under the Adapter's License You apply.
+
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+  a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+     to extract, reuse, reproduce, and Share all or a substantial
+     portion of the contents of the database for NonCommercial purposes
+     only;
+
+  b. if You include all or a substantial portion of the database
+     contents in a database in which You have Sui Generis Database
+     Rights, then the database in which You have Sui Generis Database
+     Rights (but not its individual contents) is Adapted Material,
+     including for purposes of Section 3(b); and
+
+  c. You must comply with the conditions in Section 3(a) if You Share
+     all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+  a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+     EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+     AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+     ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+     IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+     WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+     PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+     ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+     KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+     ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+  b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+     TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+     NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+     INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+     COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+     USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+     ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+     DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+     IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+  c. The disclaimer of warranties and limitation of liability provided
+     above shall be interpreted in a manner that, to the extent
+     possible, most closely approximates an absolute disclaimer and
+     waiver of all liability.
+
+
+Section 6 -- Term and Termination.
+
+  a. This Public License applies for the term of the Copyright and
+     Similar Rights licensed here. However, if You fail to comply with
+     this Public License, then Your rights under this Public License
+     terminate automatically.
+
+  b. Where Your right to use the Licensed Material has terminated under
+     Section 6(a), it reinstates:
+
+       1. automatically as of the date the violation is cured, provided
+          it is cured within 30 days of Your discovery of the
+          violation; or
+
+       2. upon express reinstatement by the Licensor.
+
+     For the avoidance of doubt, this Section 6(b) does not affect any
+     right the Licensor may have to seek remedies for Your violations
+     of this Public License.
+
+  c. For the avoidance of doubt, the Licensor may also offer the
+     Licensed Material under separate terms or conditions or stop
+     distributing the Licensed Material at any time; however, doing so
+     will not terminate this Public License.
+
+  d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+     License.
+
+
+Section 7 -- Other Terms and Conditions.
+
+  a. The Licensor shall not be bound by any additional or different
+     terms or conditions communicated by You unless expressly agreed.
+
+  b. Any arrangements, understandings, or agreements regarding the
+     Licensed Material not stated herein are separate from and
+     independent of the terms and conditions of this Public License.
+
+
+Section 8 -- Interpretation.
+
+  a. For the avoidance of doubt, this Public License does not, and
+     shall not be interpreted to, reduce, limit, restrict, or impose
+     conditions on any use of the Licensed Material that could lawfully
+     be made without permission under this Public License.
+
+  b. To the extent possible, if any provision of this Public License is
+     deemed unenforceable, it shall be automatically reformed to the
+     minimum extent necessary to make it enforceable. If the provision
+     cannot be reformed, it shall be severed from this Public License
+     without affecting the enforceability of the remaining terms and
+     conditions.
+
+  c. No term or condition of this Public License will be waived and no
+     failure to comply consented to unless expressly agreed to by the
+     Licensor.
+
+  d. Nothing in this Public License constitutes or may be interpreted
+     as a limitation upon, or waiver of, any privileges and immunities
+     that apply to the Licensor or You, including from the legal
+     processes of any jurisdiction or authority.
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.

ImageBind/README.md

@@ -0,0 +1,155 @@
+# ImageBind: One Embedding Space To Bind Them All
+
+**[FAIR, Meta AI](https://ai.facebook.com/research/)** 
+
+Rohit Girdhar*,
+Alaaeldin El-Nouby*,
+Zhuang Liu,
+Mannat Singh,
+Kalyan Vasudev Alwala,
+Armand Joulin,
+Ishan Misra*
+
+To appear at CVPR 2023 (*Highlighted paper*)
+
+[[`Paper`](https://facebookresearch.github.io/ImageBind/paper)] [[`Blog`](https://ai.facebook.com/blog/imagebind-six-modalities-binding-ai/)] [[`Demo`](https://imagebind.metademolab.com/)] [[`Supplementary Video`](https://dl.fbaipublicfiles.com/imagebind/imagebind_video.mp4)] [[`BibTex`](#citing-imagebind)]
+
+PyTorch implementation and pretrained models for ImageBind. For details, see the paper: **[ImageBind: One Embedding Space To Bind Them All](https://facebookresearch.github.io/ImageBind/paper)**.
+
+ImageBind learns a joint embedding across six different modalities - images, text, audio, depth, thermal, and IMU data. It enables novel emergent applications ‘out-of-the-box’ including cross-modal retrieval, composing modalities with arithmetic, cross-modal detection and generation.
+
+
+
+![ImageBind](https://user-images.githubusercontent.com/8495451/236859695-ffa13364-3e39-4d99-a8da-fbfab17f9a6b.gif)
+
+## ImageBind model
+
+Emergent zero-shot classification performance.
+
+<table style="margin: auto">
+  <tr>
+    <th>Model</th>
+    <th><span style="color:blue">IN1k</span></th>
+    <th><span style="color:purple">K400</span></th>
+    <th><span style="color:green">NYU-D</span></th>
+    <th><span style="color:LightBlue">ESC</span></th>
+    <th><span style="color:orange">LLVIP</span></th>
+    <th><span style="color:purple">Ego4D</span></th>
+    <th>download</th>
+  </tr>
+  <tr>
+    <td>imagebind_huge</td>
+    <td align="right">77.7</td>
+    <td align="right">50.0</td>
+    <td align="right">54.0</td>
+    <td align="right">66.9</td>
+    <td align="right">63.4</td>
+    <td align="right">25.0</td>
+    <td><a href="https://dl.fbaipublicfiles.com/imagebind/imagebind_huge.pth">checkpoint</a></td>
+  </tr>
+  
+</table>
+
+## Usage
+
+Install pytorch 1.13+ and other 3rd party dependencies.
+
+```shell
+conda create --name imagebind python=3.10 -y
+conda activate imagebind
+
+pip install .
+```
+
+For windows users, you might need to install `soundfile` for reading/writing audio files. (Thanks @congyue1977)
+
+```
+pip install soundfile
+```
+
+
+Extract and compare features across modalities (e.g. Image, Text and Audio).
+
+```python
+from imagebind import data
+import torch
+from imagebind.models import imagebind_model
+from imagebind.models.imagebind_model import ModalityType
+
+text_list=["A dog.", "A car", "A bird"]
+image_paths=[".assets/dog_image.jpg", ".assets/car_image.jpg", ".assets/bird_image.jpg"]
+audio_paths=[".assets/dog_audio.wav", ".assets/car_audio.wav", ".assets/bird_audio.wav"]
+
+device = "cuda:0" if torch.cuda.is_available() else "cpu"
+
+# Instantiate model
+model = imagebind_model.imagebind_huge(pretrained=True)
+model.eval()
+model.to(device)
+
+# Load data
+inputs = {
+    ModalityType.TEXT: data.load_and_transform_text(text_list, device),
+    ModalityType.VISION: data.load_and_transform_vision_data(image_paths, device),
+    ModalityType.AUDIO: data.load_and_transform_audio_data(audio_paths, device),
+}
+
+with torch.no_grad():
+    embeddings = model(inputs)
+
+print(
+    "Vision x Text: ",
+    torch.softmax(embeddings[ModalityType.VISION] @ embeddings[ModalityType.TEXT].T, dim=-1),
+)
+print(
+    "Audio x Text: ",
+    torch.softmax(embeddings[ModalityType.AUDIO] @ embeddings[ModalityType.TEXT].T, dim=-1),
+)
+print(
+    "Vision x Audio: ",
+    torch.softmax(embeddings[ModalityType.VISION] @ embeddings[ModalityType.AUDIO].T, dim=-1),
+)
+
+# Expected output:
+#
+# Vision x Text:
+# tensor([[9.9761e-01, 2.3694e-03, 1.8612e-05],
+#         [3.3836e-05, 9.9994e-01, 2.4118e-05],
+#         [4.7997e-05, 1.3496e-02, 9.8646e-01]])
+#
+# Audio x Text:
+# tensor([[1., 0., 0.],
+#         [0., 1., 0.],
+#         [0., 0., 1.]])
+#
+# Vision x Audio:
+# tensor([[0.8070, 0.1088, 0.0842],
+#         [0.1036, 0.7884, 0.1079],
+#         [0.0018, 0.0022, 0.9960]])
+
+```
+
+## Model card
+Please see the [model card](model_card.md) for details.
+
+## License
+
+ImageBind code and model weights are released under the CC-BY-NC 4.0 license. See [LICENSE](LICENSE) for additional details.
+
+## Contributing
+
+See [contributing](CONTRIBUTING.md) and the [code of conduct](CODE_OF_CONDUCT.md).
+
+## Citing ImageBind
+
+If you find this repository useful, please consider giving a star :star: and citation
+
+```
+@inproceedings{girdhar2023imagebind,
+  title={ImageBind: One Embedding Space To Bind Them All},
+  author={Girdhar, Rohit and El-Nouby, Alaaeldin and Liu, Zhuang
+and Singh, Mannat and Alwala, Kalyan Vasudev and Joulin, Armand and Misra, Ishan},
+  booktitle={CVPR},
+  year={2023}
+}
+```

ImageBind/build/lib/imagebind/__init__.py

@@ -0,0 +1,3 @@
+from imagebind import data
+from imagebind.models import imagebind_model
+from imagebind.models.imagebind_model import ModalityType

ImageBind/build/lib/imagebind/data.py

@@ -0,0 +1,343 @@
+#!/usr/bin/env python3
+# Portions Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+import math
+import pkg_resources
+
+import torch
+import torch.nn as nn
+import torchaudio
+from PIL import Image
+from pytorchvideo import transforms as pv_transforms
+from pytorchvideo.data.clip_sampling import ConstantClipsPerVideoSampler
+from pytorchvideo.data.encoded_video import EncodedVideo
+from torchvision import transforms
+from torchvision.transforms._transforms_video import NormalizeVideo
+
+from imagebind.models.multimodal_preprocessors import SimpleTokenizer
+
+DEFAULT_AUDIO_FRAME_SHIFT_MS = 10  # in milliseconds
+
+
+def return_bpe_path():
+    return pkg_resources.resource_filename(
+        "imagebind", "bpe/bpe_simple_vocab_16e6.txt.gz"
+    )
+
+
+def waveform2melspec(waveform, sample_rate, num_mel_bins, target_length):
+    # Based on https://github.com/YuanGongND/ast/blob/d7d8b4b8e06cdaeb6c843cdb38794c1c7692234c/src/dataloader.py#L102
+    waveform -= waveform.mean()
+    fbank = torchaudio.compliance.kaldi.fbank(
+        waveform,
+        htk_compat=True,
+        sample_frequency=sample_rate,
+        use_energy=False,
+        window_type="hanning",
+        num_mel_bins=num_mel_bins,
+        dither=0.0,
+        frame_length=25,
+        frame_shift=DEFAULT_AUDIO_FRAME_SHIFT_MS,
+    )
+    # Convert to [mel_bins, num_frames] shape
+    fbank = fbank.transpose(0, 1)
+    # Pad to target_length
+    n_frames = fbank.size(1)
+    p = target_length - n_frames
+    # if p is too large (say >20%), flash a warning
+    if abs(p) / n_frames > 0.2:
+        logging.warning(
+            "Large gap between audio n_frames(%d) and "
+            "target_length (%d). Is the audio_target_length "
+            "setting correct?",
+            n_frames,
+            target_length,
+        )
+    # cut and pad
+    if p > 0:
+        fbank = torch.nn.functional.pad(fbank, (0, p), mode="constant", value=0)
+    elif p < 0:
+        fbank = fbank[:, 0:target_length]
+    # Convert to [1, mel_bins, num_frames] shape, essentially like a 1
+    # channel image
+    fbank = fbank.unsqueeze(0)
+    return fbank
+
+
+def get_clip_timepoints(clip_sampler, duration):
+    # Read out all clips in this video
+    all_clips_timepoints = []
+    is_last_clip = False
+    end = 0.0
+    while not is_last_clip:
+        start, end, _, _, is_last_clip = clip_sampler(end, duration, annotation=None)
+        all_clips_timepoints.append((start, end))
+    return all_clips_timepoints
+
+
+def load_and_transform_vision_data(image_paths, device):
+    if image_paths is None:
+        return None
+
+    image_outputs = []
+
+    data_transform = transforms.Compose(
+        [
+            transforms.Resize(224, interpolation=transforms.InterpolationMode.BICUBIC),
+            transforms.CenterCrop(224),
+            transforms.ToTensor(),
+            transforms.Normalize(
+                mean=(0.48145466, 0.4578275, 0.40821073),
+                std=(0.26862954, 0.26130258, 0.27577711),
+            ),
+        ]
+    )
+
+    for image_path in image_paths:
+        with open(image_path, "rb") as fopen:
+            image = Image.open(fopen).convert("RGB")
+
+        image = data_transform(image).to(device)
+        image_outputs.append(image)
+    return torch.stack(image_outputs, dim=0)
+
+
+def load_and_transform_text(text, device):
+    if text is None:
+        return None
+    tokenizer = SimpleTokenizer(bpe_path=return_bpe_path())
+    tokens = [tokenizer(t).unsqueeze(0).to(device) for t in text]
+    tokens = torch.cat(tokens, dim=0)
+    return tokens
+
+
+def load_and_transform_audio_data(
+    audio_paths,
+    device,
+    num_mel_bins=128,
+    target_length=204,
+    sample_rate=16000,
+    clip_duration=2,
+    clips_per_video=3,
+    mean=-4.268,
+    std=9.138,
+):
+    if audio_paths is None:
+        return None
+
+    audio_outputs = []
+    clip_sampler = ConstantClipsPerVideoSampler(
+        clip_duration=clip_duration, clips_per_video=clips_per_video
+    )
+
+    for audio_path in audio_paths:
+        waveform, sr = torchaudio.load(audio_path)
+        if sample_rate != sr:
+            waveform = torchaudio.functional.resample(
+                waveform, orig_freq=sr, new_freq=sample_rate
+            )
+        all_clips_timepoints = get_clip_timepoints(
+            clip_sampler, waveform.size(1) / sample_rate
+        )
+        all_clips = []
+        for clip_timepoints in all_clips_timepoints:
+            waveform_clip = waveform[
+                :,
+                int(clip_timepoints[0] * sample_rate) : int(
+                    clip_timepoints[1] * sample_rate
+                ),
+            ]
+            waveform_melspec = waveform2melspec(
+                waveform_clip, sample_rate, num_mel_bins, target_length
+            )
+            all_clips.append(waveform_melspec)
+
+        normalize = transforms.Normalize(mean=mean, std=std)
+        all_clips = [normalize(ac).to(device) for ac in all_clips]
+
+        all_clips = torch.stack(all_clips, dim=0)
+        audio_outputs.append(all_clips)
+
+    return torch.stack(audio_outputs, dim=0)
+
+
+def crop_boxes(boxes, x_offset, y_offset):
+    """
+    Perform crop on the bounding boxes given the offsets.
+    Args:
+        boxes (ndarray or None): bounding boxes to perform crop. The dimension
+            is `num boxes` x 4.
+        x_offset (int): cropping offset in the x axis.
+        y_offset (int): cropping offset in the y axis.
+    Returns:
+        cropped_boxes (ndarray or None): the cropped boxes with dimension of
+            `num boxes` x 4.
+    """
+    cropped_boxes = boxes.copy()
+    cropped_boxes[:, [0, 2]] = boxes[:, [0, 2]] - x_offset
+    cropped_boxes[:, [1, 3]] = boxes[:, [1, 3]] - y_offset
+
+    return cropped_boxes
+
+
+def uniform_crop(images, size, spatial_idx, boxes=None, scale_size=None):
+    """
+    Perform uniform spatial sampling on the images and corresponding boxes.
+    Args:
+        images (tensor): images to perform uniform crop. The dimension is
+            `num frames` x `channel` x `height` x `width`.
+        size (int): size of height and weight to crop the images.
+        spatial_idx (int): 0, 1, or 2 for left, center, and right crop if width
+            is larger than height. Or 0, 1, or 2 for top, center, and bottom
+            crop if height is larger than width.
+        boxes (ndarray or None): optional. Corresponding boxes to images.
+            Dimension is `num boxes` x 4.
+        scale_size (int): optinal. If not None, resize the images to scale_size before
+            performing any crop.
+    Returns:
+        cropped (tensor): images with dimension of
+            `num frames` x `channel` x `size` x `size`.
+        cropped_boxes (ndarray or None): the cropped boxes with dimension of
+            `num boxes` x 4.
+    """
+    assert spatial_idx in [0, 1, 2]
+    ndim = len(images.shape)
+    if ndim == 3:
+        images = images.unsqueeze(0)
+    height = images.shape[2]
+    width = images.shape[3]
+
+    if scale_size is not None:
+        if width <= height:
+            width, height = scale_size, int(height / width * scale_size)
+        else:
+            width, height = int(width / height * scale_size), scale_size
+        images = torch.nn.functional.interpolate(
+            images,
+            size=(height, width),
+            mode="bilinear",
+            align_corners=False,
+        )
+
+    y_offset = int(math.ceil((height - size) / 2))
+    x_offset = int(math.ceil((width - size) / 2))
+
+    if height > width:
+        if spatial_idx == 0:
+            y_offset = 0
+        elif spatial_idx == 2:
+            y_offset = height - size
+    else:
+        if spatial_idx == 0:
+            x_offset = 0
+        elif spatial_idx == 2:
+            x_offset = width - size
+    cropped = images[:, :, y_offset : y_offset + size, x_offset : x_offset + size]
+    cropped_boxes = crop_boxes(boxes, x_offset, y_offset) if boxes is not None else None
+    if ndim == 3:
+        cropped = cropped.squeeze(0)
+    return cropped, cropped_boxes
+
+
+class SpatialCrop(nn.Module):
+    """
+    Convert the video into 3 smaller clips spatially. Must be used after the
+        temporal crops to get spatial crops, and should be used with
+        -2 in the spatial crop at the slowfast augmentation stage (so full
+        frames are passed in here). Will return a larger list with the
+        3x spatial crops as well.
+    """
+
+    def __init__(self, crop_size: int = 224, num_crops: int = 3):
+        super().__init__()
+        self.crop_size = crop_size
+        if num_crops == 3:
+            self.crops_to_ext = [0, 1, 2]
+            self.flipped_crops_to_ext = []
+        elif num_crops == 1:
+            self.crops_to_ext = [1]
+            self.flipped_crops_to_ext = []
+        else:
+            raise NotImplementedError("Nothing else supported yet")
+
+    def forward(self, videos):
+        """
+        Args:
+            videos: A list of C, T, H, W videos.
+        Returns:
+            videos: A list with 3x the number of elements. Each video converted
+                to C, T, H', W' by spatial cropping.
+        """
+        assert isinstance(videos, list), "Must be a list of videos after temporal crops"
+        assert all([video.ndim == 4 for video in videos]), "Must be (C,T,H,W)"
+        res = []
+        for video in videos:
+            for spatial_idx in self.crops_to_ext:
+                res.append(uniform_crop(video, self.crop_size, spatial_idx)[0])
+            if not self.flipped_crops_to_ext:
+                continue
+            flipped_video = transforms.functional.hflip(video)
+            for spatial_idx in self.flipped_crops_to_ext:
+                res.append(uniform_crop(flipped_video, self.crop_size, spatial_idx)[0])
+        return res
+
+
+def load_and_transform_video_data(
+    video_paths,
+    device,
+    clip_duration=2,
+    clips_per_video=5,
+    sample_rate=16000,
+):
+    if video_paths is None:
+        return None
+
+    video_outputs = []
+    video_transform = transforms.Compose(
+        [
+            pv_transforms.ShortSideScale(224),
+            NormalizeVideo(
+                mean=(0.48145466, 0.4578275, 0.40821073),
+                std=(0.26862954, 0.26130258, 0.27577711),
+            ),
+        ]
+    )
+
+    clip_sampler = ConstantClipsPerVideoSampler(
+        clip_duration=clip_duration, clips_per_video=clips_per_video
+    )
+    frame_sampler = pv_transforms.UniformTemporalSubsample(num_samples=clip_duration)
+
+    for video_path in video_paths:
+        video = EncodedVideo.from_path(
+            video_path,
+            decoder="decord",
+            decode_audio=False,
+            **{"sample_rate": sample_rate},
+        )
+
+        all_clips_timepoints = get_clip_timepoints(clip_sampler, video.duration)
+
+        all_video = []
+        for clip_timepoints in all_clips_timepoints:
+            # Read the clip, get frames
+            clip = video.get_clip(clip_timepoints[0], clip_timepoints[1])
+            if clip is None:
+                raise ValueError("No clip found")
+            video_clip = frame_sampler(clip["video"])
+            video_clip = video_clip / 255.0  # since this is float, need 0-1
+
+            all_video.append(video_clip)
+
+        all_video = [video_transform(clip) for clip in all_video]
+        all_video = SpatialCrop(224, num_crops=3)(all_video)
+
+        all_video = torch.stack(all_video, dim=0)
+        video_outputs.append(all_video)
+
+    return torch.stack(video_outputs, dim=0).to(device)

ImageBind/build/lib/imagebind/models/helpers.py

@@ -0,0 +1,140 @@
+#!/usr/bin/env python3
+# Portions Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import einops
+import numpy as np
+import torch
+import torch.nn as nn
+
+
+class Normalize(nn.Module):
+    def __init__(self, dim: int) -> None:
+        super().__init__()
+        self.dim = dim
+
+    def forward(self, x):
+        return torch.nn.functional.normalize(x, dim=self.dim, p=2)
+
+
+class LearnableLogitScaling(nn.Module):
+    def __init__(
+        self,
+        logit_scale_init: float = 1 / 0.07,
+        learnable: bool = True,
+        max_logit_scale: float = 100,
+    ) -> None:
+        super().__init__()
+        self.max_logit_scale = max_logit_scale
+        self.logit_scale_init = logit_scale_init
+        self.learnable = learnable
+        log_logit_scale = torch.ones([]) * np.log(self.logit_scale_init)
+        if learnable:
+            self.log_logit_scale = nn.Parameter(log_logit_scale)
+        else:
+            self.register_buffer("log_logit_scale", log_logit_scale)
+
+    def forward(self, x):
+        return torch.clip(self.log_logit_scale.exp(), max=self.max_logit_scale) * x
+
+    def extra_repr(self):
+        st = f"logit_scale_init={self.logit_scale_init},learnable={self.learnable}," \
+             f" max_logit_scale={self.max_logit_scale}"
+        return st
+
+
+class EinOpsRearrange(nn.Module):
+    def __init__(self, rearrange_expr: str, **kwargs) -> None:
+        super().__init__()
+        self.rearrange_expr = rearrange_expr
+        self.kwargs = kwargs
+
+    def forward(self, x):
+        assert isinstance(x, torch.Tensor)
+        return einops.rearrange(x, self.rearrange_expr, **self.kwargs)
+
+
+class VerboseNNModule(nn.Module):
+    """
+    Wrapper around nn.Module that prints registered buffers and parameter names.
+    """
+
+    @staticmethod
+    def get_readable_tensor_repr(name: str, tensor: torch.Tensor) -> str:
+        st = (
+            "("
+            + name
+            + "): "
+            + "tensor("
+            + str(tuple(tensor[1].shape))
+            + ", requires_grad="
+            + str(tensor[1].requires_grad)
+            + ")\n"
+        )
+        return st
+
+    def extra_repr(self) -> str:
+        named_modules = set()
+        for p in self.named_modules():
+            named_modules.update([p[0]])
+        named_modules = list(named_modules)
+
+        string_repr = ""
+        for p in self.named_parameters():
+            name = p[0].split(".")[0]
+            if name not in named_modules:
+                string_repr += self.get_readable_tensor_repr(name, p)
+
+        for p in self.named_buffers():
+            name = p[0].split(".")[0]
+            string_repr += self.get_readable_tensor_repr(name, p)
+
+        return string_repr
+
+
+def cast_if_src_dtype(
+    tensor: torch.Tensor, src_dtype: torch.dtype, tgt_dtype: torch.dtype
+):
+    updated = False
+    if tensor.dtype == src_dtype:
+        tensor = tensor.to(dtype=tgt_dtype)
+        updated = True
+    return tensor, updated
+
+
+class QuickGELU(nn.Module):
+    # From https://github.com/openai/CLIP/blob/d50d76daa670286dd6cacf3bcd80b5e4823fc8e1/clip/model.py#L166
+    def forward(self, x: torch.Tensor):
+        return x * torch.sigmoid(1.702 * x)
+
+
+class SelectElement(nn.Module):
+    def __init__(self, index) -> None:
+        super().__init__()
+        self.index = index
+
+    def forward(self, x):
+        assert x.ndim >= 3
+        return x[:, self.index, ...]
+
+
+class SelectEOSAndProject(nn.Module):
+    """
+    Text Pooling used in OpenCLIP
+    """
+
+    def __init__(self, proj: nn.Module) -> None:
+        super().__init__()
+        self.proj = proj
+
+    def forward(self, x, seq_len):
+        assert x.ndim == 3
+        # x is of shape B x L x D
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        x = x[torch.arange(x.shape[0]), seq_len]
+        x = self.proj(x)
+        return x

ImageBind/build/lib/imagebind/models/imagebind_model.py

@@ -0,0 +1,506 @@
+#!/usr/bin/env python3
+# Portions Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import os
+from functools import partial
+from types import SimpleNamespace
+
+import torch
+import torch.nn as nn
+
+from imagebind.models.helpers import (EinOpsRearrange, LearnableLogitScaling, Normalize,
+                            SelectElement, SelectEOSAndProject)
+from imagebind.models.multimodal_preprocessors import (AudioPreprocessor,
+                                             IMUPreprocessor, PadIm2Video,
+                                             PatchEmbedGeneric,
+                                             RGBDTPreprocessor,
+                                             SpatioTemporalPosEmbeddingHelper,
+                                             TextPreprocessor,
+                                             ThermalPreprocessor)
+from imagebind.models.transformer import MultiheadAttention, SimpleTransformer
+
+ModalityType = SimpleNamespace(
+    VISION="vision",
+    TEXT="text",
+    AUDIO="audio",
+    THERMAL="thermal",
+    DEPTH="depth",
+    IMU="imu",
+)
+
+
+class ImageBindModel(nn.Module):
+    def __init__(
+        self,
+        video_frames=2,
+        kernel_size=(2, 14, 14),
+        audio_kernel_size=16,
+        audio_stride=10,
+        out_embed_dim=768,
+        vision_embed_dim=1024,
+        vision_num_blocks=24,
+        vision_num_heads=16,
+        audio_embed_dim=768,
+        audio_num_blocks=12,
+        audio_num_heads=12,
+        audio_num_mel_bins=128,
+        audio_target_len=204,
+        audio_drop_path=0.1,
+        text_embed_dim=768,
+        text_num_blocks=12,
+        text_num_heads=12,
+        depth_embed_dim=384,
+        depth_kernel_size=16,
+        depth_num_blocks=12,
+        depth_num_heads=8,
+        depth_drop_path=0.0,
+        thermal_embed_dim=768,
+        thermal_kernel_size=16,
+        thermal_num_blocks=12,
+        thermal_num_heads=12,
+        thermal_drop_path=0.0,
+        imu_embed_dim=512,
+        imu_kernel_size=8,
+        imu_num_blocks=6,
+        imu_num_heads=8,
+        imu_drop_path=0.7,
+    ):
+        super().__init__()
+
+        self.modality_preprocessors = self._create_modality_preprocessors(
+            video_frames,
+            vision_embed_dim,
+            kernel_size,
+            text_embed_dim,
+            audio_embed_dim,
+            audio_kernel_size,
+            audio_stride,
+            audio_num_mel_bins,
+            audio_target_len,
+            depth_embed_dim,
+            depth_kernel_size,
+            thermal_embed_dim,
+            thermal_kernel_size,
+            imu_embed_dim,
+        )
+
+        self.modality_trunks = self._create_modality_trunks(
+            vision_embed_dim,
+            vision_num_blocks,
+            vision_num_heads,
+            text_embed_dim,
+            text_num_blocks,
+            text_num_heads,
+            audio_embed_dim,
+            audio_num_blocks,
+            audio_num_heads,
+            audio_drop_path,
+            depth_embed_dim,
+            depth_num_blocks,
+            depth_num_heads,
+            depth_drop_path,
+            thermal_embed_dim,
+            thermal_num_blocks,
+            thermal_num_heads,
+            thermal_drop_path,
+            imu_embed_dim,
+            imu_num_blocks,
+            imu_num_heads,
+            imu_drop_path,
+        )
+
+        self.modality_heads = self._create_modality_heads(
+            out_embed_dim,
+            vision_embed_dim,
+            text_embed_dim,
+            audio_embed_dim,
+            depth_embed_dim,
+            thermal_embed_dim,
+            imu_embed_dim,
+        )
+
+        self.modality_postprocessors = self._create_modality_postprocessors(
+            out_embed_dim
+        )
+
+    def _create_modality_preprocessors(
+        self,
+        video_frames=2,
+        vision_embed_dim=1024,
+        kernel_size=(2, 14, 14),
+        text_embed_dim=768,
+        audio_embed_dim=768,
+        audio_kernel_size=16,
+        audio_stride=10,
+        audio_num_mel_bins=128,
+        audio_target_len=204,
+        depth_embed_dim=768,
+        depth_kernel_size=16,
+        thermal_embed_dim=768,
+        thermal_kernel_size=16,
+        imu_embed_dim=512,
+    ):
+        rgbt_stem = PatchEmbedGeneric(
+            proj_stem=[
+                PadIm2Video(pad_type="repeat", ntimes=2),
+                nn.Conv3d(
+                    in_channels=3,
+                    kernel_size=kernel_size,
+                    out_channels=vision_embed_dim,
+                    stride=kernel_size,
+                    bias=False,
+                ),
+            ]
+        )
+        rgbt_preprocessor = RGBDTPreprocessor(
+            img_size=[3, video_frames, 224, 224],
+            num_cls_tokens=1,
+            pos_embed_fn=partial(SpatioTemporalPosEmbeddingHelper, learnable=True),
+            rgbt_stem=rgbt_stem,
+            depth_stem=None,
+        )
+
+        text_preprocessor = TextPreprocessor(
+            context_length=77,
+            vocab_size=49408,
+            embed_dim=text_embed_dim,
+            causal_masking=True,
+        )
+
+        audio_stem = PatchEmbedGeneric(
+            proj_stem=[
+                nn.Conv2d(
+                    in_channels=1,
+                    kernel_size=audio_kernel_size,
+                    stride=audio_stride,
+                    out_channels=audio_embed_dim,
+                    bias=False,
+                ),
+            ],
+            norm_layer=nn.LayerNorm(normalized_shape=audio_embed_dim),
+        )
+        audio_preprocessor = AudioPreprocessor(
+            img_size=[1, audio_num_mel_bins, audio_target_len],
+            num_cls_tokens=1,
+            pos_embed_fn=partial(SpatioTemporalPosEmbeddingHelper, learnable=True),
+            audio_stem=audio_stem,
+        )
+
+        depth_stem = PatchEmbedGeneric(
+            [
+                nn.Conv2d(
+                    kernel_size=depth_kernel_size,
+                    in_channels=1,
+                    out_channels=depth_embed_dim,
+                    stride=depth_kernel_size,
+                    bias=False,
+                ),
+            ],
+            norm_layer=nn.LayerNorm(normalized_shape=depth_embed_dim),
+        )
+
+        depth_preprocessor = RGBDTPreprocessor(
+            img_size=[1, 224, 224],
+            num_cls_tokens=1,
+            pos_embed_fn=partial(SpatioTemporalPosEmbeddingHelper, learnable=True),
+            rgbt_stem=None,
+            depth_stem=depth_stem,
+        )
+
+        thermal_stem = PatchEmbedGeneric(
+            [
+                nn.Conv2d(
+                    kernel_size=thermal_kernel_size,
+                    in_channels=1,
+                    out_channels=thermal_embed_dim,
+                    stride=thermal_kernel_size,
+                    bias=False,
+                ),
+            ],
+            norm_layer=nn.LayerNorm(normalized_shape=thermal_embed_dim),
+        )
+        thermal_preprocessor = ThermalPreprocessor(
+            img_size=[1, 224, 224],
+            num_cls_tokens=1,
+            pos_embed_fn=partial(SpatioTemporalPosEmbeddingHelper, learnable=True),
+            thermal_stem=thermal_stem,
+        )
+
+        imu_stem = PatchEmbedGeneric(
+            [
+                nn.Linear(
+                    in_features=48,
+                    out_features=imu_embed_dim,
+                    bias=False,
+                ),
+            ],
+            norm_layer=nn.LayerNorm(normalized_shape=imu_embed_dim),
+        )
+
+        imu_preprocessor = IMUPreprocessor(
+            img_size=[6, 2000],
+            num_cls_tokens=1,
+            kernel_size=8,
+            embed_dim=imu_embed_dim,
+            pos_embed_fn=partial(SpatioTemporalPosEmbeddingHelper, learnable=True),
+            imu_stem=imu_stem,
+        )
+
+        modality_preprocessors = {
+            ModalityType.VISION: rgbt_preprocessor,
+            ModalityType.TEXT: text_preprocessor,
+            ModalityType.AUDIO: audio_preprocessor,
+            ModalityType.DEPTH: depth_preprocessor,
+            ModalityType.THERMAL: thermal_preprocessor,
+            ModalityType.IMU: imu_preprocessor,
+        }
+
+        return nn.ModuleDict(modality_preprocessors)
+
+    def _create_modality_trunks(
+        self,
+        vision_embed_dim=1024,
+        vision_num_blocks=24,
+        vision_num_heads=16,
+        text_embed_dim=768,
+        text_num_blocks=12,
+        text_num_heads=12,
+        audio_embed_dim=768,
+        audio_num_blocks=12,
+        audio_num_heads=12,
+        audio_drop_path=0.0,
+        depth_embed_dim=768,
+        depth_num_blocks=12,
+        depth_num_heads=12,
+        depth_drop_path=0.0,
+        thermal_embed_dim=768,
+        thermal_num_blocks=12,
+        thermal_num_heads=12,
+        thermal_drop_path=0.0,
+        imu_embed_dim=512,
+        imu_num_blocks=6,
+        imu_num_heads=8,
+        imu_drop_path=0.7,
+    ):
+        def instantiate_trunk(
+            embed_dim, num_blocks, num_heads, pre_transformer_ln, add_bias_kv, drop_path
+        ):
+            return SimpleTransformer(
+                embed_dim=embed_dim,
+                num_blocks=num_blocks,
+                ffn_dropout_rate=0.0,
+                drop_path_rate=drop_path,
+                attn_target=partial(
+                    MultiheadAttention,
+                    embed_dim=embed_dim,
+                    num_heads=num_heads,
+                    bias=True,
+                    add_bias_kv=add_bias_kv,
+                ),
+                pre_transformer_layer=nn.Sequential(
+                    nn.LayerNorm(embed_dim, eps=1e-6)
+                    if pre_transformer_ln
+                    else nn.Identity(),
+                    EinOpsRearrange("b l d -> l b d"),
+                ),
+                post_transformer_layer=EinOpsRearrange("l b d -> b l d"),
+            )
+
+        modality_trunks = {}
+        modality_trunks[ModalityType.VISION] = instantiate_trunk(
+            vision_embed_dim,
+            vision_num_blocks,
+            vision_num_heads,
+            pre_transformer_ln=True,
+            add_bias_kv=False,
+            drop_path=0.0,
+        )
+        modality_trunks[ModalityType.TEXT] = instantiate_trunk(
+            text_embed_dim,
+            text_num_blocks,
+            text_num_heads,
+            pre_transformer_ln=False,
+            add_bias_kv=False,
+            drop_path=0.0,
+        )
+        modality_trunks[ModalityType.AUDIO] = instantiate_trunk(
+            audio_embed_dim,
+            audio_num_blocks,
+            audio_num_heads,
+            pre_transformer_ln=False,
+            add_bias_kv=True,
+            drop_path=audio_drop_path,
+        )
+        modality_trunks[ModalityType.DEPTH] = instantiate_trunk(
+            depth_embed_dim,
+            depth_num_blocks,
+            depth_num_heads,
+            pre_transformer_ln=False,
+            add_bias_kv=True,
+            drop_path=depth_drop_path,
+        )
+        modality_trunks[ModalityType.THERMAL] = instantiate_trunk(
+            thermal_embed_dim,
+            thermal_num_blocks,
+            thermal_num_heads,
+            pre_transformer_ln=False,
+            add_bias_kv=True,
+            drop_path=thermal_drop_path,
+        )
+        modality_trunks[ModalityType.IMU] = instantiate_trunk(
+            imu_embed_dim,
+            imu_num_blocks,
+            imu_num_heads,
+            pre_transformer_ln=False,
+            add_bias_kv=True,
+            drop_path=imu_drop_path,
+        )
+
+        return nn.ModuleDict(modality_trunks)
+
+    def _create_modality_heads(
+        self,
+        out_embed_dim,
+        vision_embed_dim,
+        text_embed_dim,
+        audio_embed_dim,
+        depth_embed_dim,
+        thermal_embed_dim,
+        imu_embed_dim,
+    ):
+        modality_heads = {}
+
+        modality_heads[ModalityType.VISION] = nn.Sequential(
+            nn.LayerNorm(normalized_shape=vision_embed_dim, eps=1e-6),
+            SelectElement(index=0),
+            nn.Linear(vision_embed_dim, out_embed_dim, bias=False),
+        )
+
+        modality_heads[ModalityType.TEXT] = SelectEOSAndProject(
+            proj=nn.Sequential(
+                nn.LayerNorm(normalized_shape=text_embed_dim, eps=1e-6),
+                nn.Linear(text_embed_dim, out_embed_dim, bias=False),
+            )
+        )
+
+        modality_heads[ModalityType.AUDIO] = nn.Sequential(
+            nn.LayerNorm(normalized_shape=audio_embed_dim, eps=1e-6),
+            SelectElement(index=0),
+            nn.Linear(audio_embed_dim, out_embed_dim, bias=False),
+        )
+
+        modality_heads[ModalityType.DEPTH] = nn.Sequential(
+            nn.LayerNorm(normalized_shape=depth_embed_dim, eps=1e-6),
+            SelectElement(index=0),
+            nn.Linear(depth_embed_dim, out_embed_dim, bias=False),
+        )
+
+        modality_heads[ModalityType.THERMAL] = nn.Sequential(
+            nn.LayerNorm(normalized_shape=thermal_embed_dim, eps=1e-6),
+            SelectElement(index=0),
+            nn.Linear(thermal_embed_dim, out_embed_dim, bias=False),
+        )
+
+        modality_heads[ModalityType.IMU] = nn.Sequential(
+            nn.LayerNorm(normalized_shape=imu_embed_dim, eps=1e-6),
+            SelectElement(index=0),
+            nn.Dropout(p=0.5),
+            nn.Linear(imu_embed_dim, out_embed_dim, bias=False),
+        )
+
+        return nn.ModuleDict(modality_heads)
+
+    def _create_modality_postprocessors(self, out_embed_dim):
+        modality_postprocessors = {}
+
+        modality_postprocessors[ModalityType.VISION] = Normalize(dim=-1)
+        modality_postprocessors[ModalityType.TEXT] = nn.Sequential(
+            Normalize(dim=-1), LearnableLogitScaling(learnable=True)
+        )
+        modality_postprocessors[ModalityType.AUDIO] = nn.Sequential(
+            Normalize(dim=-1),
+            LearnableLogitScaling(logit_scale_init=20.0, learnable=False),
+        )
+        modality_postprocessors[ModalityType.DEPTH] = nn.Sequential(
+            Normalize(dim=-1),
+            LearnableLogitScaling(logit_scale_init=5.0, learnable=False),
+        )
+        modality_postprocessors[ModalityType.THERMAL] = nn.Sequential(
+            Normalize(dim=-1),
+            LearnableLogitScaling(logit_scale_init=10.0, learnable=False),
+        )
+        modality_postprocessors[ModalityType.IMU] = nn.Sequential(
+            Normalize(dim=-1),
+            LearnableLogitScaling(logit_scale_init=5.0, learnable=False),
+        )
+
+        return nn.ModuleDict(modality_postprocessors)
+
+    def forward(self, inputs):
+        outputs = {}
+        for modality_key, modality_value in inputs.items():
+            reduce_list = (
+                modality_value.ndim >= 5
+            )  # Audio and Video inputs consist of multiple clips
+            if reduce_list:
+                B, S = modality_value.shape[:2]
+                modality_value = modality_value.reshape(
+                    B * S, *modality_value.shape[2:]
+                )
+
+            if modality_value is not None:
+                modality_value = self.modality_preprocessors[modality_key](
+                    **{modality_key: modality_value}
+                )
+                trunk_inputs = modality_value["trunk"]
+                head_inputs = modality_value["head"]
+                modality_value = self.modality_trunks[modality_key](**trunk_inputs)
+                modality_value = self.modality_heads[modality_key](
+                    modality_value, **head_inputs
+                )
+                modality_value = self.modality_postprocessors[modality_key](
+                    modality_value
+                )
+
+                if reduce_list:
+                    modality_value = modality_value.reshape(B, S, -1)
+                    modality_value = modality_value.mean(dim=1)
+
+                outputs[modality_key] = modality_value
+
+        return outputs
+
+
+def imagebind_huge(pretrained=False):
+    model = ImageBindModel(
+        vision_embed_dim=1280,
+        vision_num_blocks=32,
+        vision_num_heads=16,
+        text_embed_dim=1024,
+        text_num_blocks=24,
+        text_num_heads=16,
+        out_embed_dim=1024,
+        audio_drop_path=0.1,
+        imu_drop_path=0.7,
+    )
+
+    if pretrained:
+        if not os.path.exists(".checkpoints/imagebind_huge.pth"):
+            print(
+                "Downloading imagebind weights to .checkpoints/imagebind_huge.pth ..."
+            )
+            os.makedirs(".checkpoints", exist_ok=True)
+            torch.hub.download_url_to_file(
+                "https://dl.fbaipublicfiles.com/imagebind/imagebind_huge.pth",
+                ".checkpoints/imagebind_huge.pth",
+                progress=True,
+            )
+
+        model.load_state_dict(torch.load(".checkpoints/imagebind_huge.pth"))
+
+    return model

ImageBind/build/lib/imagebind/models/multimodal_preprocessors.py

@@ -0,0 +1,685 @@
+#!/usr/bin/env python3
+# Portions Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import gzip
+import html
+import io
+import math
+from functools import lru_cache
+from typing import Callable, List, Optional, Tuple
+
+import ftfy
+import numpy as np
+import regex as re
+import torch
+import torch.nn as nn
+from iopath.common.file_io import g_pathmgr
+from timm.models.layers import trunc_normal_
+
+from imagebind.models.helpers import VerboseNNModule, cast_if_src_dtype
+
+
+def get_sinusoid_encoding_table(n_position, d_hid):
+    """Sinusoid position encoding table"""
+
+    # TODO: make it with torch instead of numpy
+    def get_position_angle_vec(position):
+        return [
+            position / np.power(10000, 2 * (hid_j // 2) / d_hid)
+            for hid_j in range(d_hid)
+        ]
+
+    sinusoid_table = np.array(
+        [get_position_angle_vec(pos_i) for pos_i in range(n_position)]
+    )
+    sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2])  # dim 2i
+    sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2])  # dim 2i+1
+
+    return torch.FloatTensor(sinusoid_table).unsqueeze(0)
+
+
+def interpolate_pos_encoding_2d(target_spatial_size, pos_embed):
+    N = pos_embed.shape[1]
+    if N == target_spatial_size:
+        return pos_embed
+    dim = pos_embed.shape[-1]
+    # nn.functional.interpolate doesn't work with bfloat16 so we cast to float32
+    pos_embed, updated = cast_if_src_dtype(pos_embed, torch.bfloat16, torch.float32)
+    pos_embed = nn.functional.interpolate(
+        pos_embed.reshape(1, int(math.sqrt(N)), int(math.sqrt(N)), dim).permute(
+            0, 3, 1, 2
+        ),
+        scale_factor=math.sqrt(target_spatial_size / N),
+        mode="bicubic",
+    )
+    if updated:
+        pos_embed, _ = cast_if_src_dtype(pos_embed, torch.float32, torch.bfloat16)
+    pos_embed = pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
+    return pos_embed
+
+
+def interpolate_pos_encoding(
+    npatch_per_img,
+    pos_embed,
+    patches_layout,
+    input_shape=None,
+    first_patch_idx=1,
+):
+    assert first_patch_idx == 0 or first_patch_idx == 1, "there is 1 CLS token or none"
+    N = pos_embed.shape[1] - first_patch_idx  # since it's 1 if cls_token exists
+    if npatch_per_img == N:
+        return pos_embed
+
+    assert (
+        patches_layout[-1] == patches_layout[-2]
+    ), "Interpolation of pos embed not supported for non-square layouts"
+
+    class_emb = pos_embed[:, :first_patch_idx]
+    pos_embed = pos_embed[:, first_patch_idx:]
+
+    if input_shape is None or patches_layout[0] == 1:
+        # simple 2D pos embedding, no temporal component
+        pos_embed = interpolate_pos_encoding_2d(npatch_per_img, pos_embed)
+    elif patches_layout[0] > 1:
+        # pos embed has a temporal component
+        assert len(input_shape) == 4, "temporal interpolation not supported"
+        # we only support 2D interpolation in this case
+        num_frames = patches_layout[0]
+        num_spatial_tokens = patches_layout[1] * patches_layout[2]
+        pos_embed = pos_embed.view(1, num_frames, num_spatial_tokens, -1)
+        # interpolate embedding for zeroth frame
+        pos_embed = interpolate_pos_encoding_2d(
+            npatch_per_img, pos_embed[0, 0, ...].unsqueeze(0)
+        )
+    else:
+        raise ValueError("This type of interpolation isn't implemented")
+
+    return torch.cat((class_emb, pos_embed), dim=1)
+
+
+def _get_pos_embedding(
+    npatch_per_img,
+    pos_embed,
+    patches_layout,
+    input_shape,
+    first_patch_idx=1,
+):
+    pos_embed = interpolate_pos_encoding(
+        npatch_per_img,
+        pos_embed,
+        patches_layout,
+        input_shape=input_shape,
+        first_patch_idx=first_patch_idx,
+    )
+    return pos_embed
+
+
+class PatchEmbedGeneric(nn.Module):
+    """
+    PatchEmbed from Hydra
+    """
+
+    def __init__(self, proj_stem, norm_layer: Optional[nn.Module] = None):
+        super().__init__()
+
+        if len(proj_stem) > 1:
+            self.proj = nn.Sequential(*proj_stem)
+        else:
+            # Special case to be able to load pre-trained models that were
+            # trained with a standard stem
+            self.proj = proj_stem[0]
+        self.norm_layer = norm_layer
+
+    def get_patch_layout(self, img_size):
+        with torch.no_grad():
+            dummy_img = torch.zeros(
+                [
+                    1,
+                ]
+                + img_size
+            )
+            dummy_out = self.proj(dummy_img)
+        embed_dim = dummy_out.shape[1]
+        patches_layout = tuple(dummy_out.shape[2:])
+        num_patches = np.prod(patches_layout)
+        return patches_layout, num_patches, embed_dim
+
+    def forward(self, x):
+        x = self.proj(x)
+        # B C (T) H W -> B (T)HW C
+        x = x.flatten(2).transpose(1, 2)
+        if self.norm_layer is not None:
+            x = self.norm_layer(x)
+        return x
+
+
+class SpatioTemporalPosEmbeddingHelper(VerboseNNModule):
+    def __init__(
+        self,
+        patches_layout: List,
+        num_patches: int,
+        num_cls_tokens: int,
+        embed_dim: int,
+        learnable: bool,
+    ) -> None:
+        super().__init__()
+        self.num_cls_tokens = num_cls_tokens
+        self.patches_layout = patches_layout
+        self.num_patches = num_patches
+        self.num_tokens = num_cls_tokens + num_patches
+        self.learnable = learnable
+        if self.learnable:
+            self.pos_embed = nn.Parameter(torch.zeros(1, self.num_tokens, embed_dim))
+            trunc_normal_(self.pos_embed, std=0.02)
+        else:
+            self.register_buffer(
+                "pos_embed", get_sinusoid_encoding_table(self.num_tokens, embed_dim)
+            )
+
+    def get_pos_embedding(self, vision_input, all_vision_tokens):
+        input_shape = vision_input.shape
+        pos_embed = _get_pos_embedding(
+            all_vision_tokens.size(1) - self.num_cls_tokens,
+            pos_embed=self.pos_embed,
+            patches_layout=self.patches_layout,
+            input_shape=input_shape,
+            first_patch_idx=self.num_cls_tokens,
+        )
+        return pos_embed
+
+
+class RGBDTPreprocessor(VerboseNNModule):
+    def __init__(
+        self,
+        rgbt_stem: PatchEmbedGeneric,
+        depth_stem: Optional[PatchEmbedGeneric],
+        img_size: Tuple = (3, 224, 224),
+        num_cls_tokens: int = 1,
+        pos_embed_fn: Optional[Callable] = None,
+        use_type_embed: bool = False,
+        init_param_style: str = "openclip",
+    ) -> None:
+        super().__init__()
+        stem = rgbt_stem if rgbt_stem is not None else depth_stem
+        (
+            self.patches_layout,
+            self.num_patches,
+            self.embed_dim,
+        ) = stem.get_patch_layout(img_size)
+        self.rgbt_stem = rgbt_stem
+        self.depth_stem = depth_stem
+        self.use_pos_embed = pos_embed_fn is not None
+        self.use_type_embed = use_type_embed
+        self.num_cls_tokens = num_cls_tokens
+
+        if self.use_pos_embed:
+            self.pos_embedding_helper = pos_embed_fn(
+                patches_layout=self.patches_layout,
+                num_cls_tokens=num_cls_tokens,
+                num_patches=self.num_patches,
+                embed_dim=self.embed_dim,
+            )
+        if self.num_cls_tokens > 0:
+            self.cls_token = nn.Parameter(
+                torch.zeros(1, self.num_cls_tokens, self.embed_dim)
+            )
+        if self.use_type_embed:
+            self.type_embed = nn.Parameter(torch.zeros(1, 1, self.embed_dim))
+
+        self.init_parameters(init_param_style)
+
+    @torch.no_grad()
+    def init_parameters(self, init_param_style):
+        if init_param_style == "openclip":
+            # OpenCLIP style initialization
+            scale = self.embed_dim**-0.5
+            if self.use_pos_embed:
+                nn.init.normal_(self.pos_embedding_helper.pos_embed)
+                self.pos_embedding_helper.pos_embed *= scale
+
+            if self.num_cls_tokens > 0:
+                nn.init.normal_(self.cls_token)
+                self.cls_token *= scale
+        elif init_param_style == "vit":
+            self.cls_token.data.fill_(0)
+        else:
+            raise ValueError(f"Unknown init {init_param_style}")
+
+        if self.use_type_embed:
+            nn.init.normal_(self.type_embed)
+
+    def tokenize_input_and_cls_pos(self, input, stem, mask):
+        # tokens is of shape B x L x D
+        tokens = stem(input)
+        assert tokens.ndim == 3
+        assert tokens.shape[2] == self.embed_dim
+        B = tokens.shape[0]
+        if self.num_cls_tokens > 0:
+            class_tokens = self.cls_token.expand(
+                B, -1, -1
+            )  # stole class_tokens impl from Phil Wang, thanks
+            tokens = torch.cat((class_tokens, tokens), dim=1)
+        if self.use_pos_embed:
+            pos_embed = self.pos_embedding_helper.get_pos_embedding(input, tokens)
+            tokens = tokens + pos_embed
+        if self.use_type_embed:
+            tokens = tokens + self.type_embed.expand(B, -1, -1)
+        return tokens
+
+    def forward(self, vision=None, depth=None, patch_mask=None):
+        if patch_mask is not None:
+            raise NotImplementedError()
+
+        if vision is not None:
+            vision_tokens = self.tokenize_input_and_cls_pos(
+                vision, self.rgbt_stem, patch_mask
+            )
+
+        if depth is not None:
+            depth_tokens = self.tokenize_input_and_cls_pos(
+                depth, self.depth_stem, patch_mask
+            )
+
+        # aggregate tokens
+        if vision is not None and depth is not None:
+            final_tokens = vision_tokens + depth_tokens
+        else:
+            final_tokens = vision_tokens if vision is not None else depth_tokens
+        return_dict = {
+            "trunk": {
+                "tokens": final_tokens,
+            },
+            "head": {},
+        }
+        return return_dict
+
+
+class AudioPreprocessor(RGBDTPreprocessor):
+    def __init__(self, audio_stem: PatchEmbedGeneric, **kwargs) -> None:
+        super().__init__(rgbt_stem=audio_stem, depth_stem=None, **kwargs)
+
+    def forward(self, audio=None):
+        return super().forward(vision=audio)
+
+
+class ThermalPreprocessor(RGBDTPreprocessor):
+    def __init__(self, thermal_stem: PatchEmbedGeneric, **kwargs) -> None:
+        super().__init__(rgbt_stem=thermal_stem, depth_stem=None, **kwargs)
+
+    def forward(self, thermal=None):
+        return super().forward(vision=thermal)
+
+
+def build_causal_attention_mask(context_length):
+    # lazily create causal attention mask, with full attention between the vision tokens
+    # pytorch uses additive attention mask; fill with -inf
+    mask = torch.empty(context_length, context_length, requires_grad=False)
+    mask.fill_(float("-inf"))
+    mask.triu_(1)  # zero out the lower diagonal
+    return mask
+
+
+class TextPreprocessor(VerboseNNModule):
+    def __init__(
+        self,
+        vocab_size: int,
+        context_length: int,
+        embed_dim: int,
+        causal_masking: bool,
+        supply_seq_len_to_head: bool = True,
+        num_cls_tokens: int = 0,
+        init_param_style: str = "openclip",
+    ) -> None:
+        super().__init__()
+        self.vocab_size = vocab_size
+        self.context_length = context_length
+        self.token_embedding = nn.Embedding(vocab_size, embed_dim)
+        self.pos_embed = nn.Parameter(
+            torch.empty(1, self.context_length + num_cls_tokens, embed_dim)
+        )
+        self.causal_masking = causal_masking
+        if self.causal_masking:
+            mask = build_causal_attention_mask(self.context_length)
+            # register the mask as a buffer so it can be moved to the right device
+            self.register_buffer("mask", mask)
+
+        self.supply_seq_len_to_head = supply_seq_len_to_head
+        self.num_cls_tokens = num_cls_tokens
+        self.embed_dim = embed_dim
+        if num_cls_tokens > 0:
+            assert self.causal_masking is False, "Masking + CLS token isn't implemented"
+            self.cls_token = nn.Parameter(
+                torch.zeros(1, self.num_cls_tokens, embed_dim)
+            )
+
+        self.init_parameters(init_param_style)
+
+    @torch.no_grad()
+    def init_parameters(self, init_param_style="openclip"):
+        # OpenCLIP style initialization
+        nn.init.normal_(self.token_embedding.weight, std=0.02)
+        nn.init.normal_(self.pos_embed, std=0.01)
+
+        if init_param_style == "openclip":
+            # OpenCLIP style initialization
+            scale = self.embed_dim**-0.5
+            if self.num_cls_tokens > 0:
+                nn.init.normal_(self.cls_token)
+                self.cls_token *= scale
+        elif init_param_style == "vit":
+            self.cls_token.data.fill_(0)
+        else:
+            raise ValueError(f"Unknown init {init_param_style}")
+
+    def forward(self, text):
+        # text tokens are of shape B x L x D
+        text_tokens = self.token_embedding(text)
+        # concat CLS tokens if any
+        if self.num_cls_tokens > 0:
+            B = text_tokens.shape[0]
+            class_tokens = self.cls_token.expand(
+                B, -1, -1
+            )  # stole class_tokens impl from Phil Wang, thanks
+            text_tokens = torch.cat((class_tokens, text_tokens), dim=1)
+        text_tokens = text_tokens + self.pos_embed
+        return_dict = {
+            "trunk": {
+                "tokens": text_tokens,
+            },
+            "head": {},
+        }
+        # Compute sequence length after adding CLS tokens
+        if self.supply_seq_len_to_head:
+            text_lengths = text.argmax(dim=-1)
+            return_dict["head"] = {
+                "seq_len": text_lengths,
+            }
+        if self.causal_masking:
+            return_dict["trunk"].update({"attn_mask": self.mask})
+        return return_dict
+
+
+class Im2Video(nn.Module):
+    """Convert an image into a trivial video."""
+
+    def __init__(self, time_dim=2):
+        super().__init__()
+        self.time_dim = time_dim
+
+    def forward(self, x):
+        if x.ndim == 4:
+            # B, C, H, W -> B, C, T, H, W
+            return x.unsqueeze(self.time_dim)
+        elif x.ndim == 5:
+            return x
+        else:
+            raise ValueError(f"Dimension incorrect {x.shape}")
+
+
+class PadIm2Video(Im2Video):
+    def __init__(self, ntimes, pad_type, time_dim=2):
+        super().__init__(time_dim=time_dim)
+        assert ntimes > 0
+        assert pad_type in ["zero", "repeat"]
+        self.ntimes = ntimes
+        self.pad_type = pad_type
+
+    def forward(self, x):
+        x = super().forward(x)
+        if x.shape[self.time_dim] == 1:
+            if self.pad_type == "repeat":
+                new_shape = [1] * len(x.shape)
+                new_shape[self.time_dim] = self.ntimes
+                x = x.repeat(new_shape)
+            elif self.pad_type == "zero":
+                padarg = [0, 0] * len(x.shape)
+                padarg[2 * self.time_dim + 1] = self.ntimes - x.shape[self.time_dim]
+                x = nn.functional.pad(x, padarg)
+        return x
+
+
+# Modified from github.com/openai/CLIP
+@lru_cache()
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+def get_pairs(word):
+    """Return set of symbol pairs in a word.
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+class SimpleTokenizer(object):
+    def __init__(self, bpe_path: str, context_length=77):
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+
+        with g_pathmgr.open(bpe_path, "rb") as fh:
+            bpe_bytes = io.BytesIO(fh.read())
+            merges: List[str] = gzip.open(bpe_bytes).read().decode("utf-8").split("\n")
+        merges = merges[1 : 49152 - 256 - 2 + 1]
+        merges: List[Tuple[str, ...]] = [tuple(merge.split()) for merge in merges]
+        vocab = list(bytes_to_unicode().values())
+        vocab = vocab + [v + "</w>" for v in vocab]
+        for merge in merges:
+            vocab.append("".join(merge))
+        vocab.extend(["<|startoftext|>", "<|endoftext|>"])
+        self.encoder = dict(zip(vocab, range(len(vocab))))
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.bpe_ranks = dict(zip(merges, range(len(merges))))
+        self.cache = {
+            "<|startoftext|>": "<|startoftext|>",
+            "<|endoftext|>": "<|endoftext|>",
+        }
+        self.pat = re.compile(
+            r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""",
+            re.IGNORECASE,
+        )
+        self.context_length = context_length
+
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token[:-1]) + (token[-1] + "</w>",)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token + "</w>"
+
+        while True:
+            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                    new_word.extend(word[i:j])
+                    i = j
+                except:
+                    new_word.extend(word[i:])
+                    break
+
+                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
+                    new_word.append(first + second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = " ".join(word)
+        self.cache[token] = word
+        return word
+
+    def encode(self, text):
+        bpe_tokens = []
+        text = whitespace_clean(basic_clean(text)).lower()
+        for token in re.findall(self.pat, text):
+            token = "".join(self.byte_encoder[b] for b in token.encode("utf-8"))
+            bpe_tokens.extend(
+                self.encoder[bpe_token] for bpe_token in self.bpe(token).split(" ")
+            )
+        return bpe_tokens
+
+    def decode(self, tokens):
+        text = "".join([self.decoder[token] for token in tokens])
+        text = (
+            bytearray([self.byte_decoder[c] for c in text])
+            .decode("utf-8", errors="replace")
+            .replace("</w>", " ")
+        )
+        return text
+
+    def __call__(self, texts, context_length=None):
+        if not context_length:
+            context_length = self.context_length
+
+        if isinstance(texts, str):
+            texts = [texts]
+
+        sot_token = self.encoder["<|startoftext|>"]
+        eot_token = self.encoder["<|endoftext|>"]
+        all_tokens = [[sot_token] + self.encode(text) + [eot_token] for text in texts]
+        result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)
+
+        for i, tokens in enumerate(all_tokens):
+            tokens = tokens[:context_length]
+            result[i, : len(tokens)] = torch.tensor(tokens)
+
+        if len(result) == 1:
+            return result[0]
+        return result
+
+
+class IMUPreprocessor(VerboseNNModule):
+    def __init__(
+        self,
+        kernel_size: int,
+        imu_stem: PatchEmbedGeneric,
+        embed_dim: int,
+        img_size: Tuple = (6, 2000),
+        num_cls_tokens: int = 1,
+        pos_embed_fn: Optional[Callable] = None,
+        init_param_style: str = "openclip",
+    ) -> None:
+        super().__init__()
+        self.imu_stem = imu_stem
+        self.embed_dim = embed_dim
+        self.use_pos_embed = pos_embed_fn is not None
+        self.num_cls_tokens = num_cls_tokens
+        self.kernel_size = kernel_size
+        self.pos_embed = nn.Parameter(
+            torch.empty(1, (img_size[1] // kernel_size) + num_cls_tokens, embed_dim)
+        )
+
+        if self.num_cls_tokens > 0:
+            self.cls_token = nn.Parameter(
+                torch.zeros(1, self.num_cls_tokens, self.embed_dim)
+            )
+
+        self.init_parameters(init_param_style)
+
+    @torch.no_grad()
+    def init_parameters(self, init_param_style):
+        nn.init.normal_(self.pos_embed, std=0.01)
+
+        if init_param_style == "openclip":
+            # OpenCLIP style initialization
+            scale = self.embed_dim**-0.5
+
+            if self.num_cls_tokens > 0:
+                nn.init.normal_(self.cls_token)
+                self.cls_token *= scale
+        elif init_param_style == "vit":
+            self.cls_token.data.fill_(0)
+        else:
+            raise ValueError(f"Unknown init {init_param_style}")
+
+    def tokenize_input_and_cls_pos(self, input, stem):
+        # tokens is of shape B x L x D
+        tokens = stem.norm_layer(stem.proj(input))
+        assert tokens.ndim == 3
+        assert tokens.shape[2] == self.embed_dim
+        B = tokens.shape[0]
+        if self.num_cls_tokens > 0:
+            class_tokens = self.cls_token.expand(
+                B, -1, -1
+            )  # stole class_tokens impl from Phil Wang, thanks
+            tokens = torch.cat((class_tokens, tokens), dim=1)
+        if self.use_pos_embed:
+            tokens = tokens + self.pos_embed
+        return tokens
+
+    def forward(self, imu):
+        # Patchify
+        imu = imu.unfold(
+            -1,
+            self.kernel_size,
+            self.kernel_size,
+        ).permute(0, 2, 1, 3)
+        imu = imu.reshape(imu.size(0), imu.size(1), -1)
+
+        imu_tokens = self.tokenize_input_and_cls_pos(
+            imu,
+            self.imu_stem,
+        )
+
+        return_dict = {
+            "trunk": {
+                "tokens": imu_tokens,
+            },
+            "head": {},
+        }
+        return return_dict

ImageBind/build/lib/imagebind/models/transformer.py

@@ -0,0 +1,280 @@
+#!/usr/bin/env python3
+# Portions Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+# Code modified from
+# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py ;
+# https://github.com/facebookresearch/deit/blob/main/models.py
+# and https://github.com/facebookresearch/vissl/blob/main/vissl/models/trunks/vision_transformer.py
+
+
+from functools import partial
+from typing import Callable, List, Optional
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint as checkpoint
+from timm.models.layers import DropPath, trunc_normal_
+
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_heads=8,
+        qkv_bias=False,
+        qk_scale=None,
+        attn_drop=0.0,
+        proj_drop=0.0,
+    ):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        # NOTE scale factor was wrong in my original version,
+        # can set manually to be compat with prev weights
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x):
+        B, N, C = x.shape
+        qkv = (
+            self.qkv(x)
+            .reshape(B, N, 3, self.num_heads, C // self.num_heads)
+            .permute(2, 0, 3, 1, 4)
+        )
+        q, k, v = (
+            qkv[0],
+            qkv[1],
+            qkv[2],
+        )  # make torchscript happy (cannot use tensor as tuple)
+
+        attn = (q @ k.transpose(-2, -1)) * self.scale
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class Mlp(nn.Module):
+    def __init__(
+        self,
+        in_features,
+        hidden_features=None,
+        out_features=None,
+        act_layer=nn.GELU,
+        drop=0.0,
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class MultiheadAttention(nn.MultiheadAttention):
+    def forward(self, x: torch.Tensor, attn_mask: torch.Tensor):
+        return super().forward(x, x, x, need_weights=False, attn_mask=attn_mask)[0]
+
+
+class ViTAttention(Attention):
+    def forward(self, x: torch.Tensor, attn_mask: torch.Tensor):
+        assert attn_mask is None
+        return super().forward(x)
+
+
+class BlockWithMasking(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        attn_target: Callable,
+        mlp_ratio: int = 4,
+        act_layer: Callable = nn.GELU,
+        norm_layer: Callable = nn.LayerNorm,
+        ffn_dropout_rate: float = 0.0,
+        drop_path: float = 0.0,
+        layer_scale_type: Optional[str] = None,
+        layer_scale_init_value: float = 1e-4,
+    ):
+        super().__init__()
+
+        assert not isinstance(
+            attn_target, nn.Module
+        ), "attn_target should be a Callable. Otherwise attn_target is shared across blocks!"
+        self.attn = attn_target()
+        if drop_path > 0.0:
+            self.drop_path = DropPath(drop_path)
+        else:
+            self.drop_path = nn.Identity()
+        self.norm_1 = norm_layer(dim)
+        mlp_hidden_dim = int(mlp_ratio * dim)
+        self.mlp = Mlp(
+            in_features=dim,
+            hidden_features=mlp_hidden_dim,
+            act_layer=act_layer,
+            drop=ffn_dropout_rate,
+        )
+        self.norm_2 = norm_layer(dim)
+        self.layer_scale_type = layer_scale_type
+        if self.layer_scale_type is not None:
+            assert self.layer_scale_type in [
+                "per_channel",
+                "scalar",
+            ], f"Found Layer scale type {self.layer_scale_type}"
+            if self.layer_scale_type == "per_channel":
+                # one gamma value per channel
+                gamma_shape = [1, 1, dim]
+            elif self.layer_scale_type == "scalar":
+                # single gamma value for all channels
+                gamma_shape = [1, 1, 1]
+            # two gammas: for each part of the fwd in the encoder
+            self.layer_scale_gamma1 = nn.Parameter(
+                torch.ones(size=gamma_shape) * layer_scale_init_value,
+                requires_grad=True,
+            )
+            self.layer_scale_gamma2 = nn.Parameter(
+                torch.ones(size=gamma_shape) * layer_scale_init_value,
+                requires_grad=True,
+            )
+
+    def forward(self, x: torch.Tensor, attn_mask: torch.Tensor):
+        if self.layer_scale_type is None:
+            x = x + self.drop_path(self.attn(self.norm_1(x), attn_mask))
+            x = x + self.drop_path(self.mlp(self.norm_2(x)))
+        else:
+            x = (
+                x
+                + self.drop_path(self.attn(self.norm_1(x), attn_mask))
+                * self.layer_scale_gamma1
+            )
+            x = x + self.drop_path(self.mlp(self.norm_2(x))) * self.layer_scale_gamma2
+        return x
+
+
+_LAYER_NORM = partial(nn.LayerNorm, eps=1e-6)
+
+
+class SimpleTransformer(nn.Module):
+    def __init__(
+        self,
+        attn_target: Callable,
+        embed_dim: int,
+        num_blocks: int,
+        block: Callable = BlockWithMasking,
+        pre_transformer_layer: Optional[Callable] = None,
+        post_transformer_layer: Optional[Callable] = None,
+        drop_path_rate: float = 0.0,
+        drop_path_type: str = "progressive",
+        norm_layer: Callable = _LAYER_NORM,
+        mlp_ratio: int = 4,
+        ffn_dropout_rate: float = 0.0,
+        layer_scale_type: Optional[str] = None,  # from cait; possible values are None, "per_channel", "scalar"
+        layer_scale_init_value: float = 1e-4,  # from cait; float
+        weight_init_style: str = "jax",  # possible values jax or pytorch
+    ):
+        """
+        Simple Transformer with the following features
+        1. Supports masked attention
+        2. Supports DropPath
+        3. Supports LayerScale
+        4. Supports Dropout in Attention and FFN
+        5. Makes few assumptions about the input except that it is a Tensor
+        """
+        super().__init__()
+        self.pre_transformer_layer = pre_transformer_layer
+        if drop_path_type == "progressive":
+            dpr = [x.item() for x in torch.linspace(0, drop_path_rate, num_blocks)]
+        elif drop_path_type == "uniform":
+            dpr = [drop_path_rate for i in range(num_blocks)]
+        else:
+            raise ValueError(f"Unknown drop_path_type: {drop_path_type}")
+
+        self.blocks = nn.Sequential(
+            *[
+                block(
+                    dim=embed_dim,
+                    attn_target=attn_target,
+                    mlp_ratio=mlp_ratio,
+                    ffn_dropout_rate=ffn_dropout_rate,
+                    drop_path=dpr[i],
+                    norm_layer=norm_layer,
+                    layer_scale_type=layer_scale_type,
+                    layer_scale_init_value=layer_scale_init_value,
+                )
+                for i in range(num_blocks)
+            ]
+        )
+        self.post_transformer_layer = post_transformer_layer
+        self.weight_init_style = weight_init_style
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            if self.weight_init_style == "jax":
+                # Based on MAE and official Jax ViT implementation
+                torch.nn.init.xavier_uniform_(m.weight)
+            elif self.weight_init_style == "pytorch":
+                # PyTorch ViT uses trunc_normal_
+                trunc_normal_(m.weight, std=0.02)
+
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, (nn.LayerNorm)):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    def forward(
+        self,
+        tokens: torch.Tensor,
+        attn_mask: torch.Tensor = None,
+        use_checkpoint: bool = False,
+        checkpoint_every_n: int = 1,
+        checkpoint_blk_ids: Optional[List[int]] = None,
+    ):
+        """
+        Inputs
+        - tokens: data of shape N x L x D (or L x N x D depending on the attention implementation)
+        - attn: mask of shape L x L
+
+        Output
+        - x: data of shape N x L x D (or L x N x D depending on the attention implementation)
+        """
+        if self.pre_transformer_layer:
+            tokens = self.pre_transformer_layer(tokens)
+        if use_checkpoint and checkpoint_blk_ids is None:
+            checkpoint_blk_ids = [
+                blk_id
+                for blk_id in range(len(self.blocks))
+                if blk_id % checkpoint_every_n == 0
+            ]
+        if checkpoint_blk_ids:
+            checkpoint_blk_ids = set(checkpoint_blk_ids)
+        for blk_id, blk in enumerate(self.blocks):
+            if use_checkpoint and blk_id in checkpoint_blk_ids:
+                tokens = checkpoint.checkpoint(
+                    blk, tokens, attn_mask, use_reentrant=False
+                )
+            else:
+                tokens = blk(tokens, attn_mask=attn_mask)
+        if self.post_transformer_layer:
+            tokens = self.post_transformer_layer(tokens)
+        return tokens

ImageBind/imagebind.egg-info/PKG-INFO

@@ -0,0 +1,170 @@
+Metadata-Version: 2.2
+Name: imagebind
+Version: 0.1.0
+Summary: A brief description of the package
+Home-page: https://github.com/facebookresearch/ImageBind
+Classifier: Programming Language :: Python :: 3
+Classifier: License :: Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International
+Description-Content-Type: text/markdown
+License-File: LICENSE
+Dynamic: classifier
+Dynamic: description
+Dynamic: description-content-type
+Dynamic: home-page
+Dynamic: summary
+
+# ImageBind: One Embedding Space To Bind Them All
+
+**[FAIR, Meta AI](https://ai.facebook.com/research/)** 
+
+Rohit Girdhar*,
+Alaaeldin El-Nouby*,
+Zhuang Liu,
+Mannat Singh,
+Kalyan Vasudev Alwala,
+Armand Joulin,
+Ishan Misra*
+
+To appear at CVPR 2023 (*Highlighted paper*)
+
+[[`Paper`](https://facebookresearch.github.io/ImageBind/paper)] [[`Blog`](https://ai.facebook.com/blog/imagebind-six-modalities-binding-ai/)] [[`Demo`](https://imagebind.metademolab.com/)] [[`Supplementary Video`](https://dl.fbaipublicfiles.com/imagebind/imagebind_video.mp4)] [[`BibTex`](#citing-imagebind)]
+
+PyTorch implementation and pretrained models for ImageBind. For details, see the paper: **[ImageBind: One Embedding Space To Bind Them All](https://facebookresearch.github.io/ImageBind/paper)**.
+
+ImageBind learns a joint embedding across six different modalities - images, text, audio, depth, thermal, and IMU data. It enables novel emergent applications ‘out-of-the-box’ including cross-modal retrieval, composing modalities with arithmetic, cross-modal detection and generation.
+
+
+
+![ImageBind](https://user-images.githubusercontent.com/8495451/236859695-ffa13364-3e39-4d99-a8da-fbfab17f9a6b.gif)
+
+## ImageBind model
+
+Emergent zero-shot classification performance.
+
+<table style="margin: auto">
+  <tr>
+    <th>Model</th>
+    <th><span style="color:blue">IN1k</span></th>
+    <th><span style="color:purple">K400</span></th>
+    <th><span style="color:green">NYU-D</span></th>
+    <th><span style="color:LightBlue">ESC</span></th>
+    <th><span style="color:orange">LLVIP</span></th>
+    <th><span style="color:purple">Ego4D</span></th>
+    <th>download</th>
+  </tr>
+  <tr>
+    <td>imagebind_huge</td>
+    <td align="right">77.7</td>
+    <td align="right">50.0</td>
+    <td align="right">54.0</td>
+    <td align="right">66.9</td>
+    <td align="right">63.4</td>
+    <td align="right">25.0</td>
+    <td><a href="https://dl.fbaipublicfiles.com/imagebind/imagebind_huge.pth">checkpoint</a></td>
+  </tr>
+  
+</table>
+
+## Usage
+
+Install pytorch 1.13+ and other 3rd party dependencies.
+
+```shell
+conda create --name imagebind python=3.10 -y
+conda activate imagebind
+
+pip install .
+```
+
+For windows users, you might need to install `soundfile` for reading/writing audio files. (Thanks @congyue1977)
+
+```
+pip install soundfile
+```
+
+
+Extract and compare features across modalities (e.g. Image, Text and Audio).
+
+```python
+from imagebind import data
+import torch
+from imagebind.models import imagebind_model
+from imagebind.models.imagebind_model import ModalityType
+
+text_list=["A dog.", "A car", "A bird"]
+image_paths=[".assets/dog_image.jpg", ".assets/car_image.jpg", ".assets/bird_image.jpg"]
+audio_paths=[".assets/dog_audio.wav", ".assets/car_audio.wav", ".assets/bird_audio.wav"]
+
+device = "cuda:0" if torch.cuda.is_available() else "cpu"
+
+# Instantiate model
+model = imagebind_model.imagebind_huge(pretrained=True)
+model.eval()
+model.to(device)
+
+# Load data
+inputs = {
+    ModalityType.TEXT: data.load_and_transform_text(text_list, device),
+    ModalityType.VISION: data.load_and_transform_vision_data(image_paths, device),
+    ModalityType.AUDIO: data.load_and_transform_audio_data(audio_paths, device),
+}
+
+with torch.no_grad():
+    embeddings = model(inputs)
+
+print(
+    "Vision x Text: ",
+    torch.softmax(embeddings[ModalityType.VISION] @ embeddings[ModalityType.TEXT].T, dim=-1),
+)
+print(
+    "Audio x Text: ",
+    torch.softmax(embeddings[ModalityType.AUDIO] @ embeddings[ModalityType.TEXT].T, dim=-1),
+)
+print(
+    "Vision x Audio: ",
+    torch.softmax(embeddings[ModalityType.VISION] @ embeddings[ModalityType.AUDIO].T, dim=-1),
+)
+
+# Expected output:
+#
+# Vision x Text:
+# tensor([[9.9761e-01, 2.3694e-03, 1.8612e-05],
+#         [3.3836e-05, 9.9994e-01, 2.4118e-05],
+#         [4.7997e-05, 1.3496e-02, 9.8646e-01]])
+#
+# Audio x Text:
+# tensor([[1., 0., 0.],
+#         [0., 1., 0.],
+#         [0., 0., 1.]])
+#
+# Vision x Audio:
+# tensor([[0.8070, 0.1088, 0.0842],
+#         [0.1036, 0.7884, 0.1079],
+#         [0.0018, 0.0022, 0.9960]])
+
+```
+
+## Model card
+Please see the [model card](model_card.md) for details.
+
+## License
+
+ImageBind code and model weights are released under the CC-BY-NC 4.0 license. See [LICENSE](LICENSE) for additional details.
+
+## Contributing
+
+See [contributing](CONTRIBUTING.md) and the [code of conduct](CODE_OF_CONDUCT.md).
+
+## Citing ImageBind
+
+If you find this repository useful, please consider giving a star :star: and citation
+
+```
+@inproceedings{girdhar2023imagebind,
+  title={ImageBind: One Embedding Space To Bind Them All},
+  author={Girdhar, Rohit and El-Nouby, Alaaeldin and Liu, Zhuang
+and Singh, Mannat and Alwala, Kalyan Vasudev and Joulin, Armand and Misra, Ishan},
+  booktitle={CVPR},
+  year={2023}
+}
+```

ImageBind/imagebind.egg-info/SOURCES.txt

@@ -0,0 +1,15 @@
+LICENSE
+README.md
+setup.py
+imagebind/__init__.py
+imagebind/data.py
+imagebind.egg-info/PKG-INFO
+imagebind.egg-info/SOURCES.txt
+imagebind.egg-info/dependency_links.txt
+imagebind.egg-info/top_level.txt
+imagebind/bpe/bpe_simple_vocab_16e6.txt.gz
+imagebind/models/__init__.py
+imagebind/models/helpers.py
+imagebind/models/imagebind_model.py
+imagebind/models/multimodal_preprocessors.py
+imagebind/models/transformer.py

ImageBind/imagebind.egg-info/dependency_links.txt

@@ -0,0 +1 @@
+https://download.pytorch.org/whl/cu113

ImageBind/imagebind.egg-info/top_level.txt

@@ -0,0 +1 @@
+imagebind

ImageBind/imagebind/__init__.py

@@ -0,0 +1,3 @@
+from imagebind import data
+from imagebind.models import imagebind_model
+from imagebind.models.imagebind_model import ModalityType

ImageBind/imagebind/data.py

@@ -0,0 +1,343 @@
+#!/usr/bin/env python3
+# Portions Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+import math
+import pkg_resources
+
+import torch
+import torch.nn as nn
+import torchaudio
+from PIL import Image
+from pytorchvideo import transforms as pv_transforms
+from pytorchvideo.data.clip_sampling import ConstantClipsPerVideoSampler
+from pytorchvideo.data.encoded_video import EncodedVideo
+from torchvision import transforms
+from torchvision.transforms._transforms_video import NormalizeVideo
+
+from imagebind.models.multimodal_preprocessors import SimpleTokenizer
+
+DEFAULT_AUDIO_FRAME_SHIFT_MS = 10  # in milliseconds
+
+
+def return_bpe_path():
+    return pkg_resources.resource_filename(
+        "imagebind", "bpe/bpe_simple_vocab_16e6.txt.gz"
+    )
+
+
+def waveform2melspec(waveform, sample_rate, num_mel_bins, target_length):
+    # Based on https://github.com/YuanGongND/ast/blob/d7d8b4b8e06cdaeb6c843cdb38794c1c7692234c/src/dataloader.py#L102
+    waveform -= waveform.mean()
+    fbank = torchaudio.compliance.kaldi.fbank(
+        waveform,
+        htk_compat=True,
+        sample_frequency=sample_rate,
+        use_energy=False,
+        window_type="hanning",
+        num_mel_bins=num_mel_bins,
+        dither=0.0,
+        frame_length=25,
+        frame_shift=DEFAULT_AUDIO_FRAME_SHIFT_MS,
+    )
+    # Convert to [mel_bins, num_frames] shape
+    fbank = fbank.transpose(0, 1)
+    # Pad to target_length
+    n_frames = fbank.size(1)
+    p = target_length - n_frames
+    # if p is too large (say >20%), flash a warning
+    if abs(p) / n_frames > 0.2:
+        logging.warning(
+            "Large gap between audio n_frames(%d) and "
+            "target_length (%d). Is the audio_target_length "
+            "setting correct?",
+            n_frames,
+            target_length,
+        )
+    # cut and pad
+    if p > 0:
+        fbank = torch.nn.functional.pad(fbank, (0, p), mode="constant", value=0)
+    elif p < 0:
+        fbank = fbank[:, 0:target_length]
+    # Convert to [1, mel_bins, num_frames] shape, essentially like a 1
+    # channel image
+    fbank = fbank.unsqueeze(0)
+    return fbank
+
+
+def get_clip_timepoints(clip_sampler, duration):
+    # Read out all clips in this video
+    all_clips_timepoints = []
+    is_last_clip = False
+    end = 0.0
+    while not is_last_clip:
+        start, end, _, _, is_last_clip = clip_sampler(end, duration, annotation=None)
+        all_clips_timepoints.append((start, end))
+    return all_clips_timepoints
+
+
+def load_and_transform_vision_data(image_paths, device):
+    if image_paths is None:
+        return None
+
+    image_outputs = []
+
+    data_transform = transforms.Compose(
+        [
+            transforms.Resize(224, interpolation=transforms.InterpolationMode.BICUBIC),
+            transforms.CenterCrop(224),
+            transforms.ToTensor(),
+            transforms.Normalize(
+                mean=(0.48145466, 0.4578275, 0.40821073),
+                std=(0.26862954, 0.26130258, 0.27577711),
+            ),
+        ]
+    )
+
+    for image_path in image_paths:
+        with open(image_path, "rb") as fopen:
+            image = Image.open(fopen).convert("RGB")
+
+        image = data_transform(image).to(device)
+        image_outputs.append(image)
+    return torch.stack(image_outputs, dim=0)
+
+
+def load_and_transform_text(text, device):
+    if text is None:
+        return None
+    tokenizer = SimpleTokenizer(bpe_path=return_bpe_path())
+    tokens = [tokenizer(t).unsqueeze(0).to(device) for t in text]
+    tokens = torch.cat(tokens, dim=0)
+    return tokens
+
+
+def load_and_transform_audio_data(
+    audio_paths,
+    device,
+    num_mel_bins=128,
+    target_length=204,
+    sample_rate=16000,
+    clip_duration=2,
+    clips_per_video=3,
+    mean=-4.268,
+    std=9.138,
+):
+    if audio_paths is None:
+        return None
+
+    audio_outputs = []
+    clip_sampler = ConstantClipsPerVideoSampler(
+        clip_duration=clip_duration, clips_per_video=clips_per_video
+    )
+
+    for audio_path in audio_paths:
+        waveform, sr = torchaudio.load(audio_path)
+        if sample_rate != sr:
+            waveform = torchaudio.functional.resample(
+                waveform, orig_freq=sr, new_freq=sample_rate
+            )
+        all_clips_timepoints = get_clip_timepoints(
+            clip_sampler, waveform.size(1) / sample_rate
+        )
+        all_clips = []
+        for clip_timepoints in all_clips_timepoints:
+            waveform_clip = waveform[
+                :,
+                int(clip_timepoints[0] * sample_rate) : int(
+                    clip_timepoints[1] * sample_rate
+                ),
+            ]
+            waveform_melspec = waveform2melspec(
+                waveform_clip, sample_rate, num_mel_bins, target_length
+            )
+            all_clips.append(waveform_melspec)
+
+        normalize = transforms.Normalize(mean=mean, std=std)
+        all_clips = [normalize(ac).to(device) for ac in all_clips]
+
+        all_clips = torch.stack(all_clips, dim=0)
+        audio_outputs.append(all_clips)
+
+    return torch.stack(audio_outputs, dim=0)
+
+
+def crop_boxes(boxes, x_offset, y_offset):
+    """
+    Perform crop on the bounding boxes given the offsets.
+    Args:
+        boxes (ndarray or None): bounding boxes to perform crop. The dimension
+            is `num boxes` x 4.
+        x_offset (int): cropping offset in the x axis.
+        y_offset (int): cropping offset in the y axis.
+    Returns:
+        cropped_boxes (ndarray or None): the cropped boxes with dimension of
+            `num boxes` x 4.
+    """
+    cropped_boxes = boxes.copy()
+    cropped_boxes[:, [0, 2]] = boxes[:, [0, 2]] - x_offset
+    cropped_boxes[:, [1, 3]] = boxes[:, [1, 3]] - y_offset
+
+    return cropped_boxes
+
+
+def uniform_crop(images, size, spatial_idx, boxes=None, scale_size=None):
+    """
+    Perform uniform spatial sampling on the images and corresponding boxes.
+    Args:
+        images (tensor): images to perform uniform crop. The dimension is
+            `num frames` x `channel` x `height` x `width`.
+        size (int): size of height and weight to crop the images.
+        spatial_idx (int): 0, 1, or 2 for left, center, and right crop if width
+            is larger than height. Or 0, 1, or 2 for top, center, and bottom
+            crop if height is larger than width.
+        boxes (ndarray or None): optional. Corresponding boxes to images.
+            Dimension is `num boxes` x 4.
+        scale_size (int): optinal. If not None, resize the images to scale_size before
+            performing any crop.
+    Returns:
+        cropped (tensor): images with dimension of
+            `num frames` x `channel` x `size` x `size`.
+        cropped_boxes (ndarray or None): the cropped boxes with dimension of
+            `num boxes` x 4.
+    """
+    assert spatial_idx in [0, 1, 2]
+    ndim = len(images.shape)
+    if ndim == 3:
+        images = images.unsqueeze(0)
+    height = images.shape[2]
+    width = images.shape[3]
+
+    if scale_size is not None:
+        if width <= height:
+            width, height = scale_size, int(height / width * scale_size)
+        else:
+            width, height = int(width / height * scale_size), scale_size
+        images = torch.nn.functional.interpolate(
+            images,
+            size=(height, width),
+            mode="bilinear",
+            align_corners=False,
+        )
+
+    y_offset = int(math.ceil((height - size) / 2))
+    x_offset = int(math.ceil((width - size) / 2))
+
+    if height > width:
+        if spatial_idx == 0:
+            y_offset = 0
+        elif spatial_idx == 2:
+            y_offset = height - size
+    else:
+        if spatial_idx == 0:
+            x_offset = 0
+        elif spatial_idx == 2:
+            x_offset = width - size
+    cropped = images[:, :, y_offset : y_offset + size, x_offset : x_offset + size]
+    cropped_boxes = crop_boxes(boxes, x_offset, y_offset) if boxes is not None else None
+    if ndim == 3:
+        cropped = cropped.squeeze(0)
+    return cropped, cropped_boxes
+
+
+class SpatialCrop(nn.Module):
+    """
+    Convert the video into 3 smaller clips spatially. Must be used after the
+        temporal crops to get spatial crops, and should be used with
+        -2 in the spatial crop at the slowfast augmentation stage (so full
+        frames are passed in here). Will return a larger list with the
+        3x spatial crops as well.
+    """
+
+    def __init__(self, crop_size: int = 224, num_crops: int = 3):
+        super().__init__()
+        self.crop_size = crop_size
+        if num_crops == 3:
+            self.crops_to_ext = [0, 1, 2]
+            self.flipped_crops_to_ext = []
+        elif num_crops == 1:
+            self.crops_to_ext = [1]
+            self.flipped_crops_to_ext = []
+        else:
+            raise NotImplementedError("Nothing else supported yet")
+
+    def forward(self, videos):
+        """
+        Args:
+            videos: A list of C, T, H, W videos.
+        Returns:
+            videos: A list with 3x the number of elements. Each video converted
+                to C, T, H', W' by spatial cropping.
+        """
+        assert isinstance(videos, list), "Must be a list of videos after temporal crops"
+        assert all([video.ndim == 4 for video in videos]), "Must be (C,T,H,W)"
+        res = []
+        for video in videos:
+            for spatial_idx in self.crops_to_ext:
+                res.append(uniform_crop(video, self.crop_size, spatial_idx)[0])
+            if not self.flipped_crops_to_ext:
+                continue
+            flipped_video = transforms.functional.hflip(video)
+            for spatial_idx in self.flipped_crops_to_ext:
+                res.append(uniform_crop(flipped_video, self.crop_size, spatial_idx)[0])
+        return res
+
+
+def load_and_transform_video_data(
+    video_paths,
+    device,
+    clip_duration=2,
+    clips_per_video=5,
+    sample_rate=16000,
+):
+    if video_paths is None:
+        return None
+
+    video_outputs = []
+    video_transform = transforms.Compose(
+        [
+            pv_transforms.ShortSideScale(224),
+            NormalizeVideo(
+                mean=(0.48145466, 0.4578275, 0.40821073),
+                std=(0.26862954, 0.26130258, 0.27577711),
+            ),
+        ]
+    )
+
+    clip_sampler = ConstantClipsPerVideoSampler(
+        clip_duration=clip_duration, clips_per_video=clips_per_video
+    )
+    frame_sampler = pv_transforms.UniformTemporalSubsample(num_samples=clip_duration)
+
+    for video_path in video_paths:
+        video = EncodedVideo.from_path(
+            video_path,
+            decoder="decord",
+            decode_audio=False,
+            **{"sample_rate": sample_rate},
+        )
+
+        all_clips_timepoints = get_clip_timepoints(clip_sampler, video.duration)
+
+        all_video = []
+        for clip_timepoints in all_clips_timepoints:
+            # Read the clip, get frames
+            clip = video.get_clip(clip_timepoints[0], clip_timepoints[1])
+            if clip is None:
+                raise ValueError("No clip found")
+            video_clip = frame_sampler(clip["video"])
+            video_clip = video_clip / 255.0  # since this is float, need 0-1
+
+            all_video.append(video_clip)
+
+        all_video = [video_transform(clip) for clip in all_video]
+        all_video = SpatialCrop(224, num_crops=3)(all_video)
+
+        all_video = torch.stack(all_video, dim=0)
+        video_outputs.append(all_video)
+
+    return torch.stack(video_outputs, dim=0).to(device)

ImageBind/imagebind/models/helpers.py

@@ -0,0 +1,140 @@
+#!/usr/bin/env python3
+# Portions Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import einops
+import numpy as np
+import torch
+import torch.nn as nn
+
+
+class Normalize(nn.Module):
+    def __init__(self, dim: int) -> None:
+        super().__init__()
+        self.dim = dim
+
+    def forward(self, x):
+        return torch.nn.functional.normalize(x, dim=self.dim, p=2)
+
+
+class LearnableLogitScaling(nn.Module):
+    def __init__(
+        self,
+        logit_scale_init: float = 1 / 0.07,
+        learnable: bool = True,
+        max_logit_scale: float = 100,
+    ) -> None:
+        super().__init__()
+        self.max_logit_scale = max_logit_scale
+        self.logit_scale_init = logit_scale_init
+        self.learnable = learnable
+        log_logit_scale = torch.ones([]) * np.log(self.logit_scale_init)
+        if learnable:
+            self.log_logit_scale = nn.Parameter(log_logit_scale)
+        else:
+            self.register_buffer("log_logit_scale", log_logit_scale)
+
+    def forward(self, x):
+        return torch.clip(self.log_logit_scale.exp(), max=self.max_logit_scale) * x
+
+    def extra_repr(self):
+        st = f"logit_scale_init={self.logit_scale_init},learnable={self.learnable}," \
+             f" max_logit_scale={self.max_logit_scale}"
+        return st
+
+
+class EinOpsRearrange(nn.Module):
+    def __init__(self, rearrange_expr: str, **kwargs) -> None:
+        super().__init__()
+        self.rearrange_expr = rearrange_expr
+        self.kwargs = kwargs
+
+    def forward(self, x):
+        assert isinstance(x, torch.Tensor)
+        return einops.rearrange(x, self.rearrange_expr, **self.kwargs)
+
+
+class VerboseNNModule(nn.Module):
+    """
+    Wrapper around nn.Module that prints registered buffers and parameter names.
+    """
+
+    @staticmethod
+    def get_readable_tensor_repr(name: str, tensor: torch.Tensor) -> str:
+        st = (
+            "("
+            + name
+            + "): "
+            + "tensor("
+            + str(tuple(tensor[1].shape))
+            + ", requires_grad="
+            + str(tensor[1].requires_grad)
+            + ")\n"
+        )
+        return st
+
+    def extra_repr(self) -> str:
+        named_modules = set()
+        for p in self.named_modules():
+            named_modules.update([p[0]])
+        named_modules = list(named_modules)
+
+        string_repr = ""
+        for p in self.named_parameters():
+            name = p[0].split(".")[0]
+            if name not in named_modules:
+                string_repr += self.get_readable_tensor_repr(name, p)
+
+        for p in self.named_buffers():
+            name = p[0].split(".")[0]
+            string_repr += self.get_readable_tensor_repr(name, p)
+
+        return string_repr
+
+
+def cast_if_src_dtype(
+    tensor: torch.Tensor, src_dtype: torch.dtype, tgt_dtype: torch.dtype
+):
+    updated = False
+    if tensor.dtype == src_dtype:
+        tensor = tensor.to(dtype=tgt_dtype)
+        updated = True
+    return tensor, updated
+
+
+class QuickGELU(nn.Module):
+    # From https://github.com/openai/CLIP/blob/d50d76daa670286dd6cacf3bcd80b5e4823fc8e1/clip/model.py#L166
+    def forward(self, x: torch.Tensor):
+        return x * torch.sigmoid(1.702 * x)
+
+
+class SelectElement(nn.Module):
+    def __init__(self, index) -> None:
+        super().__init__()
+        self.index = index
+
+    def forward(self, x):
+        assert x.ndim >= 3
+        return x[:, self.index, ...]
+
+
+class SelectEOSAndProject(nn.Module):
+    """
+    Text Pooling used in OpenCLIP
+    """
+
+    def __init__(self, proj: nn.Module) -> None:
+        super().__init__()
+        self.proj = proj
+
+    def forward(self, x, seq_len):
+        assert x.ndim == 3
+        # x is of shape B x L x D
+        # take features from the eot embedding (eot_token is the highest number in each sequence)
+        x = x[torch.arange(x.shape[0]), seq_len]
+        x = self.proj(x)
+        return x

ImageBind/imagebind/models/imagebind_model.py

@@ -0,0 +1,506 @@
+#!/usr/bin/env python3
+# Portions Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import os
+from functools import partial
+from types import SimpleNamespace
+
+import torch
+import torch.nn as nn
+
+from imagebind.models.helpers import (EinOpsRearrange, LearnableLogitScaling, Normalize,
+                            SelectElement, SelectEOSAndProject)
+from imagebind.models.multimodal_preprocessors import (AudioPreprocessor,
+                                             IMUPreprocessor, PadIm2Video,
+                                             PatchEmbedGeneric,
+                                             RGBDTPreprocessor,
+                                             SpatioTemporalPosEmbeddingHelper,
+                                             TextPreprocessor,
+                                             ThermalPreprocessor)
+from imagebind.models.transformer import MultiheadAttention, SimpleTransformer
+
+ModalityType = SimpleNamespace(
+    VISION="vision",
+    TEXT="text",
+    AUDIO="audio",
+    THERMAL="thermal",
+    DEPTH="depth",
+    IMU="imu",
+)
+
+
+class ImageBindModel(nn.Module):
+    def __init__(
+        self,
+        video_frames=2,
+        kernel_size=(2, 14, 14),
+        audio_kernel_size=16,
+        audio_stride=10,
+        out_embed_dim=768,
+        vision_embed_dim=1024,
+        vision_num_blocks=24,
+        vision_num_heads=16,
+        audio_embed_dim=768,
+        audio_num_blocks=12,
+        audio_num_heads=12,
+        audio_num_mel_bins=128,
+        audio_target_len=204,
+        audio_drop_path=0.1,
+        text_embed_dim=768,
+        text_num_blocks=12,
+        text_num_heads=12,
+        depth_embed_dim=384,
+        depth_kernel_size=16,
+        depth_num_blocks=12,
+        depth_num_heads=8,
+        depth_drop_path=0.0,
+        thermal_embed_dim=768,
+        thermal_kernel_size=16,
+        thermal_num_blocks=12,
+        thermal_num_heads=12,
+        thermal_drop_path=0.0,
+        imu_embed_dim=512,
+        imu_kernel_size=8,
+        imu_num_blocks=6,
+        imu_num_heads=8,
+        imu_drop_path=0.7,
+    ):
+        super().__init__()
+
+        self.modality_preprocessors = self._create_modality_preprocessors(
+            video_frames,
+            vision_embed_dim,
+            kernel_size,
+            text_embed_dim,
+            audio_embed_dim,
+            audio_kernel_size,
+            audio_stride,
+            audio_num_mel_bins,
+            audio_target_len,
+            depth_embed_dim,
+            depth_kernel_size,
+            thermal_embed_dim,
+            thermal_kernel_size,
+            imu_embed_dim,
+        )
+
+        self.modality_trunks = self._create_modality_trunks(
+            vision_embed_dim,
+            vision_num_blocks,
+            vision_num_heads,
+            text_embed_dim,
+            text_num_blocks,
+            text_num_heads,
+            audio_embed_dim,
+            audio_num_blocks,
+            audio_num_heads,
+            audio_drop_path,
+            depth_embed_dim,
+            depth_num_blocks,
+            depth_num_heads,
+            depth_drop_path,
+            thermal_embed_dim,
+            thermal_num_blocks,
+            thermal_num_heads,
+            thermal_drop_path,
+            imu_embed_dim,
+            imu_num_blocks,
+            imu_num_heads,
+            imu_drop_path,
+        )
+
+        self.modality_heads = self._create_modality_heads(
+            out_embed_dim,
+            vision_embed_dim,
+            text_embed_dim,
+            audio_embed_dim,
+            depth_embed_dim,
+            thermal_embed_dim,
+            imu_embed_dim,
+        )
+
+        self.modality_postprocessors = self._create_modality_postprocessors(
+            out_embed_dim
+        )
+
+    def _create_modality_preprocessors(
+        self,
+        video_frames=2,
+        vision_embed_dim=1024,
+        kernel_size=(2, 14, 14),
+        text_embed_dim=768,
+        audio_embed_dim=768,
+        audio_kernel_size=16,
+        audio_stride=10,
+        audio_num_mel_bins=128,
+        audio_target_len=204,
+        depth_embed_dim=768,
+        depth_kernel_size=16,
+        thermal_embed_dim=768,
+        thermal_kernel_size=16,
+        imu_embed_dim=512,
+    ):
+        rgbt_stem = PatchEmbedGeneric(
+            proj_stem=[
+                PadIm2Video(pad_type="repeat", ntimes=2),
+                nn.Conv3d(
+                    in_channels=3,
+                    kernel_size=kernel_size,
+                    out_channels=vision_embed_dim,
+                    stride=kernel_size,
+                    bias=False,
+                ),
+            ]
+        )
+        rgbt_preprocessor = RGBDTPreprocessor(
+            img_size=[3, video_frames, 224, 224],
+            num_cls_tokens=1,
+            pos_embed_fn=partial(SpatioTemporalPosEmbeddingHelper, learnable=True),
+            rgbt_stem=rgbt_stem,
+            depth_stem=None,
+        )
+
+        text_preprocessor = TextPreprocessor(
+            context_length=77,
+            vocab_size=49408,
+            embed_dim=text_embed_dim,
+            causal_masking=True,
+        )
+
+        audio_stem = PatchEmbedGeneric(
+            proj_stem=[
+                nn.Conv2d(
+                    in_channels=1,
+                    kernel_size=audio_kernel_size,
+                    stride=audio_stride,
+                    out_channels=audio_embed_dim,
+                    bias=False,
+                ),
+            ],
+            norm_layer=nn.LayerNorm(normalized_shape=audio_embed_dim),
+        )
+        audio_preprocessor = AudioPreprocessor(
+            img_size=[1, audio_num_mel_bins, audio_target_len],
+            num_cls_tokens=1,
+            pos_embed_fn=partial(SpatioTemporalPosEmbeddingHelper, learnable=True),
+            audio_stem=audio_stem,
+        )
+
+        depth_stem = PatchEmbedGeneric(
+            [
+                nn.Conv2d(
+                    kernel_size=depth_kernel_size,
+                    in_channels=1,
+                    out_channels=depth_embed_dim,
+                    stride=depth_kernel_size,
+                    bias=False,
+                ),
+            ],
+            norm_layer=nn.LayerNorm(normalized_shape=depth_embed_dim),
+        )
+
+        depth_preprocessor = RGBDTPreprocessor(
+            img_size=[1, 224, 224],
+            num_cls_tokens=1,
+            pos_embed_fn=partial(SpatioTemporalPosEmbeddingHelper, learnable=True),
+            rgbt_stem=None,
+            depth_stem=depth_stem,
+        )
+
+        thermal_stem = PatchEmbedGeneric(
+            [
+                nn.Conv2d(
+                    kernel_size=thermal_kernel_size,
+                    in_channels=1,
+                    out_channels=thermal_embed_dim,
+                    stride=thermal_kernel_size,
+                    bias=False,
+                ),
+            ],
+            norm_layer=nn.LayerNorm(normalized_shape=thermal_embed_dim),
+        )
+        thermal_preprocessor = ThermalPreprocessor(
+            img_size=[1, 224, 224],
+            num_cls_tokens=1,
+            pos_embed_fn=partial(SpatioTemporalPosEmbeddingHelper, learnable=True),
+            thermal_stem=thermal_stem,
+        )
+
+        imu_stem = PatchEmbedGeneric(
+            [
+                nn.Linear(
+                    in_features=48,
+                    out_features=imu_embed_dim,
+                    bias=False,
+                ),
+            ],
+            norm_layer=nn.LayerNorm(normalized_shape=imu_embed_dim),
+        )
+
+        imu_preprocessor = IMUPreprocessor(
+            img_size=[6, 2000],
+            num_cls_tokens=1,
+            kernel_size=8,
+            embed_dim=imu_embed_dim,
+            pos_embed_fn=partial(SpatioTemporalPosEmbeddingHelper, learnable=True),
+            imu_stem=imu_stem,
+        )
+
+        modality_preprocessors = {
+            ModalityType.VISION: rgbt_preprocessor,
+            ModalityType.TEXT: text_preprocessor,
+            ModalityType.AUDIO: audio_preprocessor,
+            ModalityType.DEPTH: depth_preprocessor,
+            ModalityType.THERMAL: thermal_preprocessor,
+            ModalityType.IMU: imu_preprocessor,
+        }
+
+        return nn.ModuleDict(modality_preprocessors)
+
+    def _create_modality_trunks(
+        self,
+        vision_embed_dim=1024,
+        vision_num_blocks=24,
+        vision_num_heads=16,
+        text_embed_dim=768,
+        text_num_blocks=12,
+        text_num_heads=12,
+        audio_embed_dim=768,
+        audio_num_blocks=12,
+        audio_num_heads=12,
+        audio_drop_path=0.0,
+        depth_embed_dim=768,
+        depth_num_blocks=12,
+        depth_num_heads=12,
+        depth_drop_path=0.0,
+        thermal_embed_dim=768,
+        thermal_num_blocks=12,
+        thermal_num_heads=12,
+        thermal_drop_path=0.0,
+        imu_embed_dim=512,
+        imu_num_blocks=6,
+        imu_num_heads=8,
+        imu_drop_path=0.7,
+    ):
+        def instantiate_trunk(
+            embed_dim, num_blocks, num_heads, pre_transformer_ln, add_bias_kv, drop_path
+        ):
+            return SimpleTransformer(
+                embed_dim=embed_dim,
+                num_blocks=num_blocks,
+                ffn_dropout_rate=0.0,
+                drop_path_rate=drop_path,
+                attn_target=partial(
+                    MultiheadAttention,
+                    embed_dim=embed_dim,
+                    num_heads=num_heads,
+                    bias=True,
+                    add_bias_kv=add_bias_kv,
+                ),
+                pre_transformer_layer=nn.Sequential(
+                    nn.LayerNorm(embed_dim, eps=1e-6)
+                    if pre_transformer_ln
+                    else nn.Identity(),
+                    EinOpsRearrange("b l d -> l b d"),
+                ),
+                post_transformer_layer=EinOpsRearrange("l b d -> b l d"),
+            )
+
+        modality_trunks = {}
+        modality_trunks[ModalityType.VISION] = instantiate_trunk(
+            vision_embed_dim,
+            vision_num_blocks,
+            vision_num_heads,
+            pre_transformer_ln=True,
+            add_bias_kv=False,
+            drop_path=0.0,
+        )
+        modality_trunks[ModalityType.TEXT] = instantiate_trunk(
+            text_embed_dim,
+            text_num_blocks,
+            text_num_heads,
+            pre_transformer_ln=False,
+            add_bias_kv=False,
+            drop_path=0.0,
+        )
+        modality_trunks[ModalityType.AUDIO] = instantiate_trunk(
+            audio_embed_dim,
+            audio_num_blocks,
+            audio_num_heads,
+            pre_transformer_ln=False,
+            add_bias_kv=True,
+            drop_path=audio_drop_path,
+        )
+        modality_trunks[ModalityType.DEPTH] = instantiate_trunk(
+            depth_embed_dim,
+            depth_num_blocks,
+            depth_num_heads,
+            pre_transformer_ln=False,
+            add_bias_kv=True,
+            drop_path=depth_drop_path,
+        )
+        modality_trunks[ModalityType.THERMAL] = instantiate_trunk(
+            thermal_embed_dim,
+            thermal_num_blocks,
+            thermal_num_heads,
+            pre_transformer_ln=False,
+            add_bias_kv=True,
+            drop_path=thermal_drop_path,
+        )
+        modality_trunks[ModalityType.IMU] = instantiate_trunk(
+            imu_embed_dim,
+            imu_num_blocks,
+            imu_num_heads,
+            pre_transformer_ln=False,
+            add_bias_kv=True,
+            drop_path=imu_drop_path,
+        )
+
+        return nn.ModuleDict(modality_trunks)
+
+    def _create_modality_heads(
+        self,
+        out_embed_dim,
+        vision_embed_dim,
+        text_embed_dim,
+        audio_embed_dim,
+        depth_embed_dim,
+        thermal_embed_dim,
+        imu_embed_dim,
+    ):
+        modality_heads = {}
+
+        modality_heads[ModalityType.VISION] = nn.Sequential(
+            nn.LayerNorm(normalized_shape=vision_embed_dim, eps=1e-6),
+            SelectElement(index=0),
+            nn.Linear(vision_embed_dim, out_embed_dim, bias=False),
+        )
+
+        modality_heads[ModalityType.TEXT] = SelectEOSAndProject(
+            proj=nn.Sequential(
+                nn.LayerNorm(normalized_shape=text_embed_dim, eps=1e-6),
+                nn.Linear(text_embed_dim, out_embed_dim, bias=False),
+            )
+        )
+
+        modality_heads[ModalityType.AUDIO] = nn.Sequential(
+            nn.LayerNorm(normalized_shape=audio_embed_dim, eps=1e-6),
+            SelectElement(index=0),
+            nn.Linear(audio_embed_dim, out_embed_dim, bias=False),
+        )
+
+        modality_heads[ModalityType.DEPTH] = nn.Sequential(
+            nn.LayerNorm(normalized_shape=depth_embed_dim, eps=1e-6),
+            SelectElement(index=0),
+            nn.Linear(depth_embed_dim, out_embed_dim, bias=False),
+        )
+
+        modality_heads[ModalityType.THERMAL] = nn.Sequential(
+            nn.LayerNorm(normalized_shape=thermal_embed_dim, eps=1e-6),
+            SelectElement(index=0),
+            nn.Linear(thermal_embed_dim, out_embed_dim, bias=False),
+        )
+
+        modality_heads[ModalityType.IMU] = nn.Sequential(
+            nn.LayerNorm(normalized_shape=imu_embed_dim, eps=1e-6),
+            SelectElement(index=0),
+            nn.Dropout(p=0.5),
+            nn.Linear(imu_embed_dim, out_embed_dim, bias=False),
+        )
+
+        return nn.ModuleDict(modality_heads)
+
+    def _create_modality_postprocessors(self, out_embed_dim):
+        modality_postprocessors = {}
+
+        modality_postprocessors[ModalityType.VISION] = Normalize(dim=-1)
+        modality_postprocessors[ModalityType.TEXT] = nn.Sequential(
+            Normalize(dim=-1), LearnableLogitScaling(learnable=True)
+        )
+        modality_postprocessors[ModalityType.AUDIO] = nn.Sequential(
+            Normalize(dim=-1),
+            LearnableLogitScaling(logit_scale_init=20.0, learnable=False),
+        )
+        modality_postprocessors[ModalityType.DEPTH] = nn.Sequential(
+            Normalize(dim=-1),
+            LearnableLogitScaling(logit_scale_init=5.0, learnable=False),
+        )
+        modality_postprocessors[ModalityType.THERMAL] = nn.Sequential(
+            Normalize(dim=-1),
+            LearnableLogitScaling(logit_scale_init=10.0, learnable=False),
+        )
+        modality_postprocessors[ModalityType.IMU] = nn.Sequential(
+            Normalize(dim=-1),
+            LearnableLogitScaling(logit_scale_init=5.0, learnable=False),
+        )
+
+        return nn.ModuleDict(modality_postprocessors)
+
+    def forward(self, inputs):
+        outputs = {}
+        for modality_key, modality_value in inputs.items():
+            reduce_list = (
+                modality_value.ndim >= 5
+            )  # Audio and Video inputs consist of multiple clips
+            if reduce_list:
+                B, S = modality_value.shape[:2]
+                modality_value = modality_value.reshape(
+                    B * S, *modality_value.shape[2:]
+                )
+
+            if modality_value is not None:
+                modality_value = self.modality_preprocessors[modality_key](
+                    **{modality_key: modality_value}
+                )
+                trunk_inputs = modality_value["trunk"]
+                head_inputs = modality_value["head"]
+                modality_value = self.modality_trunks[modality_key](**trunk_inputs)
+                modality_value = self.modality_heads[modality_key](
+                    modality_value, **head_inputs
+                )
+                modality_value = self.modality_postprocessors[modality_key](
+                    modality_value
+                )
+
+                if reduce_list:
+                    modality_value = modality_value.reshape(B, S, -1)
+                    modality_value = modality_value.mean(dim=1)
+
+                outputs[modality_key] = modality_value
+
+        return outputs
+
+
+def imagebind_huge(pretrained=False):
+    model = ImageBindModel(
+        vision_embed_dim=1280,
+        vision_num_blocks=32,
+        vision_num_heads=16,
+        text_embed_dim=1024,
+        text_num_blocks=24,
+        text_num_heads=16,
+        out_embed_dim=1024,
+        audio_drop_path=0.1,
+        imu_drop_path=0.7,
+    )
+
+    if pretrained:
+        if not os.path.exists(".checkpoints/imagebind_huge.pth"):
+            print(
+                "Downloading imagebind weights to .checkpoints/imagebind_huge.pth ..."
+            )
+            os.makedirs(".checkpoints", exist_ok=True)
+            torch.hub.download_url_to_file(
+                "https://dl.fbaipublicfiles.com/imagebind/imagebind_huge.pth",
+                ".checkpoints/imagebind_huge.pth",
+                progress=True,
+            )
+
+        model.load_state_dict(torch.load(".checkpoints/imagebind_huge.pth"))
+
+    return model

ImageBind/imagebind/models/multimodal_preprocessors.py

@@ -0,0 +1,685 @@
+#!/usr/bin/env python3
+# Portions Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import gzip
+import html
+import io
+import math
+from functools import lru_cache
+from typing import Callable, List, Optional, Tuple
+
+import ftfy
+import numpy as np
+import regex as re
+import torch
+import torch.nn as nn
+from iopath.common.file_io import g_pathmgr
+from timm.models.layers import trunc_normal_
+
+from imagebind.models.helpers import VerboseNNModule, cast_if_src_dtype
+
+
+def get_sinusoid_encoding_table(n_position, d_hid):
+    """Sinusoid position encoding table"""
+
+    # TODO: make it with torch instead of numpy
+    def get_position_angle_vec(position):
+        return [
+            position / np.power(10000, 2 * (hid_j // 2) / d_hid)
+            for hid_j in range(d_hid)
+        ]
+
+    sinusoid_table = np.array(
+        [get_position_angle_vec(pos_i) for pos_i in range(n_position)]
+    )
+    sinusoid_table[:, 0::2] = np.sin(sinusoid_table[:, 0::2])  # dim 2i
+    sinusoid_table[:, 1::2] = np.cos(sinusoid_table[:, 1::2])  # dim 2i+1
+
+    return torch.FloatTensor(sinusoid_table).unsqueeze(0)
+
+
+def interpolate_pos_encoding_2d(target_spatial_size, pos_embed):
+    N = pos_embed.shape[1]
+    if N == target_spatial_size:
+        return pos_embed
+    dim = pos_embed.shape[-1]
+    # nn.functional.interpolate doesn't work with bfloat16 so we cast to float32
+    pos_embed, updated = cast_if_src_dtype(pos_embed, torch.bfloat16, torch.float32)
+    pos_embed = nn.functional.interpolate(
+        pos_embed.reshape(1, int(math.sqrt(N)), int(math.sqrt(N)), dim).permute(
+            0, 3, 1, 2
+        ),
+        scale_factor=math.sqrt(target_spatial_size / N),
+        mode="bicubic",
+    )
+    if updated:
+        pos_embed, _ = cast_if_src_dtype(pos_embed, torch.float32, torch.bfloat16)
+    pos_embed = pos_embed.permute(0, 2, 3, 1).view(1, -1, dim)
+    return pos_embed
+
+
+def interpolate_pos_encoding(
+    npatch_per_img,
+    pos_embed,
+    patches_layout,
+    input_shape=None,
+    first_patch_idx=1,
+):
+    assert first_patch_idx == 0 or first_patch_idx == 1, "there is 1 CLS token or none"
+    N = pos_embed.shape[1] - first_patch_idx  # since it's 1 if cls_token exists
+    if npatch_per_img == N:
+        return pos_embed
+
+    assert (
+        patches_layout[-1] == patches_layout[-2]
+    ), "Interpolation of pos embed not supported for non-square layouts"
+
+    class_emb = pos_embed[:, :first_patch_idx]
+    pos_embed = pos_embed[:, first_patch_idx:]
+
+    if input_shape is None or patches_layout[0] == 1:
+        # simple 2D pos embedding, no temporal component
+        pos_embed = interpolate_pos_encoding_2d(npatch_per_img, pos_embed)
+    elif patches_layout[0] > 1:
+        # pos embed has a temporal component
+        assert len(input_shape) == 4, "temporal interpolation not supported"
+        # we only support 2D interpolation in this case
+        num_frames = patches_layout[0]
+        num_spatial_tokens = patches_layout[1] * patches_layout[2]
+        pos_embed = pos_embed.view(1, num_frames, num_spatial_tokens, -1)
+        # interpolate embedding for zeroth frame
+        pos_embed = interpolate_pos_encoding_2d(
+            npatch_per_img, pos_embed[0, 0, ...].unsqueeze(0)
+        )
+    else:
+        raise ValueError("This type of interpolation isn't implemented")
+
+    return torch.cat((class_emb, pos_embed), dim=1)
+
+
+def _get_pos_embedding(
+    npatch_per_img,
+    pos_embed,
+    patches_layout,
+    input_shape,
+    first_patch_idx=1,
+):
+    pos_embed = interpolate_pos_encoding(
+        npatch_per_img,
+        pos_embed,
+        patches_layout,
+        input_shape=input_shape,
+        first_patch_idx=first_patch_idx,
+    )
+    return pos_embed
+
+
+class PatchEmbedGeneric(nn.Module):
+    """
+    PatchEmbed from Hydra
+    """
+
+    def __init__(self, proj_stem, norm_layer: Optional[nn.Module] = None):
+        super().__init__()
+
+        if len(proj_stem) > 1:
+            self.proj = nn.Sequential(*proj_stem)
+        else:
+            # Special case to be able to load pre-trained models that were
+            # trained with a standard stem
+            self.proj = proj_stem[0]
+        self.norm_layer = norm_layer
+
+    def get_patch_layout(self, img_size):
+        with torch.no_grad():
+            dummy_img = torch.zeros(
+                [
+                    1,
+                ]
+                + img_size
+            )
+            dummy_out = self.proj(dummy_img)
+        embed_dim = dummy_out.shape[1]
+        patches_layout = tuple(dummy_out.shape[2:])
+        num_patches = np.prod(patches_layout)
+        return patches_layout, num_patches, embed_dim
+
+    def forward(self, x):
+        x = self.proj(x)
+        # B C (T) H W -> B (T)HW C
+        x = x.flatten(2).transpose(1, 2)
+        if self.norm_layer is not None:
+            x = self.norm_layer(x)
+        return x
+
+
+class SpatioTemporalPosEmbeddingHelper(VerboseNNModule):
+    def __init__(
+        self,
+        patches_layout: List,
+        num_patches: int,
+        num_cls_tokens: int,
+        embed_dim: int,
+        learnable: bool,
+    ) -> None:
+        super().__init__()
+        self.num_cls_tokens = num_cls_tokens
+        self.patches_layout = patches_layout
+        self.num_patches = num_patches
+        self.num_tokens = num_cls_tokens + num_patches
+        self.learnable = learnable
+        if self.learnable:
+            self.pos_embed = nn.Parameter(torch.zeros(1, self.num_tokens, embed_dim))
+            trunc_normal_(self.pos_embed, std=0.02)
+        else:
+            self.register_buffer(
+                "pos_embed", get_sinusoid_encoding_table(self.num_tokens, embed_dim)
+            )
+
+    def get_pos_embedding(self, vision_input, all_vision_tokens):
+        input_shape = vision_input.shape
+        pos_embed = _get_pos_embedding(
+            all_vision_tokens.size(1) - self.num_cls_tokens,
+            pos_embed=self.pos_embed,
+            patches_layout=self.patches_layout,
+            input_shape=input_shape,
+            first_patch_idx=self.num_cls_tokens,
+        )
+        return pos_embed
+
+
+class RGBDTPreprocessor(VerboseNNModule):
+    def __init__(
+        self,
+        rgbt_stem: PatchEmbedGeneric,
+        depth_stem: Optional[PatchEmbedGeneric],
+        img_size: Tuple = (3, 224, 224),
+        num_cls_tokens: int = 1,
+        pos_embed_fn: Optional[Callable] = None,
+        use_type_embed: bool = False,
+        init_param_style: str = "openclip",
+    ) -> None:
+        super().__init__()
+        stem = rgbt_stem if rgbt_stem is not None else depth_stem
+        (
+            self.patches_layout,
+            self.num_patches,
+            self.embed_dim,
+        ) = stem.get_patch_layout(img_size)
+        self.rgbt_stem = rgbt_stem
+        self.depth_stem = depth_stem
+        self.use_pos_embed = pos_embed_fn is not None
+        self.use_type_embed = use_type_embed
+        self.num_cls_tokens = num_cls_tokens
+
+        if self.use_pos_embed:
+            self.pos_embedding_helper = pos_embed_fn(
+                patches_layout=self.patches_layout,
+                num_cls_tokens=num_cls_tokens,
+                num_patches=self.num_patches,
+                embed_dim=self.embed_dim,
+            )
+        if self.num_cls_tokens > 0:
+            self.cls_token = nn.Parameter(
+                torch.zeros(1, self.num_cls_tokens, self.embed_dim)
+            )
+        if self.use_type_embed:
+            self.type_embed = nn.Parameter(torch.zeros(1, 1, self.embed_dim))
+
+        self.init_parameters(init_param_style)
+
+    @torch.no_grad()
+    def init_parameters(self, init_param_style):
+        if init_param_style == "openclip":
+            # OpenCLIP style initialization
+            scale = self.embed_dim**-0.5
+            if self.use_pos_embed:
+                nn.init.normal_(self.pos_embedding_helper.pos_embed)
+                self.pos_embedding_helper.pos_embed *= scale
+
+            if self.num_cls_tokens > 0:
+                nn.init.normal_(self.cls_token)
+                self.cls_token *= scale
+        elif init_param_style == "vit":
+            self.cls_token.data.fill_(0)
+        else:
+            raise ValueError(f"Unknown init {init_param_style}")
+
+        if self.use_type_embed:
+            nn.init.normal_(self.type_embed)
+
+    def tokenize_input_and_cls_pos(self, input, stem, mask):
+        # tokens is of shape B x L x D
+        tokens = stem(input)
+        assert tokens.ndim == 3
+        assert tokens.shape[2] == self.embed_dim
+        B = tokens.shape[0]
+        if self.num_cls_tokens > 0:
+            class_tokens = self.cls_token.expand(
+                B, -1, -1
+            )  # stole class_tokens impl from Phil Wang, thanks
+            tokens = torch.cat((class_tokens, tokens), dim=1)
+        if self.use_pos_embed:
+            pos_embed = self.pos_embedding_helper.get_pos_embedding(input, tokens)
+            tokens = tokens + pos_embed
+        if self.use_type_embed:
+            tokens = tokens + self.type_embed.expand(B, -1, -1)
+        return tokens
+
+    def forward(self, vision=None, depth=None, patch_mask=None):
+        if patch_mask is not None:
+            raise NotImplementedError()
+
+        if vision is not None:
+            vision_tokens = self.tokenize_input_and_cls_pos(
+                vision, self.rgbt_stem, patch_mask
+            )
+
+        if depth is not None:
+            depth_tokens = self.tokenize_input_and_cls_pos(
+                depth, self.depth_stem, patch_mask
+            )
+
+        # aggregate tokens
+        if vision is not None and depth is not None:
+            final_tokens = vision_tokens + depth_tokens
+        else:
+            final_tokens = vision_tokens if vision is not None else depth_tokens
+        return_dict = {
+            "trunk": {
+                "tokens": final_tokens,
+            },
+            "head": {},
+        }
+        return return_dict
+
+
+class AudioPreprocessor(RGBDTPreprocessor):
+    def __init__(self, audio_stem: PatchEmbedGeneric, **kwargs) -> None:
+        super().__init__(rgbt_stem=audio_stem, depth_stem=None, **kwargs)
+
+    def forward(self, audio=None):
+        return super().forward(vision=audio)
+
+
+class ThermalPreprocessor(RGBDTPreprocessor):
+    def __init__(self, thermal_stem: PatchEmbedGeneric, **kwargs) -> None:
+        super().__init__(rgbt_stem=thermal_stem, depth_stem=None, **kwargs)
+
+    def forward(self, thermal=None):
+        return super().forward(vision=thermal)
+
+
+def build_causal_attention_mask(context_length):
+    # lazily create causal attention mask, with full attention between the vision tokens
+    # pytorch uses additive attention mask; fill with -inf
+    mask = torch.empty(context_length, context_length, requires_grad=False)
+    mask.fill_(float("-inf"))
+    mask.triu_(1)  # zero out the lower diagonal
+    return mask
+
+
+class TextPreprocessor(VerboseNNModule):
+    def __init__(
+        self,
+        vocab_size: int,
+        context_length: int,
+        embed_dim: int,
+        causal_masking: bool,
+        supply_seq_len_to_head: bool = True,
+        num_cls_tokens: int = 0,
+        init_param_style: str = "openclip",
+    ) -> None:
+        super().__init__()
+        self.vocab_size = vocab_size
+        self.context_length = context_length
+        self.token_embedding = nn.Embedding(vocab_size, embed_dim)
+        self.pos_embed = nn.Parameter(
+            torch.empty(1, self.context_length + num_cls_tokens, embed_dim)
+        )
+        self.causal_masking = causal_masking
+        if self.causal_masking:
+            mask = build_causal_attention_mask(self.context_length)
+            # register the mask as a buffer so it can be moved to the right device
+            self.register_buffer("mask", mask)
+
+        self.supply_seq_len_to_head = supply_seq_len_to_head
+        self.num_cls_tokens = num_cls_tokens
+        self.embed_dim = embed_dim
+        if num_cls_tokens > 0:
+            assert self.causal_masking is False, "Masking + CLS token isn't implemented"
+            self.cls_token = nn.Parameter(
+                torch.zeros(1, self.num_cls_tokens, embed_dim)
+            )
+
+        self.init_parameters(init_param_style)
+
+    @torch.no_grad()
+    def init_parameters(self, init_param_style="openclip"):
+        # OpenCLIP style initialization
+        nn.init.normal_(self.token_embedding.weight, std=0.02)
+        nn.init.normal_(self.pos_embed, std=0.01)
+
+        if init_param_style == "openclip":
+            # OpenCLIP style initialization
+            scale = self.embed_dim**-0.5
+            if self.num_cls_tokens > 0:
+                nn.init.normal_(self.cls_token)
+                self.cls_token *= scale
+        elif init_param_style == "vit":
+            self.cls_token.data.fill_(0)
+        else:
+            raise ValueError(f"Unknown init {init_param_style}")
+
+    def forward(self, text):
+        # text tokens are of shape B x L x D
+        text_tokens = self.token_embedding(text)
+        # concat CLS tokens if any
+        if self.num_cls_tokens > 0:
+            B = text_tokens.shape[0]
+            class_tokens = self.cls_token.expand(
+                B, -1, -1
+            )  # stole class_tokens impl from Phil Wang, thanks
+            text_tokens = torch.cat((class_tokens, text_tokens), dim=1)
+        text_tokens = text_tokens + self.pos_embed
+        return_dict = {
+            "trunk": {
+                "tokens": text_tokens,
+            },
+            "head": {},
+        }
+        # Compute sequence length after adding CLS tokens
+        if self.supply_seq_len_to_head:
+            text_lengths = text.argmax(dim=-1)
+            return_dict["head"] = {
+                "seq_len": text_lengths,
+            }
+        if self.causal_masking:
+            return_dict["trunk"].update({"attn_mask": self.mask})
+        return return_dict
+
+
+class Im2Video(nn.Module):
+    """Convert an image into a trivial video."""
+
+    def __init__(self, time_dim=2):
+        super().__init__()
+        self.time_dim = time_dim
+
+    def forward(self, x):
+        if x.ndim == 4:
+            # B, C, H, W -> B, C, T, H, W
+            return x.unsqueeze(self.time_dim)
+        elif x.ndim == 5:
+            return x
+        else:
+            raise ValueError(f"Dimension incorrect {x.shape}")
+
+
+class PadIm2Video(Im2Video):
+    def __init__(self, ntimes, pad_type, time_dim=2):
+        super().__init__(time_dim=time_dim)
+        assert ntimes > 0
+        assert pad_type in ["zero", "repeat"]
+        self.ntimes = ntimes
+        self.pad_type = pad_type
+
+    def forward(self, x):
+        x = super().forward(x)
+        if x.shape[self.time_dim] == 1:
+            if self.pad_type == "repeat":
+                new_shape = [1] * len(x.shape)
+                new_shape[self.time_dim] = self.ntimes
+                x = x.repeat(new_shape)
+            elif self.pad_type == "zero":
+                padarg = [0, 0] * len(x.shape)
+                padarg[2 * self.time_dim + 1] = self.ntimes - x.shape[self.time_dim]
+                x = nn.functional.pad(x, padarg)
+        return x
+
+
+# Modified from github.com/openai/CLIP
+@lru_cache()
+def bytes_to_unicode():
+    """
+    Returns list of utf-8 byte and a corresponding list of unicode strings.
+    The reversible bpe codes work on unicode strings.
+    This means you need a large # of unicode characters in your vocab if you want to avoid UNKs.
+    When you're at something like a 10B token dataset you end up needing around 5K for decent coverage.
+    This is a signficant percentage of your normal, say, 32K bpe vocab.
+    To avoid that, we want lookup tables between utf-8 bytes and unicode strings.
+    And avoids mapping to whitespace/control characters the bpe code barfs on.
+    """
+    bs = (
+        list(range(ord("!"), ord("~") + 1))
+        + list(range(ord("¡"), ord("¬") + 1))
+        + list(range(ord("®"), ord("ÿ") + 1))
+    )
+    cs = bs[:]
+    n = 0
+    for b in range(2**8):
+        if b not in bs:
+            bs.append(b)
+            cs.append(2**8 + n)
+            n += 1
+    cs = [chr(n) for n in cs]
+    return dict(zip(bs, cs))
+
+
+def get_pairs(word):
+    """Return set of symbol pairs in a word.
+    Word is represented as tuple of symbols (symbols being variable-length strings).
+    """
+    pairs = set()
+    prev_char = word[0]
+    for char in word[1:]:
+        pairs.add((prev_char, char))
+        prev_char = char
+    return pairs
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r"\s+", " ", text)
+    text = text.strip()
+    return text
+
+
+class SimpleTokenizer(object):
+    def __init__(self, bpe_path: str, context_length=77):
+        self.byte_encoder = bytes_to_unicode()
+        self.byte_decoder = {v: k for k, v in self.byte_encoder.items()}
+
+        with g_pathmgr.open(bpe_path, "rb") as fh:
+            bpe_bytes = io.BytesIO(fh.read())
+            merges: List[str] = gzip.open(bpe_bytes).read().decode("utf-8").split("\n")
+        merges = merges[1 : 49152 - 256 - 2 + 1]
+        merges: List[Tuple[str, ...]] = [tuple(merge.split()) for merge in merges]
+        vocab = list(bytes_to_unicode().values())
+        vocab = vocab + [v + "</w>" for v in vocab]
+        for merge in merges:
+            vocab.append("".join(merge))
+        vocab.extend(["<|startoftext|>", "<|endoftext|>"])
+        self.encoder = dict(zip(vocab, range(len(vocab))))
+        self.decoder = {v: k for k, v in self.encoder.items()}
+        self.bpe_ranks = dict(zip(merges, range(len(merges))))
+        self.cache = {
+            "<|startoftext|>": "<|startoftext|>",
+            "<|endoftext|>": "<|endoftext|>",
+        }
+        self.pat = re.compile(
+            r"""<\|startoftext\|>|<\|endoftext\|>|'s|'t|'re|'ve|'m|'ll|'d|[\p{L}]+|[\p{N}]|[^\s\p{L}\p{N}]+""",
+            re.IGNORECASE,
+        )
+        self.context_length = context_length
+
+    def bpe(self, token):
+        if token in self.cache:
+            return self.cache[token]
+        word = tuple(token[:-1]) + (token[-1] + "</w>",)
+        pairs = get_pairs(word)
+
+        if not pairs:
+            return token + "</w>"
+
+        while True:
+            bigram = min(pairs, key=lambda pair: self.bpe_ranks.get(pair, float("inf")))
+            if bigram not in self.bpe_ranks:
+                break
+            first, second = bigram
+            new_word = []
+            i = 0
+            while i < len(word):
+                try:
+                    j = word.index(first, i)
+                    new_word.extend(word[i:j])
+                    i = j
+                except:
+                    new_word.extend(word[i:])
+                    break
+
+                if word[i] == first and i < len(word) - 1 and word[i + 1] == second:
+                    new_word.append(first + second)
+                    i += 2
+                else:
+                    new_word.append(word[i])
+                    i += 1
+            new_word = tuple(new_word)
+            word = new_word
+            if len(word) == 1:
+                break
+            else:
+                pairs = get_pairs(word)
+        word = " ".join(word)
+        self.cache[token] = word
+        return word
+
+    def encode(self, text):
+        bpe_tokens = []
+        text = whitespace_clean(basic_clean(text)).lower()
+        for token in re.findall(self.pat, text):
+            token = "".join(self.byte_encoder[b] for b in token.encode("utf-8"))
+            bpe_tokens.extend(
+                self.encoder[bpe_token] for bpe_token in self.bpe(token).split(" ")
+            )
+        return bpe_tokens
+
+    def decode(self, tokens):
+        text = "".join([self.decoder[token] for token in tokens])
+        text = (
+            bytearray([self.byte_decoder[c] for c in text])
+            .decode("utf-8", errors="replace")
+            .replace("</w>", " ")
+        )
+        return text
+
+    def __call__(self, texts, context_length=None):
+        if not context_length:
+            context_length = self.context_length
+
+        if isinstance(texts, str):
+            texts = [texts]
+
+        sot_token = self.encoder["<|startoftext|>"]
+        eot_token = self.encoder["<|endoftext|>"]
+        all_tokens = [[sot_token] + self.encode(text) + [eot_token] for text in texts]
+        result = torch.zeros(len(all_tokens), context_length, dtype=torch.long)
+
+        for i, tokens in enumerate(all_tokens):
+            tokens = tokens[:context_length]
+            result[i, : len(tokens)] = torch.tensor(tokens)
+
+        if len(result) == 1:
+            return result[0]
+        return result
+
+
+class IMUPreprocessor(VerboseNNModule):
+    def __init__(
+        self,
+        kernel_size: int,
+        imu_stem: PatchEmbedGeneric,
+        embed_dim: int,
+        img_size: Tuple = (6, 2000),
+        num_cls_tokens: int = 1,
+        pos_embed_fn: Optional[Callable] = None,
+        init_param_style: str = "openclip",
+    ) -> None:
+        super().__init__()
+        self.imu_stem = imu_stem
+        self.embed_dim = embed_dim
+        self.use_pos_embed = pos_embed_fn is not None
+        self.num_cls_tokens = num_cls_tokens
+        self.kernel_size = kernel_size
+        self.pos_embed = nn.Parameter(
+            torch.empty(1, (img_size[1] // kernel_size) + num_cls_tokens, embed_dim)
+        )
+
+        if self.num_cls_tokens > 0:
+            self.cls_token = nn.Parameter(
+                torch.zeros(1, self.num_cls_tokens, self.embed_dim)
+            )
+
+        self.init_parameters(init_param_style)
+
+    @torch.no_grad()
+    def init_parameters(self, init_param_style):
+        nn.init.normal_(self.pos_embed, std=0.01)
+
+        if init_param_style == "openclip":
+            # OpenCLIP style initialization
+            scale = self.embed_dim**-0.5
+
+            if self.num_cls_tokens > 0:
+                nn.init.normal_(self.cls_token)
+                self.cls_token *= scale
+        elif init_param_style == "vit":
+            self.cls_token.data.fill_(0)
+        else:
+            raise ValueError(f"Unknown init {init_param_style}")
+
+    def tokenize_input_and_cls_pos(self, input, stem):
+        # tokens is of shape B x L x D
+        tokens = stem.norm_layer(stem.proj(input))
+        assert tokens.ndim == 3
+        assert tokens.shape[2] == self.embed_dim
+        B = tokens.shape[0]
+        if self.num_cls_tokens > 0:
+            class_tokens = self.cls_token.expand(
+                B, -1, -1
+            )  # stole class_tokens impl from Phil Wang, thanks
+            tokens = torch.cat((class_tokens, tokens), dim=1)
+        if self.use_pos_embed:
+            tokens = tokens + self.pos_embed
+        return tokens
+
+    def forward(self, imu):
+        # Patchify
+        imu = imu.unfold(
+            -1,
+            self.kernel_size,
+            self.kernel_size,
+        ).permute(0, 2, 1, 3)
+        imu = imu.reshape(imu.size(0), imu.size(1), -1)
+
+        imu_tokens = self.tokenize_input_and_cls_pos(
+            imu,
+            self.imu_stem,
+        )
+
+        return_dict = {
+            "trunk": {
+                "tokens": imu_tokens,
+            },
+            "head": {},
+        }
+        return return_dict

ImageBind/imagebind/models/transformer.py

@@ -0,0 +1,280 @@
+#!/usr/bin/env python3
+# Portions Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+# Code modified from
+# https://github.com/rwightman/pytorch-image-models/blob/master/timm/models/vision_transformer.py ;
+# https://github.com/facebookresearch/deit/blob/main/models.py
+# and https://github.com/facebookresearch/vissl/blob/main/vissl/models/trunks/vision_transformer.py
+
+
+from functools import partial
+from typing import Callable, List, Optional
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint as checkpoint
+from timm.models.layers import DropPath, trunc_normal_
+
+
+class Attention(nn.Module):
+    def __init__(
+        self,
+        dim,
+        num_heads=8,
+        qkv_bias=False,
+        qk_scale=None,
+        attn_drop=0.0,
+        proj_drop=0.0,
+    ):
+        super().__init__()
+        self.num_heads = num_heads
+        head_dim = dim // num_heads
+        # NOTE scale factor was wrong in my original version,
+        # can set manually to be compat with prev weights
+        self.scale = qk_scale or head_dim**-0.5
+
+        self.qkv = nn.Linear(dim, dim * 3, bias=qkv_bias)
+        self.attn_drop = nn.Dropout(attn_drop)
+        self.proj = nn.Linear(dim, dim)
+        self.proj_drop = nn.Dropout(proj_drop)
+
+    def forward(self, x):
+        B, N, C = x.shape
+        qkv = (
+            self.qkv(x)
+            .reshape(B, N, 3, self.num_heads, C // self.num_heads)
+            .permute(2, 0, 3, 1, 4)
+        )
+        q, k, v = (
+            qkv[0],
+            qkv[1],
+            qkv[2],
+        )  # make torchscript happy (cannot use tensor as tuple)
+
+        attn = (q @ k.transpose(-2, -1)) * self.scale
+        attn = attn.softmax(dim=-1)
+        attn = self.attn_drop(attn)
+
+        x = (attn @ v).transpose(1, 2).reshape(B, N, C)
+        x = self.proj(x)
+        x = self.proj_drop(x)
+        return x
+
+
+class Mlp(nn.Module):
+    def __init__(
+        self,
+        in_features,
+        hidden_features=None,
+        out_features=None,
+        act_layer=nn.GELU,
+        drop=0.0,
+    ):
+        super().__init__()
+        out_features = out_features or in_features
+        hidden_features = hidden_features or in_features
+        self.fc1 = nn.Linear(in_features, hidden_features)
+        self.act = act_layer()
+        self.fc2 = nn.Linear(hidden_features, out_features)
+        self.drop = nn.Dropout(drop)
+
+    def forward(self, x):
+        x = self.fc1(x)
+        x = self.act(x)
+        x = self.drop(x)
+        x = self.fc2(x)
+        x = self.drop(x)
+        return x
+
+
+class MultiheadAttention(nn.MultiheadAttention):
+    def forward(self, x: torch.Tensor, attn_mask: torch.Tensor):
+        return super().forward(x, x, x, need_weights=False, attn_mask=attn_mask)[0]
+
+
+class ViTAttention(Attention):
+    def forward(self, x: torch.Tensor, attn_mask: torch.Tensor):
+        assert attn_mask is None
+        return super().forward(x)
+
+
+class BlockWithMasking(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        attn_target: Callable,
+        mlp_ratio: int = 4,
+        act_layer: Callable = nn.GELU,
+        norm_layer: Callable = nn.LayerNorm,
+        ffn_dropout_rate: float = 0.0,
+        drop_path: float = 0.0,
+        layer_scale_type: Optional[str] = None,
+        layer_scale_init_value: float = 1e-4,
+    ):
+        super().__init__()
+
+        assert not isinstance(
+            attn_target, nn.Module
+        ), "attn_target should be a Callable. Otherwise attn_target is shared across blocks!"
+        self.attn = attn_target()
+        if drop_path > 0.0:
+            self.drop_path = DropPath(drop_path)
+        else:
+            self.drop_path = nn.Identity()
+        self.norm_1 = norm_layer(dim)
+        mlp_hidden_dim = int(mlp_ratio * dim)
+        self.mlp = Mlp(
+            in_features=dim,
+            hidden_features=mlp_hidden_dim,
+            act_layer=act_layer,
+            drop=ffn_dropout_rate,
+        )
+        self.norm_2 = norm_layer(dim)
+        self.layer_scale_type = layer_scale_type
+        if self.layer_scale_type is not None:
+            assert self.layer_scale_type in [
+                "per_channel",
+                "scalar",
+            ], f"Found Layer scale type {self.layer_scale_type}"
+            if self.layer_scale_type == "per_channel":
+                # one gamma value per channel
+                gamma_shape = [1, 1, dim]
+            elif self.layer_scale_type == "scalar":
+                # single gamma value for all channels
+                gamma_shape = [1, 1, 1]
+            # two gammas: for each part of the fwd in the encoder
+            self.layer_scale_gamma1 = nn.Parameter(
+                torch.ones(size=gamma_shape) * layer_scale_init_value,
+                requires_grad=True,
+            )
+            self.layer_scale_gamma2 = nn.Parameter(
+                torch.ones(size=gamma_shape) * layer_scale_init_value,
+                requires_grad=True,
+            )
+
+    def forward(self, x: torch.Tensor, attn_mask: torch.Tensor):
+        if self.layer_scale_type is None:
+            x = x + self.drop_path(self.attn(self.norm_1(x), attn_mask))
+            x = x + self.drop_path(self.mlp(self.norm_2(x)))
+        else:
+            x = (
+                x
+                + self.drop_path(self.attn(self.norm_1(x), attn_mask))
+                * self.layer_scale_gamma1
+            )
+            x = x + self.drop_path(self.mlp(self.norm_2(x))) * self.layer_scale_gamma2
+        return x
+
+
+_LAYER_NORM = partial(nn.LayerNorm, eps=1e-6)
+
+
+class SimpleTransformer(nn.Module):
+    def __init__(
+        self,
+        attn_target: Callable,
+        embed_dim: int,
+        num_blocks: int,
+        block: Callable = BlockWithMasking,
+        pre_transformer_layer: Optional[Callable] = None,
+        post_transformer_layer: Optional[Callable] = None,
+        drop_path_rate: float = 0.0,
+        drop_path_type: str = "progressive",
+        norm_layer: Callable = _LAYER_NORM,
+        mlp_ratio: int = 4,
+        ffn_dropout_rate: float = 0.0,
+        layer_scale_type: Optional[str] = None,  # from cait; possible values are None, "per_channel", "scalar"
+        layer_scale_init_value: float = 1e-4,  # from cait; float
+        weight_init_style: str = "jax",  # possible values jax or pytorch
+    ):
+        """
+        Simple Transformer with the following features
+        1. Supports masked attention
+        2. Supports DropPath
+        3. Supports LayerScale
+        4. Supports Dropout in Attention and FFN
+        5. Makes few assumptions about the input except that it is a Tensor
+        """
+        super().__init__()
+        self.pre_transformer_layer = pre_transformer_layer
+        if drop_path_type == "progressive":
+            dpr = [x.item() for x in torch.linspace(0, drop_path_rate, num_blocks)]
+        elif drop_path_type == "uniform":
+            dpr = [drop_path_rate for i in range(num_blocks)]
+        else:
+            raise ValueError(f"Unknown drop_path_type: {drop_path_type}")
+
+        self.blocks = nn.Sequential(
+            *[
+                block(
+                    dim=embed_dim,
+                    attn_target=attn_target,
+                    mlp_ratio=mlp_ratio,
+                    ffn_dropout_rate=ffn_dropout_rate,
+                    drop_path=dpr[i],
+                    norm_layer=norm_layer,
+                    layer_scale_type=layer_scale_type,
+                    layer_scale_init_value=layer_scale_init_value,
+                )
+                for i in range(num_blocks)
+            ]
+        )
+        self.post_transformer_layer = post_transformer_layer
+        self.weight_init_style = weight_init_style
+        self.apply(self._init_weights)
+
+    def _init_weights(self, m):
+        if isinstance(m, nn.Linear):
+            if self.weight_init_style == "jax":
+                # Based on MAE and official Jax ViT implementation
+                torch.nn.init.xavier_uniform_(m.weight)
+            elif self.weight_init_style == "pytorch":
+                # PyTorch ViT uses trunc_normal_
+                trunc_normal_(m.weight, std=0.02)
+
+            if m.bias is not None:
+                nn.init.constant_(m.bias, 0)
+        elif isinstance(m, (nn.LayerNorm)):
+            nn.init.constant_(m.bias, 0)
+            nn.init.constant_(m.weight, 1.0)
+
+    def forward(
+        self,
+        tokens: torch.Tensor,
+        attn_mask: torch.Tensor = None,
+        use_checkpoint: bool = False,
+        checkpoint_every_n: int = 1,
+        checkpoint_blk_ids: Optional[List[int]] = None,
+    ):
+        """
+        Inputs
+        - tokens: data of shape N x L x D (or L x N x D depending on the attention implementation)
+        - attn: mask of shape L x L
+
+        Output
+        - x: data of shape N x L x D (or L x N x D depending on the attention implementation)
+        """
+        if self.pre_transformer_layer:
+            tokens = self.pre_transformer_layer(tokens)
+        if use_checkpoint and checkpoint_blk_ids is None:
+            checkpoint_blk_ids = [
+                blk_id
+                for blk_id in range(len(self.blocks))
+                if blk_id % checkpoint_every_n == 0
+            ]
+        if checkpoint_blk_ids:
+            checkpoint_blk_ids = set(checkpoint_blk_ids)
+        for blk_id, blk in enumerate(self.blocks):
+            if use_checkpoint and blk_id in checkpoint_blk_ids:
+                tokens = checkpoint.checkpoint(
+                    blk, tokens, attn_mask, use_reentrant=False
+                )
+            else:
+                tokens = blk(tokens, attn_mask=attn_mask)
+        if self.post_transformer_layer:
+            tokens = self.post_transformer_layer(tokens)
+        return tokens

ImageBind/model_card.md

@@ -0,0 +1,94 @@
+# Model Card for ImageBind
+
+Multimodal joint embedding model for image/video, text, audio, depth, IMU, and thermal images.
+Input any of the six modalities and get the same sized embedding that can be used for cross-modal and multimodal tasks.
+
+# Model Details
+
+## Model Description
+
+<!-- Provide a longer summary of what this model is/does. -->
+Multimodal joint embedding model for image/video, text, audio, depth, IMU, and thermal images
+
+- **Developed by:** Meta AI
+- **Model type:** Multimodal model
+- **Language(s) (NLP):** en
+- **License:** CC BY-NC-SA 4.0
+- **Resources for more information:**
+    - [GitHub Repo](https://github.com/facebookresearch/ImageBind)
+
+
+# Uses
+
+<!-- Address questions around how the model is intended to be used, including the foreseeable users of the model and those affected by the model. -->
+This model is intended only for research purposes. It provides a joint embedding space for different modalities -- image/video, text, audio, depth, IMU and thermal images.
+We hope that these joint embeddings can be used for a variety of different cross-modal research, e.g., cross-modal retrieval and combining embeddings from different modalities.
+
+## Out-of-Scope Use
+
+<!-- This section addresses misuse, malicious use, and uses that the model will not work well for. -->
+<!-- If the user enters content, print that. If not, but they enter a task in the list, use that. If neither, say "more info needed." -->
+
+This model is *NOT* intended to be used in any real world application -- commercial or otherwise.
+It may produce harmful associations with different inputs.
+The model needs to be investigated and likely re-trained on specific data for any such application.
+The model is expected to work better on web-based visual data since it was trained on such data.
+The text encoder is likely to work only on English language text because of the underlying training datasets.
+
+# Bias, Risks, and Limitations
+
+<!-- This section is meant to convey both technical and sociotechnical limitations. -->
+Open-domain joint embedding models are prone to producing specific biases, e.g., study from [CLIP](https://github.com/openai/CLIP/blob/main/model-card.md#bias-and-fairness).
+Since our model uses such models as initialization, it will exhibit such biases too.
+Moreover, for learning joint embeddings for other modalities such as audio, thermal, depth, and IMU we leverage datasets that are relatively small. These joint embeddings are thus limited to the concepts present in the datasets. For example, the thermal datasets we used are limited to outdoor street scenes, while the depth datasets are limited to indoor scenes.
+
+
+
+# Training Details
+
+## Training Data
+
+<!-- This should link to a Data Card, perhaps with a short stub of information on what the training data is all about as well as documentation related to data pre-processing or additional filtering. -->
+
+ImageBind uses image-paired data for training -- (image, X) where X is one of text, audio, depth, IMU or thermal data.
+In particular, we initialize and freeze the image and text encoders using an OpenCLIP ViT-H encoder.
+We train audio embeddings using Audioset, depth embeddings using the SUN RGB-D dataset, IMU using the Ego4D dataset and thermal embeddings using the LLVIP dataset.
+We provide the exact training data details in the paper.
+
+
+## Training Procedure
+
+<!-- This relates heavily to the Technical Specifications. Content here should link to that section when it is relevant to the training procedure. -->
+Please refer to the research paper and github repo for exact details on this.
+
+# Evaluation
+
+## Testing Data, Factors & Metrics
+
+We evaluate the model on a variety of different classification benchmarks for each modality.
+The evaluation details are presented in the paper.
+The models performance is measured using standard classification metrics such as accuracy and mAP.
+
+# Citation
+
+<!-- If there is a paper or blog post introducing the model, the APA and Bibtex information for that should go in this section. -->
+
+**BibTeX:**
+```
+@inproceedings{girdhar2023imagebind,
+  title={ImageBind: One Embedding Space To Bind Them All},
+  author={Girdhar, Rohit and El-Nouby, Alaaeldin and Liu, Zhuang
+and Singh, Mannat and Alwala, Kalyan Vasudev and Joulin, Armand and Misra, Ishan},
+  booktitle={CVPR},
+  year={2023}
+}
+```
+
+
+# Model Card Contact
+
+Please reach out to the authors at: rgirdhar@meta.com imisra@meta.com alaaelnouby@gmail.com
+
+# How to Get Started with the Model
+
+Our github repo provides a simple example to extract embeddings from images, audio etc.

ImageBind/setup.py

@@ -0,0 +1,23 @@
+from setuptools import setup, find_packages
+
+with open('requirements.txt') as f:
+    required = f.read().splitlines()
+
+setup(
+    name='imagebind',
+    version='0.1.0',
+    packages=find_packages(),
+    package_data={
+        'imagebind': ['bpe/bpe_simple_vocab_16e6.txt.gz'],
+    },
+    description='A brief description of the package',
+    long_description=open('README.md', encoding='utf-8').read(),
+    long_description_content_type="text/markdown",
+    url='https://github.com/facebookresearch/ImageBind',
+    classifiers=[
+        'Programming Language :: Python :: 3',
+        'License :: Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International',
+    ],
+    install_requires=required,
+    dependency_links=['https://download.pytorch.org/whl/cu113'],
+)

LICENSE

@@ -0,0 +1,437 @@
+Attribution-NonCommercial-ShareAlike 4.0 International
+
+=======================================================================
+
+Creative Commons Corporation ("Creative Commons") is not a law firm and
+does not provide legal services or legal advice. Distribution of
+Creative Commons public licenses does not create a lawyer-client or
+other relationship. Creative Commons makes its licenses and related
+information available on an "as-is" basis. Creative Commons gives no
+warranties regarding its licenses, any material licensed under their
+terms and conditions, or any related information. Creative Commons
+disclaims all liability for damages resulting from their use to the
+fullest extent possible.
+
+Using Creative Commons Public Licenses
+
+Creative Commons public licenses provide a standard set of terms and
+conditions that creators and other rights holders may use to share
+original works of authorship and other material subject to copyright
+and certain other rights specified in the public license below. The
+following considerations are for informational purposes only, are not
+exhaustive, and do not form part of our licenses.
+
+     Considerations for licensors: Our public licenses are
+     intended for use by those authorized to give the public
+     permission to use material in ways otherwise restricted by
+     copyright and certain other rights. Our licenses are
+     irrevocable. Licensors should read and understand the terms
+     and conditions of the license they choose before applying it.
+     Licensors should also secure all rights necessary before
+     applying our licenses so that the public can reuse the
+     material as expected. Licensors should clearly mark any
+     material not subject to the license. This includes other CC-
+     licensed material, or material used under an exception or
+     limitation to copyright. More considerations for licensors:
+    wiki.creativecommons.org/Considerations_for_licensors
+
+     Considerations for the public: By using one of our public
+     licenses, a licensor grants the public permission to use the
+     licensed material under specified terms and conditions. If
+     the licensor's permission is not necessary for any reason--for
+     example, because of any applicable exception or limitation to
+     copyright--then that use is not regulated by the license. Our
+     licenses grant only permissions under copyright and certain
+     other rights that a licensor has authority to grant. Use of
+     the licensed material may still be restricted for other
+     reasons, including because others have copyright or other
+     rights in the material. A licensor may make special requests,
+     such as asking that all changes be marked or described.
+     Although not required by our licenses, you are encouraged to
+     respect those requests where reasonable. More considerations
+     for the public:
+    wiki.creativecommons.org/Considerations_for_licensees
+
+=======================================================================
+
+Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International
+Public License
+
+By exercising the Licensed Rights (defined below), You accept and agree
+to be bound by the terms and conditions of this Creative Commons
+Attribution-NonCommercial-ShareAlike 4.0 International Public License
+("Public License"). To the extent this Public License may be
+interpreted as a contract, You are granted the Licensed Rights in
+consideration of Your acceptance of these terms and conditions, and the
+Licensor grants You such rights in consideration of benefits the
+Licensor receives from making the Licensed Material available under
+these terms and conditions.
+
+
+Section 1 -- Definitions.
+
+  a. Adapted Material means material subject to Copyright and Similar
+     Rights that is derived from or based upon the Licensed Material
+     and in which the Licensed Material is translated, altered,
+     arranged, transformed, or otherwise modified in a manner requiring
+     permission under the Copyright and Similar Rights held by the
+     Licensor. For purposes of this Public License, where the Licensed
+     Material is a musical work, performance, or sound recording,
+     Adapted Material is always produced where the Licensed Material is
+     synched in timed relation with a moving image.
+
+  b. Adapter's License means the license You apply to Your Copyright
+     and Similar Rights in Your contributions to Adapted Material in
+     accordance with the terms and conditions of this Public License.
+
+  c. BY-NC-SA Compatible License means a license listed at
+     creativecommons.org/compatiblelicenses, approved by Creative
+     Commons as essentially the equivalent of this Public License.
+
+  d. Copyright and Similar Rights means copyright and/or similar rights
+     closely related to copyright including, without limitation,
+     performance, broadcast, sound recording, and Sui Generis Database
+     Rights, without regard to how the rights are labeled or
+     categorized. For purposes of this Public License, the rights
+     specified in Section 2(b)(1)-(2) are not Copyright and Similar
+     Rights.
+
+  e. Effective Technological Measures means those measures that, in the
+     absence of proper authority, may not be circumvented under laws
+     fulfilling obligations under Article 11 of the WIPO Copyright
+     Treaty adopted on December 20, 1996, and/or similar international
+     agreements.
+
+  f. Exceptions and Limitations means fair use, fair dealing, and/or
+     any other exception or limitation to Copyright and Similar Rights
+     that applies to Your use of the Licensed Material.
+
+  g. License Elements means the license attributes listed in the name
+     of a Creative Commons Public License. The License Elements of this
+     Public License are Attribution, NonCommercial, and ShareAlike.
+
+  h. Licensed Material means the artistic or literary work, database,
+     or other material to which the Licensor applied this Public
+     License.
+
+  i. Licensed Rights means the rights granted to You subject to the
+     terms and conditions of this Public License, which are limited to
+     all Copyright and Similar Rights that apply to Your use of the
+     Licensed Material and that the Licensor has authority to license.
+
+  j. Licensor means the individual(s) or entity(ies) granting rights
+     under this Public License.
+
+  k. NonCommercial means not primarily intended for or directed towards
+     commercial advantage or monetary compensation. For purposes of
+     this Public License, the exchange of the Licensed Material for
+     other material subject to Copyright and Similar Rights by digital
+     file-sharing or similar means is NonCommercial provided there is
+     no payment of monetary compensation in connection with the
+     exchange.
+
+  l. Share means to provide material to the public by any means or
+     process that requires permission under the Licensed Rights, such
+     as reproduction, public display, public performance, distribution,
+     dissemination, communication, or importation, and to make material
+     available to the public including in ways that members of the
+     public may access the material from a place and at a time
+     individually chosen by them.
+
+  m. Sui Generis Database Rights means rights other than copyright
+     resulting from Directive 96/9/EC of the European Parliament and of
+     the Council of 11 March 1996 on the legal protection of databases,
+     as amended and/or succeeded, as well as other essentially
+     equivalent rights anywhere in the world.
+
+  n. You means the individual or entity exercising the Licensed Rights
+     under this Public License. Your has a corresponding meaning.
+
+
+Section 2 -- Scope.
+
+  a. License grant.
+
+       1. Subject to the terms and conditions of this Public License,
+          the Licensor hereby grants You a worldwide, royalty-free,
+          non-sublicensable, non-exclusive, irrevocable license to
+          exercise the Licensed Rights in the Licensed Material to:
+
+            a. reproduce and Share the Licensed Material, in whole or
+               in part, for NonCommercial purposes only; and
+
+            b. produce, reproduce, and Share Adapted Material for
+               NonCommercial purposes only.
+
+       2. Exceptions and Limitations. For the avoidance of doubt, where
+          Exceptions and Limitations apply to Your use, this Public
+          License does not apply, and You do not need to comply with
+          its terms and conditions.
+
+       3. Term. The term of this Public License is specified in Section
+          6(a).
+
+       4. Media and formats; technical modifications allowed. The
+          Licensor authorizes You to exercise the Licensed Rights in
+          all media and formats whether now known or hereafter created,
+          and to make technical modifications necessary to do so. The
+          Licensor waives and/or agrees not to assert any right or
+          authority to forbid You from making technical modifications
+          necessary to exercise the Licensed Rights, including
+          technical modifications necessary to circumvent Effective
+          Technological Measures. For purposes of this Public License,
+          simply making modifications authorized by this Section 2(a)
+          (4) never produces Adapted Material.
+
+       5. Downstream recipients.
+
+            a. Offer from the Licensor -- Licensed Material. Every
+               recipient of the Licensed Material automatically
+               receives an offer from the Licensor to exercise the
+               Licensed Rights under the terms and conditions of this
+               Public License.
+
+            b. Additional offer from the Licensor -- Adapted Material.
+               Every recipient of Adapted Material from You
+               automatically receives an offer from the Licensor to
+               exercise the Licensed Rights in the Adapted Material
+               under the conditions of the Adapter's License You apply.
+
+            c. No downstream restrictions. You may not offer or impose
+               any additional or different terms or conditions on, or
+               apply any Effective Technological Measures to, the
+               Licensed Material if doing so restricts exercise of the
+               Licensed Rights by any recipient of the Licensed
+               Material.
+
+       6. No endorsement. Nothing in this Public License constitutes or
+          may be construed as permission to assert or imply that You
+          are, or that Your use of the Licensed Material is, connected
+          with, or sponsored, endorsed, or granted official status by,
+          the Licensor or others designated to receive attribution as
+          provided in Section 3(a)(1)(A)(i).
+
+  b. Other rights.
+
+       1. Moral rights, such as the right of integrity, are not
+          licensed under this Public License, nor are publicity,
+          privacy, and/or other similar personality rights; however, to
+          the extent possible, the Licensor waives and/or agrees not to
+          assert any such rights held by the Licensor to the limited
+          extent necessary to allow You to exercise the Licensed
+          Rights, but not otherwise.
+
+       2. Patent and trademark rights are not licensed under this
+          Public License.
+
+       3. To the extent possible, the Licensor waives any right to
+          collect royalties from You for the exercise of the Licensed
+          Rights, whether directly or through a collecting society
+          under any voluntary or waivable statutory or compulsory
+          licensing scheme. In all other cases the Licensor expressly
+          reserves any right to collect such royalties, including when
+          the Licensed Material is used other than for NonCommercial
+          purposes.
+
+
+Section 3 -- License Conditions.
+
+Your exercise of the Licensed Rights is expressly made subject to the
+following conditions.
+
+  a. Attribution.
+
+       1. If You Share the Licensed Material (including in modified
+          form), You must:
+
+            a. retain the following if it is supplied by the Licensor
+               with the Licensed Material:
+
+                 i. identification of the creator(s) of the Licensed
+                    Material and any others designated to receive
+                    attribution, in any reasonable manner requested by
+                    the Licensor (including by pseudonym if
+                    designated);
+
+                ii. a copyright notice;
+
+               iii. a notice that refers to this Public License;
+
+                iv. a notice that refers to the disclaimer of
+                    warranties;
+
+                 v. a URI or hyperlink to the Licensed Material to the
+                    extent reasonably practicable;
+
+            b. indicate if You modified the Licensed Material and
+               retain an indication of any previous modifications; and
+
+            c. indicate the Licensed Material is licensed under this
+               Public License, and include the text of, or the URI or
+               hyperlink to, this Public License.
+
+       2. You may satisfy the conditions in Section 3(a)(1) in any
+          reasonable manner based on the medium, means, and context in
+          which You Share the Licensed Material. For example, it may be
+          reasonable to satisfy the conditions by providing a URI or
+          hyperlink to a resource that includes the required
+          information.
+       3. If requested by the Licensor, You must remove any of the
+          information required by Section 3(a)(1)(A) to the extent
+          reasonably practicable.
+
+  b. ShareAlike.
+
+     In addition to the conditions in Section 3(a), if You Share
+     Adapted Material You produce, the following conditions also apply.
+
+       1. The Adapter's License You apply must be a Creative Commons
+          license with the same License Elements, this version or
+          later, or a BY-NC-SA Compatible License.
+
+       2. You must include the text of, or the URI or hyperlink to, the
+          Adapter's License You apply. You may satisfy this condition
+          in any reasonable manner based on the medium, means, and
+          context in which You Share Adapted Material.
+
+       3. You may not offer or impose any additional or different terms
+          or conditions on, or apply any Effective Technological
+          Measures to, Adapted Material that restrict exercise of the
+          rights granted under the Adapter's License You apply.
+
+
+Section 4 -- Sui Generis Database Rights.
+
+Where the Licensed Rights include Sui Generis Database Rights that
+apply to Your use of the Licensed Material:
+
+  a. for the avoidance of doubt, Section 2(a)(1) grants You the right
+     to extract, reuse, reproduce, and Share all or a substantial
+     portion of the contents of the database for NonCommercial purposes
+     only;
+
+  b. if You include all or a substantial portion of the database
+     contents in a database in which You have Sui Generis Database
+     Rights, then the database in which You have Sui Generis Database
+     Rights (but not its individual contents) is Adapted Material,
+     including for purposes of Section 3(b); and
+
+  c. You must comply with the conditions in Section 3(a) if You Share
+     all or a substantial portion of the contents of the database.
+
+For the avoidance of doubt, this Section 4 supplements and does not
+replace Your obligations under this Public License where the Licensed
+Rights include other Copyright and Similar Rights.
+
+
+Section 5 -- Disclaimer of Warranties and Limitation of Liability.
+
+  a. UNLESS OTHERWISE SEPARATELY UNDERTAKEN BY THE LICENSOR, TO THE
+     EXTENT POSSIBLE, THE LICENSOR OFFERS THE LICENSED MATERIAL AS-IS
+     AND AS-AVAILABLE, AND MAKES NO REPRESENTATIONS OR WARRANTIES OF
+     ANY KIND CONCERNING THE LICENSED MATERIAL, WHETHER EXPRESS,
+     IMPLIED, STATUTORY, OR OTHER. THIS INCLUDES, WITHOUT LIMITATION,
+     WARRANTIES OF TITLE, MERCHANTABILITY, FITNESS FOR A PARTICULAR
+     PURPOSE, NON-INFRINGEMENT, ABSENCE OF LATENT OR OTHER DEFECTS,
+     ACCURACY, OR THE PRESENCE OR ABSENCE OF ERRORS, WHETHER OR NOT
+     KNOWN OR DISCOVERABLE. WHERE DISCLAIMERS OF WARRANTIES ARE NOT
+     ALLOWED IN FULL OR IN PART, THIS DISCLAIMER MAY NOT APPLY TO YOU.
+
+  b. TO THE EXTENT POSSIBLE, IN NO EVENT WILL THE LICENSOR BE LIABLE
+     TO YOU ON ANY LEGAL THEORY (INCLUDING, WITHOUT LIMITATION,
+     NEGLIGENCE) OR OTHERWISE FOR ANY DIRECT, SPECIAL, INDIRECT,
+     INCIDENTAL, CONSEQUENTIAL, PUNITIVE, EXEMPLARY, OR OTHER LOSSES,
+     COSTS, EXPENSES, OR DAMAGES ARISING OUT OF THIS PUBLIC LICENSE OR
+     USE OF THE LICENSED MATERIAL, EVEN IF THE LICENSOR HAS BEEN
+     ADVISED OF THE POSSIBILITY OF SUCH LOSSES, COSTS, EXPENSES, OR
+     DAMAGES. WHERE A LIMITATION OF LIABILITY IS NOT ALLOWED IN FULL OR
+     IN PART, THIS LIMITATION MAY NOT APPLY TO YOU.
+
+  c. The disclaimer of warranties and limitation of liability provided
+     above shall be interpreted in a manner that, to the extent
+     possible, most closely approximates an absolute disclaimer and
+     waiver of all liability.
+
+
+Section 6 -- Term and Termination.
+
+  a. This Public License applies for the term of the Copyright and
+     Similar Rights licensed here. However, if You fail to comply with
+     this Public License, then Your rights under this Public License
+     terminate automatically.
+
+  b. Where Your right to use the Licensed Material has terminated under
+     Section 6(a), it reinstates:
+
+       1. automatically as of the date the violation is cured, provided
+          it is cured within 30 days of Your discovery of the
+          violation; or
+
+       2. upon express reinstatement by the Licensor.
+
+     For the avoidance of doubt, this Section 6(b) does not affect any
+     right the Licensor may have to seek remedies for Your violations
+     of this Public License.
+
+  c. For the avoidance of doubt, the Licensor may also offer the
+     Licensed Material under separate terms or conditions or stop
+     distributing the Licensed Material at any time; however, doing so
+     will not terminate this Public License.
+
+  d. Sections 1, 5, 6, 7, and 8 survive termination of this Public
+     License.
+
+
+Section 7 -- Other Terms and Conditions.
+
+  a. The Licensor shall not be bound by any additional or different
+     terms or conditions communicated by You unless expressly agreed.
+
+  b. Any arrangements, understandings, or agreements regarding the
+     Licensed Material not stated herein are separate from and
+     independent of the terms and conditions of this Public License.
+
+
+Section 8 -- Interpretation.
+
+  a. For the avoidance of doubt, this Public License does not, and
+     shall not be interpreted to, reduce, limit, restrict, or impose
+     conditions on any use of the Licensed Material that could lawfully
+     be made without permission under this Public License.
+
+  b. To the extent possible, if any provision of this Public License is
+     deemed unenforceable, it shall be automatically reformed to the
+     minimum extent necessary to make it enforceable. If the provision
+     cannot be reformed, it shall be severed from this Public License
+     without affecting the enforceability of the remaining terms and
+     conditions.
+
+  c. No term or condition of this Public License will be waived and no
+     failure to comply consented to unless expressly agreed to by the
+     Licensor.
+
+  d. Nothing in this Public License constitutes or may be interpreted
+     as a limitation upon, or waiver of, any privileges and immunities
+     that apply to the Licensor or You, including from the legal
+     processes of any jurisdiction or authority.
+
+=======================================================================
+
+Creative Commons is not a party to its public
+licenses. Notwithstanding, Creative Commons may elect to apply one of
+its public licenses to material it publishes and in those instances
+will be considered the “Licensor.” The text of the Creative Commons
+public licenses is dedicated to the public domain under the CC0 Public
+Domain Dedication. Except for the limited purpose of indicating that
+material is shared under a Creative Commons public license or as
+otherwise permitted by the Creative Commons policies published at
+creativecommons.org/policies, Creative Commons does not authorize the
+use of the trademark "Creative Commons" or any other trademark or logo
+of Creative Commons without its prior written consent including,
+without limitation, in connection with any unauthorized modifications
+to any of its public licenses or any other arrangements,
+understandings, or agreements concerning use of licensed material. For
+the avoidance of doubt, this paragraph does not form part of the
+public licenses.
+
+Creative Commons may be contacted at creativecommons.org.

README.md

@@ -0,0 +1,215 @@
+# VideoRAG: Retrieval-Augmented Generation with Extreme Long-Context Videos
+
+<img src='VideoRAG_cover.png' />
+
+ This is the PyTorch implementation for VideoRAG proposed in this paper:
+
+ >**VideoRAG: Retrieval-Augmented Generation with Extreme Long-Context Videos**  
+ >Xubin Ren*, Lingrui Xu*, Long Xia, Shuaiqiang Wang, Dawei Yin, Chao Huang†
+
+\* denotes equal contribution.
+† denotes corresponding author
+
+ In this paper, we proposed a retrieval-augmented generation framework specifically designed for processing and understanding **extremely long-context videos**.
+
+## VideoRAG Framework
+
+<p align="center">
+<img src="VideoRAG.png" alt="VideoRAG" />
+</p>
+
+VideoRAG introduces a novel dual-channel architecture that synergistically combines graph-driven textual knowledge grounding for modeling cross-video semantic relationships with hierarchical multimodal context encoding to preserve spatiotemporal visual patterns, enabling unbounded-length video understanding through dynamically constructed knowledge graphs that maintain semantic coherence across multi-video contexts while optimizing retrieval efficiency via adaptive multimodal fusion mechanisms.
+
+## Installation
+
+To utilize VideoRAG, please first create a conda environment with the following commands:
+```bash
+conda create --name videorag python=3.11
+conda activate videorag
+
+pip install numpy==1.26.4
+pip install torch==2.1.2 torchvision==0.16.2 torchaudio==2.1.2
+pip install accelerate==0.30.1
+pip install bitsandbytes==0.43.1
+pip install moviepy==1.0.3
+pip install git+https://github.com/facebookresearch/pytorchvideo.git@28fe037d212663c6a24f373b94cc5d478c8c1a1d
+pip install timm==0.6.7 ftfy regex einops fvcore eva-decord==0.6.1 iopath matplotlib types-regex cartopy
+pip install ctranslate2==4.4.0 faster_whisper neo4j hnswlib xxhash nano-vectordb
+pip install transformers==4.37.1
+pip install tiktoken openai tenacity
+
+# Install ImageBind using the provided code in this repository, where we have removed the requirements.txt to avoid environment conflicts.
+cd ImageBind
+pip install .
+```
+
+Then, please download the necessary checkpoints in **the repository's root folder** for MiniCPM-V, Whisper, and ImageBind as follows:
+```bash
+# minicpm-v
+git clone https://huggingface.co/openbmb/MiniCPM-V-2_6-int4
+
+# whisper
+git clone https://huggingface.co/Systran/faster-distil-whisper-large-v3
+
+# imagebind
+mkdir .checkpoints
+cd .checkpoints
+wget https://dl.fbaipublicfiles.com/imagebind/imagebind_huge.pth
+cd ../
+```
+
+## Quick Start
+
+VideoRAG is capable of extracting knowledge from multiple videos and answering queries based on those videos. Now, try VideoRAG with your own videos 🤗.
+
+> [!NOTE]
+> Currently, VideoRAG has only been tested in an English environment. To process videos in multiple languages, it is recommended to modify the  ```WhisperModel``` in [asr.py](VideoRAG/videorag/_videoutil/asr.py). For more details, please refer to [faster-whisper](https://github.com/systran/faster-whisper).
+
+**At first**, let the VideoRAG extract and indexing the knowledge from given videos (Only one GPU with 24GB of memory is sufficient, such as the RTX 3090):
+```python
+import os
+import logging
+import warnings
+import multiprocessing
+
+warnings.filterwarnings("ignore")
+logging.getLogger("httpx").setLevel(logging.WARNING)
+
+# Please enter your openai key
+os.environ["OPENAI_API_KEY"] = ""
+
+from videorag._llm import *
+from videorag import VideoRAG, QueryParam
+
+
+if __name__ == '__main__':
+    multiprocessing.set_start_method('spawn')
+
+    # Please enter your video file path in this list; there is no limit on the length.
+    # Here is an example; you can use your own videos instead.
+    video_paths = [
+        'movies/Iron-Man.mp4',
+        'movies/Spider-Man.mkv',
+    ]
+    videorag = VideoRAG(cheap_model_func=gpt_4o_mini_complete, best_model_func=gpt_4o_mini_complete, working_dir=f"./videorag-workdir")
+    videorag.insert_video(video_path_list=video_paths)
+```
+
+**Then**, ask any questions about the videos! Here is an exmaple:
+```python
+import os
+import logging
+import warnings
+import multiprocessing
+
+warnings.filterwarnings("ignore")
+logging.getLogger("httpx").setLevel(logging.WARNING)
+
+# Please enter your openai key
+os.environ["OPENAI_API_KEY"] = ""
+
+from videorag._llm import *
+from videorag import VideoRAG, QueryParam
+
+
+if __name__ == '__main__':
+    multiprocessing.set_start_method('spawn')
+
+    query = 'What is the relationship between Iron Man and Spider-Man? How do they meet, and how does Iron Man help Spider-Man?'
+    param = QueryParam(mode="videorag")
+    # if param.wo_reference = False, VideoRAG will add reference to video clips in the response
+    param.wo_reference = True
+
+    videorag = VideoRAG(cheap_model_func=gpt_4o_mini_complete, best_model_func=gpt_4o_mini_complete, working_dir=f"./videorag-workdir")
+    videorag.load_caption_model(debug=False)
+    response = videorag.query(query=query, param=param)
+    print(response)
+```
+
+## Evaluation
+
+### LongerVideos
+We constructed the LongerVideos benchmark to evaluate the model's performance in comprehending multiple long-context videos and answering open-ended queries. All the videos are open-access videos on YouTube, and we record the URLs of the collections of videos as well as the corresponding queries in the [JSON](longervideos/dataset.json) file.
+
+| Video Type       | #video list | #video | #query | #avg. queries per list | #overall duration      |
+|------------------|------------:|-------:|-------:|-----------------------:|-------------------------|
+| **Lecture**      | 12          | 135    | 376    | 31.3                   | ~ 64.3 hours           |
+| **Documentary**  | 5           | 12     | 114    | 22.8                   | ~ 28.5 hours           |
+| **Entertainment**| 5           | 17     | 112    | 22.4                   | ~ 41.9 hours           |
+| **All**          | 22          | 164    | 602    | 27.4                   | ~ 134.6 hours          |
+
+### Process LongerVideos with VideoRAG
+
+First, we need to download all the videos in 720p resolution using the `yt-dlp` library. You can use the following commands to download all the videos in LongerVideos:
+
+```shell
+cd longervideos
+python prepare_videos.py # create collection folders
+sh download.sh # downloading videos
+```
+
+Then, you can run the following example command to process and answer queries for LongerVideos with VideoRAG:
+
+```shell
+# Please enter your openai_key in line 18 at first
+python videorag_experiment.py --collection 4-rag-lecture --cuda 0
+```
+
+### Evaluation
+
+We conduct win-rate comparisons as well as quantitative comparisons with RAG-based baselines and long-context video understanding methods separately. **NaiveRAG, GraphRAG and LightRAG** are implemented using the `nano-graphrag` library, which is consistent with our VideoRAG, ensuring a fair comparison.
+
+In this part, we directly provided the **answers from all the methods** (including VideoRAG) as well as the evaluation codes for experiment reproduction. Please utilize the following commands to download the answers:
+
+```shell
+cd reproduce
+wget https://archive.org/download/videorag/all_answers.zip
+unzip all_answers
+```
+
+#### Win-Rate Comparison
+
+We conduct the win-rate comparison with RAG-based baselines. To reproduce the results, please follow these steps:
+
+```shell
+cd reproduce/winrate_comparison
+
+# First Step: Upload the batch request to OpenAI (remember to enter your key in the file, same for the following steps).
+python batch_winrate_eval_upload.py
+
+# Second Step: Download the results. Please enter the batch ID and then the output file ID in the file. Generally, you need to run this twice: first to obtain the output file ID, and then to download it.
+python batch_winrate_eval_download.py
+
+# Third Step: Parsing the results. Please the output file ID in the file.
+python batch_winrate_eval_download.py
+
+# Fourth Step: Calculate the results. Please enter the parsed result file name in the file.
+python batch_winrate_eval_calculate.py
+
+```
+
+#### Quantitative Comparison
+
+We conduct a quantitative comparison, which extends the win-rate comparison by assigning a 5-point score to long-context video understanding methods. We use the answers from NaiveRAG as the baseline response for scoring each query. To reproduce the results, please follow these steps:
+
+```shell
+cd reproduce/quantitative_comparison
+
+# First Step: Upload the batch request to OpenAI (remember to enter your key in the file, same for the following steps).
+python batch_winrate_quant_upload.py
+
+# Second Step: Download the results. Please enter the batch ID and then the output file ID in the file. Generally, you need to run this twice: first to obtain the output file ID, and then to download it.
+python batch_winrate_quant_download.py
+
+# Third Step: Parsing the results. Please the output file ID in the file.
+python batch_winrate_quant_download.py
+
+# Fourth Step: Calculate the results. Please enter the parsed result file name in the file.
+python batch_winrate_quant_calculate.py
+```
+
+### Acknowledgement
+You may refer to related work that serves as foundations for our framework and code repository, 
+[nano-graphrag](https://github.com/gusye1234/nano-graphrag) and [LightRAG](https://github.com/HKUDS/LightRAG). Thanks for their wonderful works.
+
+**Thank you for your interest in our work!**

longervideos/download.sh

@@ -0,0 +1,11 @@
+#!/bin/bash
+
+set -ex
+
+courses="0-fights-in-animal-kingdom 1-nature-scenes 2-climate-week-at-columbia-engineering 3-black-myth-wukong 4-rag-lecture 5-ai-agent-lecture 6-daubechies-wavelet-lecture 7-daubechies-art-and-mathematics-lecture 8-tech-ceo-lecture 9-dspy-lecture 10-trading-for-beginners 11-primetime-emmy-awards 12-journey-through-china 13-fia-awards 14-education-united-nations 15-game-awards 16-ahp-superdecision 17-decision-making-science 18-elon-musk 19-jeff-bezos 20-12-days-of-openai 21-autogen"
+
+for course in $courses; do
+    mkdir -p ./$course/videos
+    yt-dlp -o "%(id)s.%(ext)s" -S "res:720" -a "./$course/videos.txt" -P "./$course/videos" >> ./download_log.txt 2>&1
+    wait
+done

longervideos/prepare_data.py

@@ -0,0 +1,18 @@
+import os
+import json
+
+with open('./dataset.json', 'rb') as f:
+    longervideos = json.load(f)
+
+collections = []
+for _id in longervideos:
+    collection = longervideos[_id][0]
+    collection_name = f"{_id}-{collection['description']}"
+    collections.append(collection_name)
+    os.makedirs(os.path.join(collection_name, 'videos'), exist_ok=True)
+    with open(os.path.join(collection_name, 'videos.txt'), 'w') as f:
+        for i in range(len(collection['video_url'])):
+            _url = collection['video_url'][i]
+            f.write(f'{_url}')
+            if i != len(collection['video_url']) - 1:
+                f.write(f'\n')

longervideos/videorag_experiment.py

@@ -0,0 +1,55 @@
+import os
+import json
+import logging
+import warnings
+import multiprocessing
+
+warnings.filterwarnings("ignore")
+logging.getLogger("httpx").setLevel(logging.WARNING)
+
+import argparse
+parser = argparse.ArgumentParser(description="Set sub-category and CUDA device.")
+parser.add_argument('--collection', type=str, default='4-rag-lecture')
+parser.add_argument('--cuda', type=str, default='0')
+args = parser.parse_args()
+sub_category = args.sub_category
+
+os.environ["CUDA_VISIBLE_DEVICES"] = args.cuda
+os.environ["OPENAI_API_KEY"] = ""
+
+from videorag._llm import *
+from videorag import VideoRAG, QueryParam
+
+if __name__ == '__main__':
+    multiprocessing.set_start_method('spawn')
+    
+    ## learn
+    video_base_path = f'./{sub_category}/videos/'
+    video_files = sorted(os.listdir(video_base_path))
+    video_paths = [os.path.join(video_base_path, f) for f in video_files]
+    videorag = VideoRAG(cheap_model_func=gpt_4o_mini_complete, best_model_func=gpt_4o_mini_complete, working_dir=f"./videorag-workdir/{sub_category}")
+    videorag.insert_video(video_path_list=video_paths)
+    
+    ## inference
+    with open(f'./dataset.json', 'r') as f:
+        longervideos = json.load(f)
+    
+    videorag = VideoRAG(cheap_model_func=gpt_4o_mini_complete, best_model_func=gpt_4o_mini_complete, working_dir=f"./videorag-workdir/{sub_category}")
+    videorag.load_caption_model(debug=False)
+    
+    answer_folder = f'./videorag-answers/{sub_category}'
+    os.makedirs(answer_folder, exist_ok=True)
+    
+    collection_id = sub_category.split('-')[0]
+    querys = longervideos[collection_id][0]['questions']
+    for i in range(len(querys)):
+        query_id = querys[i]['id']
+        query = querys[i]['question']
+        param = QueryParam(mode="videorag")
+        param.wo_reference = True
+        print("Query: ", query)
+        
+        response = videorag.query(query=query, param=param)
+        print(response)
+        with open(os.path.join(answer_folder, f'/answer_{query_id}.md'), 'w') as f:
+            f.write(response)

reproduce/quantitative_comparison/batch_quant_eval_calculate.py

@@ -0,0 +1,100 @@
+import os
+import json
+import random
+import numpy as np
+from tqdm import tqdm
+from copy import deepcopy
+
+baseline_model = 'naiverag'
+evaluate_model = [
+    'llamavid',
+    'videoagent',
+    'notebooklm',
+    'videorag'
+]
+
+metrics = ['Comprehensiveness', 'Empowerment', 'Trustworthiness', 'Depth', 'Density', 'Overall Score']
+
+base_dir = 'overall_comparison_video_understanding'
+# Please enter the parsed result files ending with .json below.
+result_file = [
+    '',
+    '',
+    '',
+    '',
+    ''
+]
+
+domain_list = ['lecture', 'documentary', 'entertainment']
+
+with open('../../longervideos/dataset.json', 'r') as f:
+    all_data = json.load(f)
+
+overall_score = {}
+for _model in evaluate_model:
+    overall_score[_model] = {}
+    for _metric in metrics:
+        overall_score[_model][_metric] = []
+
+category_domain_dict = {}
+for category_id in all_data:
+    _domain = all_data[category_id][0]['type']
+    category_domain_dict[category_id] = _domain
+        
+domain_score = {}
+for domain in domain_list:
+    domain_score[domain] = {}
+    for _model in evaluate_model:
+        domain_score[domain][_model] = {}
+        for _metric in metrics:
+            domain_score[domain][_model][_metric] = []
+
+query_count = 0 
+for category_id in tqdm(all_data):
+    category = f"{category_id}-{all_data[category_id][0]['description']}"
+    querys = all_data[category_id][0]['questions']
+    query_count += len(querys)
+    
+    score = {}
+    for _model in evaluate_model:
+        score[_model] = {}
+        for _metric in metrics:
+            score[_model][_metric] = []
+    
+    for _file in result_file:
+        result_path = f'./batch_requests/{base_dir}/{_file}'
+        with open(result_path, 'r') as f:
+            results = json.loads(f.read())
+        
+        for i in range(len(querys)):
+            for _model in evaluate_model:
+                query_id = querys[i]['id']
+                evaluation_result = results[f'{category}++query{query_id}++base++answers-{baseline_model}++evaluate++answers-{_model}']
+
+                for _metric in metrics:
+                    _metric_score = evaluation_result[_metric]['Score']
+                    score[_model][_metric].append(_metric_score)
+                    overall_score[_model][_metric].append(_metric_score)
+                    domain_score[category_domain_dict[category_id]][_model][_metric].append(_metric_score)
+                    
+with open(f'batch_requests/{base_dir}/{base_dir}.txt', 'a') as f:
+    print(query_count)
+    f.write(f'{query_count}\n')
+    for _model in evaluate_model:
+        print(_model)
+        f.write(_model + '\n')
+        for _domain in domain_list:
+            print(_domain)
+            f.write(_domain + '\n')
+            for _metric in metrics:
+                print(f'{np.array(domain_score[_domain][_model][_metric]).mean():.2f}', _metric)
+                f.write(f'{np.array(domain_score[_domain][_model][_metric]).mean():.2f} {_metric}\n')
+            print('----')
+            f.write('----\n')
+        print('All')
+        f.write('All\n')
+        for _metric in metrics:
+            print(f'{np.array(overall_score[_model][_metric]).mean():.2f}', _metric)
+            f.write(f'{np.array(overall_score[_model][_metric]).mean():.2f} {_metric}\n')
+        print('====' * 8)
+        f.write('====' * 8 + '\n')

reproduce/quantitative_comparison/batch_quant_eval_download.py

@@ -0,0 +1,53 @@
+import os
+os.environ["OPENAI_API_KEY"] = ""
+import re
+import time
+import json
+import jsonlines
+import tiktoken
+
+from tqdm import tqdm
+from openai import OpenAI
+
+client = OpenAI()
+
+def obtain_ouput_file_id(batches):
+    for batch in batches:
+        print(client.batches.retrieve(batch))
+        print(client.batches.retrieve(batch).output_file_id)
+
+def download_result(result_files, base_dir):
+    for _file in result_files:
+        content = client.files.content(_file).content
+        with open(f"batch_requests/{base_dir}/{_file}.temp", "wb") as f:
+            f.write(content)
+        results = []
+        with open(f"batch_requests/{base_dir}/{_file}.temp", 'r') as f:
+            for line in tqdm(f):
+                json_object = json.loads(line.strip())
+                results.append(json_object)
+        with open(f"batch_requests/{base_dir}/{_file}.json", "w") as json_file:
+            json.dump(results, json_file, indent=4)
+        os.remove(f"batch_requests/{base_dir}/{_file}.temp")
+
+# ================================
+
+# Please enter the relevant batch ID here to obtain the output file ID.
+batches = [
+    '',
+    '',
+    '',
+    '',
+    ''
+]
+obtain_ouput_file_id(batches)
+
+# Second Step: Please enter the output file ID below to download the output files.
+# result_files = [
+#     '',
+#     '',
+#     '',
+#     '',
+#     ''
+# ]
+# download_result(result_files, base_dir='overall_comparison_video_understanding')

reproduce/quantitative_comparison/batch_quant_eval_parse.py

@@ -0,0 +1,88 @@
+import os
+os.environ["OPENAI_API_KEY"] = ""
+import time
+import json
+import threading
+from tqdm import tqdm
+from openai import OpenAI
+from setproctitle import setproctitle
+
+base_dir = 'overall_comparison_video_understanding'
+# The JSON file contains the batch of requests created when the batch request was uploaded.
+request_file = ''
+# Please enter the output file ID below, which corresponds to the downloaded output files.
+result_files = [
+    '',
+    '',
+    '',
+    '',
+    ''
+]
+
+setproctitle(f"parse-result-{base_dir}")
+print(f"Start parsing result files in {base_dir}...")
+
+def check_response_valid(data):
+    valid_keys = ['Comprehensiveness', 'Empowerment', 'Trustworthiness', 'Depth', 'Density', 'Overall Score']
+    assert len(data) == 6
+    assert set(list(data.keys())) == set(valid_keys)
+    for _key in valid_keys:
+        assert data[_key]["Score"] in [1, 2, 3, 4, 5]
+        assert "Explanation" in list(data[_key].keys())
+
+def process_file(_file, request_dict):
+    client = OpenAI()
+    with open(f'batch_requests/{base_dir}/{_file}.json', 'r') as f:
+        data = json.load(f)
+    assert len(data) == len(request_dict)
+    parse_results = {}
+    for i in range(len(data)):
+        dp = data[i]
+        custom_id = dp["custom_id"]
+        try:
+            json_data = json.loads(dp["response"]["body"]["choices"][0]["message"]["content"])
+            check_response_valid(json_data)
+            parse_results[custom_id] = json_data
+        except Exception as e:
+            print(f"{_file} ({i}/{len(data)}) Find error when parsing {custom_id} ({e}), re-request OpenAI")
+            while True:
+                try:
+                    response = client.chat.completions.create(
+                        model=request_dict[custom_id]["model"],
+                        messages=request_dict[custom_id]["messages"],
+                        response_format=request_dict[custom_id]["response_format"]
+                    )
+                    json_data = json.loads(response.choices[0].message.content)
+                    check_response_valid(json_data)
+                    parse_results[custom_id] = json_data
+                    print(f"{_file} ({i}/{len(data)}) success re-request!")
+                    time.sleep(1)
+                    break
+                except Exception as e:
+                    print(f"{_file} ({i}/{len(data)}) {e}")
+                    print(f"{_file} ({i}/{len(data)}) continue re-request OpenAI")
+                    continue
+    with open(f'batch_requests/{base_dir}/{_file}-parse-result.json', 'w') as f:
+        json.dump(parse_results, f, indent=4, ensure_ascii=False)
+
+request_dict = {}
+with open(f'batch_requests/{base_dir}/{request_file}', 'r') as f:
+    for _line in f.readlines():
+        json_data = json.loads(_line)
+        request_dict[json_data["custom_id"]] = {
+            "model": json_data["body"]["model"],
+            "messages": json_data["body"]["messages"],
+            "response_format": json_data["body"]["response_format"]
+        }
+
+thread_list = []
+for _file in result_files:
+    thread = threading.Thread(target=process_file, args=(_file, request_dict))
+    thread_list.append(thread)
+
+for thread in thread_list:
+    thread.setDaemon(True)
+    thread.start()
+    
+for thread in thread_list:
+    thread.join()

reproduce/quantitative_comparison/batch_quant_eval_upload.py

@@ -0,0 +1,172 @@
+import os
+os.environ["OPENAI_API_KEY"] = ""
+import re
+import time
+import json
+import jsonlines
+import tiktoken
+import itertools
+from pydantic import BaseModel, Field
+from typing import Literal
+
+from tqdm import tqdm
+from openai import OpenAI
+from openai.lib._pydantic import to_strict_json_schema
+from openai.lib._parsing._completions import type_to_response_format_param
+
+encoding = tiktoken.encoding_for_model('gpt-4o-mini')
+
+sys_prompt = """
+---Role---
+You are an expert evaluating an answer against a baseline answer based on these criteria: **Comprehensiveness**, **Empowerment**, **Trustworthiness**, **Depth** and **Density**.
+"""
+
+prompt = """
+You are an expert evaluating an answer against a baseline answer based on these criteria: **Comprehensiveness**, **Empowerment**, **Trustworthiness**, **Depth** and **Density**.
+
+- **Comprehensiveness**: How much detail does the answer provide to cover all aspects and details of the question?
+- **Empowerment**: How well does the answer help the reader understand and make informed judgments about the topic?
+- **Trustworthiness**: Does the answer provide sufficient detail and align with common knowledge, enhancing its credibility?
+- **Depth**: Does the answer provide in-depth analysis or details, rather than just superficial information?
+- **Density**: Does the answer contain relevant information without less informative or redundant content?
+
+For the evaluated answer labeled "Evaluation Answer," assign a score from 1 to 5 for each criterion compared to the baseline answer labeled "Baseline Answer." Then, assign an overall score based on these criteria.
+The evaluation scores are defined as follows:
+- 1: Strongly worse than the baseline answer
+- 2: Weakly worse than the baseline answer
+- 3: Moderate compared to the baseline answer
+- 4: Weakly better than the baseline answer
+- 5: Strongly better than the baseline answer
+
+
+Here is the question:
+{query}
+
+Here are the answers:
+
+**Baseline Answer:**
+{baseline_answer}
+
+**Evaluation Answer:**
+{evaluation_answer}
+
+
+Evaluate the answer using the criteria listed above and provide detailed explanations for the scores.
+
+Output your evaluation in the following JSON format:
+
+{{
+    "Comprehensiveness": {{
+        "Score": "[1 - 5]",
+        "Explanation": "[Provide explanation here]"
+    }},
+    "Empowerment": {{
+        "Score": "[1 - 5]",
+        "Explanation": "[Provide explanation here]"
+    }},
+    "Trustworthiness": {{
+        "Score": "[1 - 5]",
+        "Explanation": "[Provide explanation here]"
+    }},
+    "Depth": {{
+        "Score": "[1 - 5]",
+        "Explanation": "[Provide explanation here]"
+    }},
+    "Density": {{
+        "Score": "[1 - 5]",
+        "Explanation": "[Provide explanation here]"
+    }}
+    "Overall Score": {{
+        "Score": "[1 - 5]",
+        "Explanation": "[Provide explanation here]"
+    }}
+}}
+"""
+
+class Criterion(BaseModel):
+    Score: int
+    Explanation: str
+
+class Result(BaseModel):
+    Comprehensiveness: Criterion
+    Empowerment: Criterion
+    Trustworthiness: Criterion
+    Depth: Criterion
+    Density: Criterion
+    Overall_Score: Criterion = Field(alias="Overall Score")
+
+result_response_format = type_to_response_format_param(Result)
+
+if __name__ == "__main__":
+    with open('../../longervideos/dataset.json', 'r') as f:
+        questions = json.load(f)
+        
+    baseline_answer_dir = 'answers-naiverag'
+    base_dir = 'overall_comparison_video_understanding'
+    evaluation_answer_dir = [ 
+        'answers-videorag',
+        'answers-notebooklm',
+        'answers-llamavid',
+        'answers-videoagent'
+    ]
+    
+    requests = []
+    total_token_count = 0
+    for _id in questions:
+        video_list_name = questions[_id][0]['description']
+        video_querys = questions[_id][0]['questions']
+        data_path = f"../all_answers/{_id}-{video_list_name}"
+        for _evaluation_answer_dir in evaluation_answer_dir:
+            baseline_work_dir = os.path.join(data_path, baseline_answer_dir)
+            evaluation_work_dir = os.path.join(data_path, _evaluation_answer_dir)
+            for i in range(len(questions[_id][0]['questions'])):
+                # query
+                query_id = questions[_id][0]['questions'][i]["id"]
+                query = questions[_id][0]['questions'][i]["question"]
+                # baseline answer
+                with open(os.path.join(baseline_work_dir, f'answer_{query_id}.md'), 'r') as f:
+                    baseline_answer = f.read()
+                # evaluation answer
+                with open(os.path.join(evaluation_work_dir, f'answer_{query_id}.md'), 'r') as f:
+                    evaluation_answer = f.read()
+                request_prompt = prompt.format(query=query, baseline_answer=baseline_answer, evaluation_answer=evaluation_answer)
+                
+                request_data = {
+                    "custom_id": f"{_id}-{video_list_name}++query{query_id}++base++{baseline_answer_dir}++evaluate++{_evaluation_answer_dir}",
+                    "method": "POST",
+                    "url": "/v1/chat/completions",
+                    "body": {
+                        "model": "gpt-4o-mini",
+                        "messages": [
+                            {"role": "system", "content": sys_prompt},
+                            {"role": "user", "content": request_prompt},
+                        ],
+                        "response_format": result_response_format
+                    },
+                }
+                requests.append(request_data)
+                total_token_count += len(encoding.encode(request_prompt))
+    
+    run_time = 5
+    os.makedirs(f'batch_requests/{base_dir}', exist_ok=True)
+    request_json_file_path = f'batch_requests/{base_dir}/{int(time.time())}.json'
+    with jsonlines.open(request_json_file_path, mode="w") as writer:
+        for request in requests:
+            writer.write(request)
+    print(f"Batch API requests written to {request_json_file_path}")
+    print(f"Price: {total_token_count / 1000000 * 0.075 * run_time}$")
+    
+    for k in range(run_time):
+        client = OpenAI()
+        batch_input_file = client.files.create(
+            file=open(request_json_file_path, "rb"), purpose="batch"
+        )
+        batch_input_file_id = batch_input_file.id
+        
+        batch = client.batches.create(
+            input_file_id=batch_input_file_id,
+            endpoint="/v1/chat/completions",
+            completion_window="24h",
+            metadata={"description": f"runtime{k} - a very nice and successful eval job: {request_json_file_path}"},
+        )
+        print(f"RunTime {k}: Batch {batch.id} has been created.")

reproduce/winrate_comparison/batch_winrate_eval_calculate.py

@@ -0,0 +1,99 @@
+import json
+from tqdm import tqdm
+
+# the model_a in fixed as videorag
+model_a = 'videorag'
+# pick the model_b from ['naiverag', 'graphrag-local', 'graphrag-global', 'lightrag-hybrid']
+model_b = 'naiverag'
+
+metrics = ['Comprehensiveness', 'Empowerment', 'Trustworthiness', 'Depth', 'Density', 'Overall Winner']
+
+base_dir = 'overall_comparison_rag'
+# Please enter the parsed result files ending with .json below.
+result_file = [
+    '',
+    '',
+    '',
+    '',
+    ''
+]
+
+domain_list = ['lecture', 'documentary', 'entertainment']
+
+with open('../../longervideos/dataset.json', 'r') as f:
+    all_data = json.load(f)
+
+overall_win_count = {}
+for _metric in metrics:
+    overall_win_count[_metric] = {'a': 0, 'b': 0}
+
+category_domain_dict = {}
+for category_id in all_data:
+    _domain = all_data[category_id][0]['type']
+    category_domain_dict[category_id] = _domain
+        
+domain_win_count = {}
+for domain in domain_list:
+    domain_win_count[domain] = {}
+    for _metric in metrics:
+        domain_win_count[domain][_metric] = {'a': 0, 'b': 0}
+
+query_count = 0 
+for category_id in tqdm(all_data):
+    category = f"{category_id}-{all_data[category_id][0]['description']}"
+    querys = all_data[category_id][0]['questions']
+    query_count += len(querys)
+    win_count = {}
+    for _metric in metrics:
+        win_count[_metric] = {'a': 0, 'b': 0}
+    
+    for _file in result_file:
+        result_path = f'./batch_requests/{base_dir}/{_file}'
+        with open(result_path, 'r') as f:
+            results = json.loads(f.read())
+        
+        for i in range(len(querys)):
+            query_id = querys[i]['id']
+            ori_result = results[f'{category}++query{query_id}++answers-{model_a}++answers-{model_b}++ori']
+            rev_result = results[f'{category}++query{query_id}++answers-{model_b}++answers-{model_a}++rev']
+            assert ori_result[_metric]['Winner'] in ['Answer 1', 'Answer 2']
+            # original order
+            for _metric in metrics:
+                winner = 'a' if ('1' in ori_result[_metric]['Winner']) else 'b'
+                win_count[_metric][winner] += 1
+                domain_win_count[category_domain_dict[category_id]][_metric][winner] += 1
+                overall_win_count[_metric][winner] += 1
+            # reverse order
+            for _metric in metrics:
+                winner = 'b' if ('1' in rev_result[_metric]['Winner']) else 'a'
+                win_count[_metric][winner] += 1
+                domain_win_count[category_domain_dict[category_id]][_metric][winner] += 1
+                overall_win_count[_metric][winner] += 1
+            
+    
+with open(f'batch_requests/{base_dir}/{base_dir}.txt', 'a') as f:
+    print(query_count)
+    print('a', model_a)
+    f.write('a ' + model_a + '\n')
+    print('b', model_b)
+    f.write('b ' + model_b + '\n')
+    for domain in domain_list:
+        print(f'(left) {model_a} : (right) {model_b} \t {domain}')
+        f.write(f'(left) {model_a} : (right) {model_b} \t {domain}' + '\n')
+        for _metric in metrics:
+            total_count = domain_win_count[domain][_metric]['a'] + domain_win_count[domain][_metric]['b']
+            win_a_percentage = (domain_win_count[domain][_metric]['a'] / total_count) * 100
+            win_b_percentage = (domain_win_count[domain][_metric]['b'] / total_count) * 100
+            print(f'{win_a_percentage:.2f}% : {win_b_percentage:.2f}%', domain_win_count[domain][_metric], _metric)
+            f.write(f'{win_a_percentage:.2f}% : {win_b_percentage:.2f}% {domain_win_count[domain][_metric]} {_metric} \n')
+        print('----'*8)
+        f.write('----'*8 + '\n')
+    print(f'(left) {model_a} : (right) {model_b} \t overall comparision')
+    f.write(f'(left) {model_a} : (right) {model_b} \t overall comparision\n')
+    for _metric in metrics:
+        total_count = overall_win_count[_metric]['a'] + overall_win_count[_metric]['b']
+        win_a_percentage = (overall_win_count[_metric]['a'] / total_count) * 100
+        win_b_percentage = (overall_win_count[_metric]['b'] / total_count) * 100
+        print(f'{win_a_percentage:.2f}% : {win_b_percentage:.2f}%', overall_win_count[_metric], _metric)
+        f.write(f'{win_a_percentage:.2f}% : {win_b_percentage:.2f}% {overall_win_count[_metric]} {_metric} \n')
+    f.write('====' * 8 + '\n\n')

reproduce/winrate_comparison/batch_winrate_eval_download.py

@@ -0,0 +1,53 @@
+import os
+os.environ["OPENAI_API_KEY"] = ""
+import re
+import time
+import json
+import jsonlines
+import tiktoken
+
+from tqdm import tqdm
+from openai import OpenAI
+
+client = OpenAI()
+
+def obtain_ouput_file_id(batches):
+    for batch in batches:
+        print(client.batches.retrieve(batch))
+        print(client.batches.retrieve(batch).output_file_id)
+
+def download_result(result_files, base_dir):
+    for _file in result_files:
+        content = client.files.content(_file).content
+        with open(f"batch_requests/{base_dir}/{_file}.temp", "wb") as f:
+            f.write(content)
+        results = []
+        with open(f"batch_requests/{base_dir}/{_file}.temp", 'r') as f:
+            for line in tqdm(f):
+                json_object = json.loads(line.strip())
+                results.append(json_object)
+        with open(f"batch_requests/{base_dir}/{_file}.json", "w") as json_file:
+            json.dump(results, json_file, indent=4)
+        os.remove(f"batch_requests/{base_dir}/{_file}.temp")
+
+# ================================
+
+# Please enter the relevant batch ID here to obtain the output file ID.
+batches = [
+    '',
+    '',
+    '',
+    '',
+    ''
+]
+obtain_ouput_file_id(batches)
+
+# Second Step: Please enter the output file ID below to download the output files.
+# result_files = [
+#     '',
+#     '',
+#     '',
+#     '',
+#     ''
+# ]
+# download_result(result_files, base_dir='overall_comparison_rag')

reproduce/winrate_comparison/batch_winrate_eval_parse.py

@@ -0,0 +1,89 @@
+import os
+os.environ["OPENAI_API_KEY"] = ""
+import time
+import json
+import threading
+from tqdm import tqdm
+from openai import OpenAI
+from setproctitle import setproctitle
+
+# overall comparison - rag methods
+base_dir = 'overall_comparison_rag'
+# The JSON file contains the batch of requests created when the batch request was uploaded.
+request_file = '' 
+# Please enter the output file ID below, which corresponds to the downloaded output files.
+result_files = [
+    '',
+    '',
+    '',
+    '',
+    ''
+]
+
+setproctitle(f"parse-result-{base_dir}")
+print(f"Start parsing result files in {base_dir}...")
+
+def check_response_valid(data):
+    valid_keys = ['Comprehensiveness', 'Empowerment', 'Trustworthiness', 'Depth', 'Density', 'Overall Winner']
+    assert len(data) == 6
+    assert set(list(data.keys())) == set(valid_keys)
+    for _key in valid_keys:
+        assert data[_key]["Winner"] in ['Answer 1', 'Answer 2']
+        assert "Explanation" in list(data[_key].keys())
+
+def process_file(_file, request_dict):
+    client = OpenAI()
+    with open(f'batch_requests/{base_dir}/{_file}.json', 'r') as f:
+        data = json.load(f)
+    assert len(data) == len(request_dict)
+    parse_results = {}
+    for i in range(len(data)):
+        dp = data[i]
+        custom_id = dp["custom_id"]
+        try:
+            json_data = json.loads(dp["response"]["body"]["choices"][0]["message"]["content"])
+            check_response_valid(json_data)
+            parse_results[custom_id] = json_data
+        except Exception as e:
+            print(f"{_file} ({i}/{len(data)}) Find error when parsing {custom_id} ({e}), re-request OpenAI")
+            while True:
+                try:
+                    response = client.chat.completions.create(
+                        model=request_dict[custom_id]["model"],
+                        messages=request_dict[custom_id]["messages"],
+                        response_format=request_dict[custom_id]["response_format"]
+                    )
+                    json_data = json.loads(response.choices[0].message.content)
+                    check_response_valid(json_data)
+                    parse_results[custom_id] = json_data
+                    print(f"{_file} ({i}/{len(data)}) success re-request!")
+                    time.sleep(1)
+                    break
+                except Exception as e:
+                    print(f"{_file} ({i}/{len(data)}) {e}")
+                    print(f"{_file} ({i}/{len(data)}) continue re-request OpenAI")
+                    continue
+    with open(f'batch_requests/{base_dir}/{_file}-parse-result.json', 'w') as f:
+        json.dump(parse_results, f, indent=4, ensure_ascii=False)
+
+request_dict = {}
+with open(f'batch_requests/{base_dir}/{request_file}', 'r') as f:
+    for _line in f.readlines():
+        json_data = json.loads(_line)
+        request_dict[json_data["custom_id"]] = {
+            "model": json_data["body"]["model"],
+            "messages": json_data["body"]["messages"],
+            "response_format": json_data["body"]["response_format"]
+        }
+
+thread_list = []
+for _file in result_files:
+    thread = threading.Thread(target=process_file, args=(_file, request_dict))
+    thread_list.append(thread)
+
+for thread in thread_list:
+    thread.setDaemon(True)
+    thread.start()
+    
+for thread in thread_list:
+    thread.join()

reproduce/winrate_comparison/batch_winrate_eval_upload.py

@@ -0,0 +1,181 @@
+import os
+os.environ["OPENAI_API_KEY"] = ""
+import re
+import time
+import json
+import jsonlines
+import tiktoken
+from pydantic import BaseModel, Field
+from typing import Literal
+
+from tqdm import tqdm
+from openai import OpenAI
+from openai.lib._pydantic import to_strict_json_schema
+from openai.lib._parsing._completions import type_to_response_format_param
+
+encoding = tiktoken.encoding_for_model('gpt-4o-mini')
+
+sys_prompt = """
+---Role---
+You are an expert tasked with evaluating two answers to the same question based on these criteria: **Comprehensiveness**, **Empowerment**, **Trustworthiness**, **Depth** and **Density**.
+"""
+
+prompt = """
+You will evaluate two answers to the same question based on these criteria: **Comprehensiveness**, **Empowerment**, **Trustworthiness**, **Depth** and **Density**.
+
+- **Comprehensiveness**: How much detail does the answer provide to cover all aspects and details of the question?
+- **Empowerment**: How well does the answer help the reader understand and make informed judgments about the topic?
+- **Trustworthiness**: Does the answer provide sufficient detail and align with common knowledge, enhancing its credibility?
+- **Depth**: Does the answer provide in-depth analysis or details, rather than just superficial information?
+- **Density**: Does the answer contain relevant information without less informative or redundant content?
+
+For each criterion, choose the better answer (either Answer 1 or Answer 2) and explain why. Then, select an overall winner based on these criteria.
+
+Here is the question:
+{query}
+
+Here are the two answers:
+
+**Answer 1:**
+{answer1}
+
+**Answer 2:**
+{answer2}
+
+Evaluate both answers using the criteria listed above and provide detailed explanations for each criterion.
+
+Output your evaluation in the following JSON format:
+
+{{
+    "Comprehensiveness": {{
+        "Winner": "[Answer 1 or Answer 2]",
+        "Explanation": "[Provide explanation here]"
+    }},
+    "Empowerment": {{
+        "Winner": "[Answer 1 or Answer 2]",
+        "Explanation": "[Provide explanation here]"
+    }},
+    "Trustworthiness": {{
+        "Winner": "[Answer 1 or Answer 2]",
+        "Explanation": "[Provide explanation here]"
+    }},
+    "Depth": {{
+        "Winner": "[Answer 1 or Answer 2]",
+        "Explanation": "[Provide explanation here]"
+    }},
+    "Density": {{
+        "Winner": "[Answer 1 or Answer 2]",
+        "Explanation": "[Provide explanation here]"
+    }},
+    "Overall Winner": {{
+        "Winner": "[Answer 1 or Answer 2]",
+        "Explanation": "[Summarize why this answer is the overall winner on the above criteria]"
+    }}
+}}
+"""
+
+class Criterion(BaseModel):
+    Winner: Literal["Answer 1", "Answer 2"]
+    Explanation: str
+
+class Result(BaseModel):
+    Comprehensiveness: Criterion
+    Empowerment: Criterion
+    Trustworthiness: Criterion
+    Depth: Criterion
+    Density: Criterion
+    Overall_Winner: Criterion = Field(alias="Overall Winner")
+
+result_response_format = type_to_response_format_param(Result)
+
+if __name__ == "__main__":
+    with open('../../longervideos/dataset.json', 'r') as f:
+        questions = json.load(f)
+    our_answer_dir = 'answers-videorag'
+    
+    # overall comparsion - rag
+    base_dir = 'overall_comparison_rag'
+    com_answer_dir = [
+        'answers-naiverag', 
+        'answers-graphrag-local', 
+        'answers-graphrag-global',
+        'answers-lightrag-hybrid',
+    ]
+    
+    requests = []
+    total_token_count = 0
+    for _id in questions:
+        video_list_name = questions[_id][0]['description']
+        video_querys = questions[_id][0]['questions']
+        data_path = f"../all_answers/{_id}-{video_list_name}"
+        for _com_answer_dir in com_answer_dir:
+            our_work_dir = os.path.join(data_path, our_answer_dir)
+            com_work_dir = os.path.join(data_path, _com_answer_dir)
+            for i in range(len(questions[_id][0]['questions'])):
+                # query
+                query_id = questions[_id][0]['questions'][i]["id"]
+                query = questions[_id][0]['questions'][i]["question"]
+                # our answer
+                with open(os.path.join(our_work_dir, f'answer_{query_id}.md'), 'r') as f:
+                    our_answer = f.read()
+                # com answer
+                with open(os.path.join(com_work_dir, f'answer_{query_id}.md'), 'r') as f:
+                    com_answer = f.read()
+                ori_prompt = prompt.format(query=query, answer1=our_answer, answer2=com_answer)
+                rev_prompt = prompt.format(query=query, answer1=com_answer, answer2=our_answer)
+                
+                ori_request_data = {
+                    "custom_id": f"{_id}-{video_list_name}++query{query_id}++{our_answer_dir}++{_com_answer_dir}++ori",
+                    "method": "POST",
+                    "url": "/v1/chat/completions",
+                    "body": {
+                        "model": "gpt-4o-mini",
+                        "messages": [
+                            {"role": "system", "content": sys_prompt},
+                            {"role": "user", "content": ori_prompt},
+                        ],
+                        "response_format": result_response_format
+                    },
+                }
+                rev_request_data = {
+                    "custom_id": f"{_id}-{video_list_name}++query{query_id}++{_com_answer_dir}++{our_answer_dir}++rev",
+                    "method": "POST",
+                    "url": "/v1/chat/completions",
+                    "body": {
+                        "model": "gpt-4o-mini",
+                        "messages": [
+                            {"role": "system", "content": sys_prompt},
+                            {"role": "user", "content": rev_prompt},
+                        ],
+                        "response_format": result_response_format
+                    },
+                }
+                requests.append(ori_request_data)
+                requests.append(rev_request_data)
+                
+                total_token_count += len(encoding.encode(ori_prompt))
+                total_token_count += len(encoding.encode(rev_prompt))
+    
+    run_time = 5
+    os.makedirs(f'batch_requests/{base_dir}', exist_ok=True)
+    request_json_file_path = f'batch_requests/{base_dir}/{int(time.time())}.json'
+    with jsonlines.open(request_json_file_path, mode="w") as writer:
+        for request in requests:
+            writer.write(request)
+    print(f"Batch API requests written to {request_json_file_path}")
+    print(f"Price: {total_token_count / 1000000 * 0.075 * run_time}$")
+    
+    for k in range(run_time):
+        client = OpenAI()
+        batch_input_file = client.files.create(
+            file=open(request_json_file_path, "rb"), purpose="batch"
+        )
+        batch_input_file_id = batch_input_file.id
+        
+        batch = client.batches.create(
+            input_file_id=batch_input_file_id,
+            endpoint="/v1/chat/completions",
+            completion_window="24h",
+            metadata={"description": f"runtime{k}: {request_json_file_path}"},
+        )
+        print(f"RunTime {k}: Batch {batch.id} has been created.")

videorag/__init__.py

@@ -0,0 +1 @@
+from .videorag import VideoRAG, QueryParam

videorag/_llm.py

@@ -0,0 +1,178 @@
+import numpy as np
+
+from openai import AsyncOpenAI, AsyncAzureOpenAI, APIConnectionError, RateLimitError
+
+from tenacity import (
+    retry,
+    stop_after_attempt,
+    wait_exponential,
+    retry_if_exception_type,
+)
+import os
+
+from ._utils import compute_args_hash, wrap_embedding_func_with_attrs
+from .base import BaseKVStorage
+
+global_openai_async_client = None
+global_azure_openai_async_client = None
+
+
+def get_openai_async_client_instance():
+    global global_openai_async_client
+    if global_openai_async_client is None:
+        global_openai_async_client = AsyncOpenAI()
+    return global_openai_async_client
+
+
+def get_azure_openai_async_client_instance():
+    global global_azure_openai_async_client
+    if global_azure_openai_async_client is None:
+        global_azure_openai_async_client = AsyncAzureOpenAI()
+    return global_azure_openai_async_client
+
+
+@retry(
+    stop=stop_after_attempt(5),
+    wait=wait_exponential(multiplier=1, min=4, max=10),
+    retry=retry_if_exception_type((RateLimitError, APIConnectionError)),
+)
+async def openai_complete_if_cache(
+    model, prompt, system_prompt=None, history_messages=[], **kwargs
+) -> str:
+    openai_async_client = get_openai_async_client_instance()
+    hashing_kv: BaseKVStorage = kwargs.pop("hashing_kv", None)
+    messages = []
+    if system_prompt:
+        messages.append({"role": "system", "content": system_prompt})
+    messages.extend(history_messages)
+    messages.append({"role": "user", "content": prompt})
+    if hashing_kv is not None:
+        args_hash = compute_args_hash(model, messages)
+        if_cache_return = await hashing_kv.get_by_id(args_hash)
+        if if_cache_return is not None:
+            return if_cache_return["return"]
+
+    response = await openai_async_client.chat.completions.create(
+        model=model, messages=messages, **kwargs
+    )
+
+    if hashing_kv is not None:
+        await hashing_kv.upsert(
+            {args_hash: {"return": response.choices[0].message.content, "model": model}}
+        )
+        await hashing_kv.index_done_callback()
+    return response.choices[0].message.content
+
+
+async def gpt_4o_complete(
+    prompt, system_prompt=None, history_messages=[], **kwargs
+) -> str:
+    return await openai_complete_if_cache(
+        "gpt-4o",
+        prompt,
+        system_prompt=system_prompt,
+        history_messages=history_messages,
+        **kwargs,
+    )
+
+
+async def gpt_4o_mini_complete(
+    prompt, system_prompt=None, history_messages=[], **kwargs
+) -> str:
+    return await openai_complete_if_cache(
+        "gpt-4o-mini",
+        prompt,
+        system_prompt=system_prompt,
+        history_messages=history_messages,
+        **kwargs,
+    )
+
+
+@wrap_embedding_func_with_attrs(embedding_dim=1536, max_token_size=8192)
+@retry(
+    stop=stop_after_attempt(5),
+    wait=wait_exponential(multiplier=1, min=4, max=10),
+    retry=retry_if_exception_type((RateLimitError, APIConnectionError)),
+)
+async def openai_embedding(texts: list[str]) -> np.ndarray:
+    openai_async_client = get_openai_async_client_instance()
+    response = await openai_async_client.embeddings.create(
+        model="text-embedding-3-small", input=texts, encoding_format="float"
+    )
+    return np.array([dp.embedding for dp in response.data])
+
+
+@retry(
+    stop=stop_after_attempt(3),
+    wait=wait_exponential(multiplier=1, min=4, max=10),
+    retry=retry_if_exception_type((RateLimitError, APIConnectionError)),
+)
+async def azure_openai_complete_if_cache(
+    deployment_name, prompt, system_prompt=None, history_messages=[], **kwargs
+) -> str:
+    azure_openai_client = get_azure_openai_async_client_instance()
+    hashing_kv: BaseKVStorage = kwargs.pop("hashing_kv", None)
+    messages = []
+    if system_prompt:
+        messages.append({"role": "system", "content": system_prompt})
+    messages.extend(history_messages)
+    messages.append({"role": "user", "content": prompt})
+    if hashing_kv is not None:
+        args_hash = compute_args_hash(deployment_name, messages)
+        if_cache_return = await hashing_kv.get_by_id(args_hash)
+        if if_cache_return is not None:
+            return if_cache_return["return"]
+
+    response = await azure_openai_client.chat.completions.create(
+        model=deployment_name, messages=messages, **kwargs
+    )
+
+    if hashing_kv is not None:
+        await hashing_kv.upsert(
+            {
+                args_hash: {
+                    "return": response.choices[0].message.content,
+                    "model": deployment_name,
+                }
+            }
+        )
+        await hashing_kv.index_done_callback()
+    return response.choices[0].message.content
+
+
+async def azure_gpt_4o_complete(
+    prompt, system_prompt=None, history_messages=[], **kwargs
+) -> str:
+    return await azure_openai_complete_if_cache(
+        "gpt-4o",
+        prompt,
+        system_prompt=system_prompt,
+        history_messages=history_messages,
+        **kwargs,
+    )
+
+
+async def azure_gpt_4o_mini_complete(
+    prompt, system_prompt=None, history_messages=[], **kwargs
+) -> str:
+    return await azure_openai_complete_if_cache(
+        "gpt-4o-mini",
+        prompt,
+        system_prompt=system_prompt,
+        history_messages=history_messages,
+        **kwargs,
+    )
+
+
+@wrap_embedding_func_with_attrs(embedding_dim=1536, max_token_size=8192)
+@retry(
+    stop=stop_after_attempt(3),
+    wait=wait_exponential(multiplier=1, min=4, max=10),
+    retry=retry_if_exception_type((RateLimitError, APIConnectionError)),
+)
+async def azure_openai_embedding(texts: list[str]) -> np.ndarray:
+    azure_openai_client = get_azure_openai_async_client_instance()
+    response = await azure_openai_client.embeddings.create(
+        model="text-embedding-3-small", input=texts, encoding_format="float"
+    )
+    return np.array([dp.embedding for dp in response.data])

videorag/_op.py

@@ -0,0 +1,744 @@
+import re
+import json
+import openai
+import asyncio
+import tiktoken
+from typing import Union
+from collections import Counter, defaultdict
+from ._splitter import SeparatorSplitter
+from ._utils import (
+    logger,
+    clean_str,
+    compute_mdhash_id,
+    decode_tokens_by_tiktoken,
+    encode_string_by_tiktoken,
+    is_float_regex,
+    list_of_list_to_csv,
+    pack_user_ass_to_openai_messages,
+    split_string_by_multi_markers,
+    truncate_list_by_token_size,
+)
+from .base import (
+    BaseGraphStorage,
+    BaseKVStorage,
+    BaseVectorStorage,
+    SingleCommunitySchema,
+    CommunitySchema,
+    TextChunkSchema,
+    QueryParam,
+)
+from .prompt import GRAPH_FIELD_SEP, PROMPTS
+from ._videoutil import (
+    retrieved_segment_caption,
+)
+
+def chunking_by_token_size(
+    tokens_list: list[list[int]],
+    doc_keys,
+    tiktoken_model,
+    overlap_token_size=128,
+    max_token_size=1024,
+):
+
+    results = []
+    for index, tokens in enumerate(tokens_list):
+        chunk_token = []
+        lengths = []
+        for start in range(0, len(tokens), max_token_size - overlap_token_size):
+
+            chunk_token.append(tokens[start : start + max_token_size])
+            lengths.append(min(max_token_size, len(tokens) - start))
+
+        # here somehow tricky, since the whole chunk tokens is list[list[list[int]]] for corpus(doc(chunk)),so it can't be decode entirely
+        chunk_token = tiktoken_model.decode_batch(chunk_token)
+        for i, chunk in enumerate(chunk_token):
+
+            results.append(
+                {
+                    "tokens": lengths[i],
+                    "content": chunk.strip(),
+                    "chunk_order_index": i,
+                    "full_doc_id": doc_keys[index],
+                }
+            )
+
+    return results
+
+
+def chunking_by_video_segments(
+    tokens_list: list[list[int]],
+    doc_keys,
+    tiktoken_model,
+    max_token_size=1024,
+):
+    # make sure each segment is not larger than max_token_size
+    for index in range(len(tokens_list)):
+        if len(tokens_list[index]) > max_token_size:
+            tokens_list[index] = tokens_list[index][:max_token_size]
+    
+    results = []
+    chunk_token = []
+    chunk_segment_ids = []
+    chunk_order_index = 0
+    for index, tokens in enumerate(tokens_list):
+        
+        if len(chunk_token) + len(tokens) <= max_token_size:
+            # add new segment
+            chunk_token += tokens.copy()
+            chunk_segment_ids.append(doc_keys[index])
+        else:
+            # save the current chunk
+            chunk = tiktoken_model.decode(chunk_token)
+            results.append(
+                {
+                    "tokens": len(chunk_token),
+                    "content": chunk.strip(),
+                    "chunk_order_index": chunk_order_index,
+                    "video_segment_id": chunk_segment_ids,
+                }
+            )
+            # new chunk with current segment as begin
+            chunk_token = []
+            chunk_segment_ids = []
+            chunk_token += tokens.copy()
+            chunk_segment_ids.append(doc_keys[index])
+            chunk_order_index += 1
+    
+    # save the last chunk
+    if len(chunk_token) > 0:
+        chunk = tiktoken_model.decode(chunk_token)
+        results.append(
+            {
+                "tokens": len(chunk_token),
+                "content": chunk.strip(),
+                "chunk_order_index": chunk_order_index,
+                "video_segment_id": chunk_segment_ids,
+            }
+        )
+    
+    return results
+    
+    
+def chunking_by_seperators(
+    tokens_list: list[list[int]],
+    doc_keys,
+    tiktoken_model,
+    overlap_token_size=128,
+    max_token_size=1024,
+):
+
+    splitter = SeparatorSplitter(
+        separators=[
+            tiktoken_model.encode(s) for s in PROMPTS["default_text_separator"]
+        ],
+        chunk_size=max_token_size,
+        chunk_overlap=overlap_token_size,
+    )
+    results = []
+    for index, tokens in enumerate(tokens_list):
+        chunk_token = splitter.split_tokens(tokens)
+        lengths = [len(c) for c in chunk_token]
+
+        # here somehow tricky, since the whole chunk tokens is list[list[list[int]]] for corpus(doc(chunk)),so it can't be decode entirely
+        chunk_token = tiktoken_model.decode_batch(chunk_token)
+        for i, chunk in enumerate(chunk_token):
+
+            results.append(
+                {
+                    "tokens": lengths[i],
+                    "content": chunk.strip(),
+                    "chunk_order_index": i,
+                    "full_doc_id": doc_keys[index],
+                }
+            )
+
+    return results
+
+
+def get_chunks(new_videos, chunk_func=chunking_by_video_segments, **chunk_func_params):
+    inserting_chunks = {}
+
+    new_videos_list = list(new_videos.keys())
+    for video_name in new_videos_list:
+        segment_id_list = list(new_videos[video_name].keys())
+        docs = [new_videos[video_name][index]["content"] for index in segment_id_list]
+        doc_keys = [f'{video_name}_{index}' for index in segment_id_list]
+
+        ENCODER = tiktoken.encoding_for_model("gpt-4o")
+        tokens = ENCODER.encode_batch(docs, num_threads=16)
+        chunks = chunk_func(
+            tokens, doc_keys=doc_keys, tiktoken_model=ENCODER, **chunk_func_params
+        )
+
+        for chunk in chunks:
+            inserting_chunks.update(
+                {compute_mdhash_id(chunk["content"], prefix="chunk-"): chunk}
+            )
+
+    return inserting_chunks
+
+
+async def _handle_entity_relation_summary(
+    entity_or_relation_name: str,
+    description: str,
+    global_config: dict,
+) -> str:
+    use_llm_func: callable = global_config["cheap_model_func"]
+    llm_max_tokens = global_config["cheap_model_max_token_size"]
+    tiktoken_model_name = global_config["tiktoken_model_name"]
+    summary_max_tokens = global_config["entity_summary_to_max_tokens"]
+
+    tokens = encode_string_by_tiktoken(description, model_name=tiktoken_model_name)
+    if len(tokens) < summary_max_tokens:  # No need for summary
+        return description
+    prompt_template = PROMPTS["summarize_entity_descriptions"]
+    use_description = decode_tokens_by_tiktoken(
+        tokens[:llm_max_tokens], model_name=tiktoken_model_name
+    )
+    context_base = dict(
+        entity_name=entity_or_relation_name,
+        description_list=use_description.split(GRAPH_FIELD_SEP),
+    )
+    use_prompt = prompt_template.format(**context_base)
+    logger.debug(f"Trigger summary: {entity_or_relation_name}")
+    summary = await use_llm_func(use_prompt, max_tokens=summary_max_tokens)
+    return summary
+
+
+async def _handle_single_entity_extraction(
+    record_attributes: list[str],
+    chunk_key: str,
+):
+    if len(record_attributes) < 4 or record_attributes[0] != '"entity"':
+        return None
+    # add this record as a node in the G
+    entity_name = clean_str(record_attributes[1].upper())
+    if not entity_name.strip():
+        return None
+    entity_type = clean_str(record_attributes[2].upper())
+    entity_description = clean_str(record_attributes[3])
+    entity_source_id = chunk_key
+    return dict(
+        entity_name=entity_name,
+        entity_type=entity_type,
+        description=entity_description,
+        source_id=entity_source_id,
+    )
+
+
+async def _handle_single_relationship_extraction(
+    record_attributes: list[str],
+    chunk_key: str,
+):
+    if len(record_attributes) < 5 or record_attributes[0] != '"relationship"':
+        return None
+    # add this record as edge
+    source = clean_str(record_attributes[1].upper())
+    target = clean_str(record_attributes[2].upper())
+    edge_description = clean_str(record_attributes[3])
+    edge_source_id = chunk_key
+    weight = (
+        float(record_attributes[-1]) if is_float_regex(record_attributes[-1]) else 1.0
+    )
+    return dict(
+        src_id=source,
+        tgt_id=target,
+        weight=weight,
+        description=edge_description,
+        source_id=edge_source_id,
+    )
+
+
+async def _merge_nodes_then_upsert(
+    entity_name: str,
+    nodes_data: list[dict],
+    knowledge_graph_inst: BaseGraphStorage,
+    global_config: dict,
+):
+    already_entitiy_types = []
+    already_source_ids = []
+    already_description = []
+
+    already_node = await knowledge_graph_inst.get_node(entity_name)
+    if already_node is not None:
+        already_entitiy_types.append(already_node["entity_type"])
+        already_source_ids.extend(
+            split_string_by_multi_markers(already_node["source_id"], [GRAPH_FIELD_SEP])
+        )
+        already_description.append(already_node["description"])
+
+    entity_type = sorted(
+        Counter(
+            [dp["entity_type"] for dp in nodes_data] + already_entitiy_types
+        ).items(),
+        key=lambda x: x[1],
+        reverse=True,
+    )[0][0]
+    description = GRAPH_FIELD_SEP.join(
+        sorted(set([dp["description"] for dp in nodes_data] + already_description))
+    )
+    source_id = GRAPH_FIELD_SEP.join(
+        set([dp["source_id"] for dp in nodes_data] + already_source_ids)
+    )
+    description = await _handle_entity_relation_summary(
+        entity_name, description, global_config
+    )
+    node_data = dict(
+        entity_type=entity_type,
+        description=description,
+        source_id=source_id,
+    )
+    await knowledge_graph_inst.upsert_node(
+        entity_name,
+        node_data=node_data,
+    )
+    node_data["entity_name"] = entity_name
+    return node_data
+
+
+async def _merge_edges_then_upsert(
+    src_id: str,
+    tgt_id: str,
+    edges_data: list[dict],
+    knowledge_graph_inst: BaseGraphStorage,
+    global_config: dict,
+):
+    already_weights = []
+    already_source_ids = []
+    already_description = []
+    already_order = []
+    if await knowledge_graph_inst.has_edge(src_id, tgt_id):
+        already_edge = await knowledge_graph_inst.get_edge(src_id, tgt_id)
+        already_weights.append(already_edge["weight"])
+        already_source_ids.extend(
+            split_string_by_multi_markers(already_edge["source_id"], [GRAPH_FIELD_SEP])
+        )
+        already_description.append(already_edge["description"])
+        already_order.append(already_edge.get("order", 1))
+
+    # [numberchiffre]: `Relationship.order` is only returned from DSPy's predictions
+    order = min([dp.get("order", 1) for dp in edges_data] + already_order)
+    weight = sum([dp["weight"] for dp in edges_data] + already_weights)
+    description = GRAPH_FIELD_SEP.join(
+        sorted(set([dp["description"] for dp in edges_data] + already_description))
+    )
+    source_id = GRAPH_FIELD_SEP.join(
+        set([dp["source_id"] for dp in edges_data] + already_source_ids)
+    )
+    for need_insert_id in [src_id, tgt_id]:
+        if not (await knowledge_graph_inst.has_node(need_insert_id)):
+            await knowledge_graph_inst.upsert_node(
+                need_insert_id,
+                node_data={
+                    "source_id": source_id,
+                    "description": description,
+                    "entity_type": '"UNKNOWN"',
+                },
+            )
+    description = await _handle_entity_relation_summary(
+        (src_id, tgt_id), description, global_config
+    )
+    await knowledge_graph_inst.upsert_edge(
+        src_id,
+        tgt_id,
+        edge_data=dict(
+            weight=weight, description=description, source_id=source_id, order=order
+        ),
+    )
+    return_edge_data = dict(
+        src_tgt=(src_id, tgt_id),
+        description=description,
+        weight=weight
+    )
+    return return_edge_data
+
+
+async def extract_entities(
+    chunks: dict[str, TextChunkSchema],
+    knowledge_graph_inst: BaseGraphStorage,
+    entity_vdb: BaseVectorStorage,
+    global_config: dict,
+) -> Union[BaseGraphStorage, None]:
+    use_llm_func: callable = global_config["best_model_func"]
+    entity_extract_max_gleaning = global_config["entity_extract_max_gleaning"]
+    
+    ordered_chunks = list(chunks.items())
+
+    entity_extract_prompt = PROMPTS["entity_extraction"]
+    context_base = dict(
+        tuple_delimiter=PROMPTS["DEFAULT_TUPLE_DELIMITER"],
+        record_delimiter=PROMPTS["DEFAULT_RECORD_DELIMITER"],
+        completion_delimiter=PROMPTS["DEFAULT_COMPLETION_DELIMITER"],
+        entity_types=",".join(PROMPTS["DEFAULT_ENTITY_TYPES"]),
+    )
+    continue_prompt = PROMPTS["entiti_continue_extraction"]
+    if_loop_prompt = PROMPTS["entiti_if_loop_extraction"]
+
+    already_processed = 0
+    already_entities = 0
+    already_relations = 0
+
+    async def _process_single_content(chunk_key_dp: tuple[str, TextChunkSchema]):
+        nonlocal already_processed, already_entities, already_relations
+        chunk_key = chunk_key_dp[0]
+        chunk_dp = chunk_key_dp[1]
+        content = chunk_dp["content"]
+        hint_prompt = entity_extract_prompt.format(**context_base, input_text=content)
+        final_result = await use_llm_func(hint_prompt)
+
+        history = pack_user_ass_to_openai_messages(hint_prompt, final_result)
+        for now_glean_index in range(entity_extract_max_gleaning):
+            glean_result = await use_llm_func(continue_prompt, history_messages=history)
+
+            history += pack_user_ass_to_openai_messages(continue_prompt, glean_result)
+            final_result += glean_result
+            if now_glean_index == entity_extract_max_gleaning - 1:
+                break
+
+            if_loop_result: str = await use_llm_func(
+                if_loop_prompt, history_messages=history
+            )
+            if_loop_result = if_loop_result.strip().strip('"').strip("'").lower()
+            if if_loop_result != "yes":
+                break
+
+        records = split_string_by_multi_markers(
+            final_result,
+            [context_base["record_delimiter"], context_base["completion_delimiter"]],
+        )
+
+        maybe_nodes = defaultdict(list)
+        maybe_edges = defaultdict(list)
+        for record in records:
+            record = re.search(r"\((.*)\)", record)
+            if record is None:
+                continue
+            record = record.group(1)
+            record_attributes = split_string_by_multi_markers(
+                record, [context_base["tuple_delimiter"]]
+            )
+            if_entities = await _handle_single_entity_extraction(
+                record_attributes, chunk_key
+            )
+            if if_entities is not None:
+                maybe_nodes[if_entities["entity_name"]].append(if_entities)
+                continue
+
+            if_relation = await _handle_single_relationship_extraction(
+                record_attributes, chunk_key
+            )
+            if if_relation is not None:
+                maybe_edges[(if_relation["src_id"], if_relation["tgt_id"])].append(
+                    if_relation
+                )
+        already_processed += 1
+        already_entities += len(maybe_nodes)
+        already_relations += len(maybe_edges)
+        now_ticks = PROMPTS["process_tickers"][
+            already_processed % len(PROMPTS["process_tickers"])
+        ]
+        print(
+            f"{now_ticks} Processed {already_processed} chunks, {already_entities} entities(duplicated), {already_relations} relations(duplicated)\r",
+            end="",
+            flush=True,
+        )
+        return dict(maybe_nodes), dict(maybe_edges)
+
+    # use_llm_func is wrapped in ascynio.Semaphore, limiting max_async callings
+    results = await asyncio.gather(
+        *[_process_single_content(c) for c in ordered_chunks]
+    )
+    print()  # clear the progress bar
+    maybe_nodes = defaultdict(list)
+    maybe_edges = defaultdict(list)
+    for m_nodes, m_edges in results:
+        for k, v in m_nodes.items():
+            maybe_nodes[k].extend(v)
+        for k, v in m_edges.items():
+            # it's undirected graph
+            maybe_edges[tuple(sorted(k))].extend(v)
+    all_entities_data = await asyncio.gather(
+        *[
+            _merge_nodes_then_upsert(k, v, knowledge_graph_inst, global_config)
+            for k, v in maybe_nodes.items()
+        ]
+    )
+    all_edges_data = await asyncio.gather(
+        *[
+            _merge_edges_then_upsert(k[0], k[1], v, knowledge_graph_inst, global_config)
+            for k, v in maybe_edges.items()
+        ]
+    )
+    if not len(all_entities_data):
+        logger.warning("Didn't extract any entities, maybe your LLM is not working")
+        return None
+    if entity_vdb is not None:
+        data_for_vdb = {
+            compute_mdhash_id(dp["entity_name"], prefix="ent-"): {
+                "content": dp["entity_name"] + dp["description"],
+                "entity_name": dp["entity_name"],
+            }
+            for dp in all_entities_data
+        }
+        await entity_vdb.upsert(data_for_vdb)
+    return knowledge_graph_inst, all_entities_data, all_edges_data
+
+
+async def _find_most_related_segments_from_entities(
+    topk_chunks: int,
+    node_datas: list[dict],
+    text_chunks_db: BaseKVStorage[TextChunkSchema],
+    knowledge_graph_inst: BaseGraphStorage,
+):
+    text_units = [
+        split_string_by_multi_markers(dp["source_id"], [GRAPH_FIELD_SEP])
+        for dp in node_datas
+    ]
+    edges = await asyncio.gather(
+        *[knowledge_graph_inst.get_node_edges(dp["entity_name"]) for dp in node_datas]
+    )
+    all_one_hop_nodes = set()
+    for this_edges in edges:
+        if not this_edges:
+            continue
+        all_one_hop_nodes.update([e[1] for e in this_edges])
+    all_one_hop_nodes = list(all_one_hop_nodes)
+    all_one_hop_nodes_data = await asyncio.gather(
+        *[knowledge_graph_inst.get_node(e) for e in all_one_hop_nodes]
+    )
+    all_one_hop_text_units_lookup = {
+        k: set(split_string_by_multi_markers(v["source_id"], [GRAPH_FIELD_SEP]))
+        for k, v in zip(all_one_hop_nodes, all_one_hop_nodes_data)
+        if v is not None
+    }
+    all_text_units_lookup = {}
+    for index, (this_text_units, this_edges) in enumerate(zip(text_units, edges)):
+        for c_id in this_text_units:
+            if c_id in all_text_units_lookup:
+                continue
+            relation_counts = 0
+            for e in this_edges:
+                if (
+                    e[1] in all_one_hop_text_units_lookup
+                    and c_id in all_one_hop_text_units_lookup[e[1]]
+                ):
+                    relation_counts += 1
+            all_text_units_lookup[c_id] = {
+                "data": await text_chunks_db.get_by_id(c_id),
+                "order": index,
+                "relation_counts": relation_counts,
+            }
+    if any([v is None for v in all_text_units_lookup.values()]):
+        logger.warning("Text chunks are missing, maybe the storage is damaged")
+    all_text_units = [
+        {"id": k, **v} for k, v in all_text_units_lookup.items() if v is not None
+    ]
+    sorted_text_units = sorted(
+        all_text_units, key=lambda x: -x["relation_counts"]
+    )[:topk_chunks]
+    
+    chunk_related_segments = set()
+    for _chunk_data in sorted_text_units:
+        for s_id in _chunk_data['data']['video_segment_id']:
+            chunk_related_segments.add(s_id)
+    
+    return chunk_related_segments
+
+async def _refine_entity_retrieval_query(
+    query,
+    query_param: QueryParam,
+    global_config: dict,
+):
+    use_llm_func: callable = global_config["cheap_model_func"]
+    query_rewrite_prompt = PROMPTS["query_rewrite_for_entity_retrieval"]
+    query_rewrite_prompt = query_rewrite_prompt.format(input_text=query)
+    final_result = await use_llm_func(query_rewrite_prompt)
+    return final_result
+
+async def _refine_visual_retrieval_query(
+    query,
+    query_param: QueryParam,
+    global_config: dict,
+):
+    use_llm_func: callable = global_config["cheap_model_func"]
+    query_rewrite_prompt = PROMPTS["query_rewrite_for_visual_retrieval"]
+    query_rewrite_prompt = query_rewrite_prompt.format(input_text=query)
+    final_result = await use_llm_func(query_rewrite_prompt)
+    return final_result
+
+async def _extract_keywords_query(
+    query,
+    query_param: QueryParam,
+    global_config: dict,
+):
+    use_llm_func: callable = global_config["cheap_model_func"]
+    keywords_prompt = PROMPTS["keywords_extraction"]
+    keywords_prompt = keywords_prompt.format(input_text=query)
+    final_result = await use_llm_func(keywords_prompt)
+    return final_result
+
+async def videorag_query(
+    query,
+    entities_vdb,
+    text_chunks_db,
+    chunks_vdb,
+    video_path_db,
+    video_segments,
+    video_segment_feature_vdb,
+    knowledge_graph_inst,
+    caption_model,
+    caption_tokenizer,
+    query_param: QueryParam,
+    global_config: dict,
+) -> str:
+    use_model_func = global_config["best_model_func"]
+    query = query
+    
+    # naive chunks
+    results = await chunks_vdb.query(query, top_k=query_param.top_k)
+    if not len(results):
+        return PROMPTS["fail_response"]
+    chunks_ids = [r["id"] for r in results]
+    chunks = await text_chunks_db.get_by_ids(chunks_ids)
+
+    maybe_trun_chunks = truncate_list_by_token_size(
+        chunks,
+        key=lambda x: x["content"],
+        max_token_size=query_param.naive_max_token_for_text_unit,
+    )
+    logger.info(f"Truncate {len(chunks)} to {len(maybe_trun_chunks)} chunks")
+    section = "-----New Chunk-----\n".join([c["content"] for c in maybe_trun_chunks])
+    retreived_chunk_context = section
+    
+    # visual retrieval
+    query_for_entity_retrieval = await _refine_entity_retrieval_query(
+        query,
+        query_param,
+        global_config,
+    )
+    entity_results = await entities_vdb.query(query_for_entity_retrieval, top_k=query_param.top_k)
+    entity_retrieved_segments = set()
+    if len(entity_results):
+        node_datas = await asyncio.gather(
+            *[knowledge_graph_inst.get_node(r["entity_name"]) for r in entity_results]
+        )
+        if not all([n is not None for n in node_datas]):
+            logger.warning("Some nodes are missing, maybe the storage is damaged")
+        node_degrees = await asyncio.gather(
+            *[knowledge_graph_inst.node_degree(r["entity_name"]) for r in entity_results]
+        )
+        node_datas = [
+            {**n, "entity_name": k["entity_name"], "rank": d}
+            for k, n, d in zip(entity_results, node_datas, node_degrees)
+            if n is not None
+        ]
+        entity_retrieved_segments = entity_retrieved_segments.union(await _find_most_related_segments_from_entities(
+            global_config["retrieval_topk_chunks"], node_datas, text_chunks_db, knowledge_graph_inst
+        ))
+    
+    # visual retrieval
+    query_for_visual_retrieval = await _refine_visual_retrieval_query(
+        query,
+        query_param,
+        global_config,
+    )
+    segment_results = await video_segment_feature_vdb.query(query_for_visual_retrieval)
+    visual_retrieved_segments = set()
+    if len(segment_results):
+        for n in segment_results:
+            visual_retrieved_segments.add(n['__id__'])
+    
+    # caption
+    retrieved_segments = list(entity_retrieved_segments.union(visual_retrieved_segments))
+    retrieved_segments = sorted(
+        retrieved_segments,
+        key=lambda x: (
+            '_'.join(x.split('_')[:-1]), # video_name
+            eval(x.split('_')[-1]) # index
+        )
+    )
+    print(query_for_entity_retrieval)
+    print(f"Retrieved Text Segments {entity_retrieved_segments}")
+    print(query_for_visual_retrieval)
+    print(f"Retrieved Visual Segments {visual_retrieved_segments}")
+    
+    already_processed = 0
+    async def _filter_single_segment(knowledge: str, segment_key_dp: tuple[str, str]):
+        nonlocal use_model_func, already_processed
+        segment_key = segment_key_dp[0]
+        segment_content = segment_key_dp[1]
+        filter_prompt = PROMPTS["filtering_segment"]
+        filter_prompt = filter_prompt.format(caption=segment_content, knowledge=knowledge)
+        result = await use_model_func(filter_prompt)
+        already_processed += 1
+        now_ticks = PROMPTS["process_tickers"][
+            already_processed % len(PROMPTS["process_tickers"])
+        ]
+        print(
+            f"{now_ticks} Checked {already_processed} segments\r",
+            end="",
+            flush=True,
+        )
+        return (segment_key, result)
+    
+    rough_captions = {}
+    for s_id in retrieved_segments:
+        video_name = '_'.join(s_id.split('_')[:-1])
+        index = s_id.split('_')[-1]
+        rough_captions[s_id] = video_segments._data[video_name][index]["content"]
+    results = await asyncio.gather(
+        *[_filter_single_segment(query, (s_id, rough_captions[s_id])) for s_id in rough_captions]
+    )
+    remain_segments = [x[0] for x in results if 'yes' in x[1].lower()]
+    print(f"{len(remain_segments)} Video Segments remain after filtering")
+    if len(remain_segments) == 0:
+        print("Since no segments remain after filtering, we utilized all the retrieved segments.")
+        remain_segments = retrieved_segments
+    print(f"Remain segments {remain_segments}")
+    
+    # visual retrieval
+    keywords_for_caption = await _extract_keywords_query(
+        query,
+        query_param,
+        global_config,
+    )
+    print(f"Keywords: {keywords_for_caption}")
+    caption_results = retrieved_segment_caption(
+        caption_model,
+        caption_tokenizer,
+        keywords_for_caption,
+        remain_segments,
+        video_path_db,
+        video_segments,
+        num_sampled_frames=global_config['fine_num_frames_per_segment']
+    )
+
+    ## data table
+    text_units_section_list = [["video_name", "start_time", "end_time", "content"]]
+    for s_id in caption_results:
+        video_name = '_'.join(s_id.split('_')[:-1])
+        index = s_id.split('_')[-1]
+        start_time = eval(video_segments._data[video_name][index]["time"].split('-')[0])
+        end_time = eval(video_segments._data[video_name][index]["time"].split('-')[1])
+        start_time = f"{start_time // 3600}:{(start_time % 3600) // 60}:{start_time % 60}"
+        end_time = f"{end_time // 3600}:{(end_time % 3600) // 60}:{end_time % 60}"
+        text_units_section_list.append([video_name, start_time, end_time, caption_results[s_id]])
+    text_units_context = list_of_list_to_csv(text_units_section_list)
+
+    retreived_video_context = f"\n-----Retrieved Knowledge From Videos-----\n```csv\n{text_units_context}\n```\n"
+    
+    if query_param.wo_reference:
+        sys_prompt_temp = PROMPTS["videorag_response_wo_reference"]
+    else:
+        sys_prompt_temp = PROMPTS["videorag_response"]
+        
+    sys_prompt = sys_prompt_temp.format(
+        video_data=retreived_video_context,
+        chunk_data=retreived_chunk_context,
+        response_type=query_param.response_type
+    )
+    response = await use_model_func(
+        query,
+        system_prompt=sys_prompt,
+    )
+    return response

videorag/_splitter.py

@@ -0,0 +1,94 @@
+from typing import List, Optional, Union, Literal
+
+class SeparatorSplitter:
+    def __init__(
+        self,
+        separators: Optional[List[List[int]]] = None,
+        keep_separator: Union[bool, Literal["start", "end"]] = "end",
+        chunk_size: int = 4000,
+        chunk_overlap: int = 200,
+        length_function: callable = len,
+    ):
+        self._separators = separators or []
+        self._keep_separator = keep_separator
+        self._chunk_size = chunk_size
+        self._chunk_overlap = chunk_overlap
+        self._length_function = length_function
+
+    def split_tokens(self, tokens: List[int]) -> List[List[int]]:
+        splits = self._split_tokens_with_separators(tokens)
+        return self._merge_splits(splits)
+
+    def _split_tokens_with_separators(self, tokens: List[int]) -> List[List[int]]:
+        splits = []
+        current_split = []
+        i = 0
+        while i < len(tokens):
+            separator_found = False
+            for separator in self._separators:
+                if tokens[i:i+len(separator)] == separator:
+                    if self._keep_separator in [True, "end"]:
+                        current_split.extend(separator)
+                    if current_split:
+                        splits.append(current_split)
+                        current_split = []
+                    if self._keep_separator == "start":
+                        current_split.extend(separator)
+                    i += len(separator)
+                    separator_found = True
+                    break
+            if not separator_found:
+                current_split.append(tokens[i])
+                i += 1
+        if current_split:
+            splits.append(current_split)
+        return [s for s in splits if s]
+
+    def _merge_splits(self, splits: List[List[int]]) -> List[List[int]]:
+        if not splits:
+            return []
+
+        merged_splits = []
+        current_chunk = []
+
+        for split in splits:
+            if not current_chunk:
+                current_chunk = split
+            elif self._length_function(current_chunk) + self._length_function(split) <= self._chunk_size:
+                current_chunk.extend(split)
+            else:
+                merged_splits.append(current_chunk)
+                current_chunk = split
+
+        if current_chunk:
+            merged_splits.append(current_chunk)
+
+        if len(merged_splits) == 1 and self._length_function(merged_splits[0]) > self._chunk_size:
+            return self._split_chunk(merged_splits[0])
+
+        if self._chunk_overlap > 0:
+            return self._enforce_overlap(merged_splits)
+        
+        return merged_splits
+
+    def _split_chunk(self, chunk: List[int]) -> List[List[int]]:
+        result = []
+        for i in range(0, len(chunk), self._chunk_size - self._chunk_overlap):
+            new_chunk = chunk[i:i + self._chunk_size]
+            if len(new_chunk) > self._chunk_overlap:  # 只有当 chunk 长度大于 overlap 时才添加
+                result.append(new_chunk)
+        return result
+
+    def _enforce_overlap(self, chunks: List[List[int]]) -> List[List[int]]:
+        result = []
+        for i, chunk in enumerate(chunks):
+            if i == 0:
+                result.append(chunk)
+            else:
+                overlap = chunks[i-1][-self._chunk_overlap:]
+                new_chunk = overlap + chunk
+                if self._length_function(new_chunk) > self._chunk_size:
+                    new_chunk = new_chunk[:self._chunk_size]
+                result.append(new_chunk)
+        return result
+

videorag/_storage/__init__.py

@@ -0,0 +1,5 @@
+from .gdb_networkx import NetworkXStorage
+from .gdb_neo4j import Neo4jStorage
+from .vdb_hnswlib import HNSWVectorStorage
+from .vdb_nanovectordb import NanoVectorDBStorage, NanoVectorDBVideoSegmentStorage
+from .kv_json import JsonKVStorage

videorag/_storage/gdb_neo4j.py

@@ -0,0 +1,330 @@
+import json
+import asyncio
+from collections import defaultdict
+from neo4j import AsyncGraphDatabase
+from dataclasses import dataclass
+from typing import Union
+from ..base import BaseGraphStorage, SingleCommunitySchema
+from .._utils import logger
+from ..prompt import GRAPH_FIELD_SEP
+
+neo4j_lock = asyncio.Lock()
+
+
+def make_path_idable(path):
+    return path.replace(".", "_").replace("/", "__").replace("-", "_")
+
+
+@dataclass
+class Neo4jStorage(BaseGraphStorage):
+    def __post_init__(self):
+        self.neo4j_url = self.global_config["addon_params"].get("neo4j_url", None)
+        self.neo4j_auth = self.global_config["addon_params"].get("neo4j_auth", None)
+        self.namespace = (
+            f"{make_path_idable(self.global_config['working_dir'])}__{self.namespace}"
+        )
+        logger.info(f"Using the label {self.namespace} for Neo4j as identifier")
+        if self.neo4j_url is None or self.neo4j_auth is None:
+            raise ValueError("Missing neo4j_url or neo4j_auth in addon_params")
+        self.async_driver = AsyncGraphDatabase.driver(
+            self.neo4j_url, auth=self.neo4j_auth
+        )
+
+    # async def create_database(self):
+    #     async with self.async_driver.session() as session:
+    #         try:
+    #             constraints = await session.run("SHOW CONSTRAINTS")
+    #             # TODO I don't know why CREATE CONSTRAINT IF NOT EXISTS still trigger error
+    #             # so have to check if the constrain exists
+    #             constrain_exists = False
+
+    #             async for record in constraints:
+    #                 if (
+    #                     self.namespace in record["labelsOrTypes"]
+    #                     and "id" in record["properties"]
+    #                     and record["type"] == "UNIQUENESS"
+    #                 ):
+    #                     constrain_exists = True
+    #                     break
+    #             if not constrain_exists:
+    #                 await session.run(
+    #                     f"CREATE CONSTRAINT FOR (n:{self.namespace}) REQUIRE n.id IS UNIQUE"
+    #                 )
+    #                 logger.info(f"Add constraint for namespace: {self.namespace}")
+
+    #         except Exception as e:
+    #             logger.error(f"Error accessing or setting up the database: {str(e)}")
+    #             raise
+
+    async def _init_workspace(self):
+        await self.async_driver.verify_authentication()
+        await self.async_driver.verify_connectivity()
+        # TODOLater: create database if not exists always cause an error when async
+        # await self.create_database()
+
+    async def index_start_callback(self):
+        logger.info("Init Neo4j workspace")
+        await self._init_workspace()
+
+    async def has_node(self, node_id: str) -> bool:
+        async with self.async_driver.session() as session:
+            result = await session.run(
+                f"MATCH (n:{self.namespace}) WHERE n.id = $node_id RETURN COUNT(n) > 0 AS exists",
+                node_id=node_id,
+            )
+            record = await result.single()
+            return record["exists"] if record else False
+
+    async def has_edge(self, source_node_id: str, target_node_id: str) -> bool:
+        async with self.async_driver.session() as session:
+            result = await session.run(
+                f"MATCH (s:{self.namespace})-[r]->(t:{self.namespace}) "
+                "WHERE s.id = $source_id AND t.id = $target_id "
+                "RETURN COUNT(r) > 0 AS exists",
+                source_id=source_node_id,
+                target_id=target_node_id,
+            )
+            record = await result.single()
+            return record["exists"] if record else False
+
+    async def node_degree(self, node_id: str) -> int:
+        async with self.async_driver.session() as session:
+            result = await session.run(
+                f"MATCH (n:{self.namespace}) WHERE n.id = $node_id "
+                f"RETURN COUNT {{(n)-[]-(:{self.namespace})}} AS degree",
+                node_id=node_id,
+            )
+            record = await result.single()
+            return record["degree"] if record else 0
+
+    async def edge_degree(self, src_id: str, tgt_id: str) -> int:
+        async with self.async_driver.session() as session:
+            result = await session.run(
+                f"MATCH (s:{self.namespace}), (t:{self.namespace}) "
+                "WHERE s.id = $src_id AND t.id = $tgt_id "
+                f"RETURN COUNT {{(s)-[]-(:{self.namespace})}} + COUNT {{(t)-[]-(:{self.namespace})}} AS degree",
+                src_id=src_id,
+                tgt_id=tgt_id,
+            )
+            record = await result.single()
+            return record["degree"] if record else 0
+
+    async def get_node(self, node_id: str) -> Union[dict, None]:
+        async with self.async_driver.session() as session:
+            result = await session.run(
+                f"MATCH (n:{self.namespace}) WHERE n.id = $node_id RETURN properties(n) AS node_data",
+                node_id=node_id,
+            )
+            record = await result.single()
+            raw_node_data = record["node_data"] if record else None
+        if raw_node_data is None:
+            return None
+        raw_node_data["clusters"] = json.dumps(
+            [
+                {
+                    "level": index,
+                    "cluster": cluster_id,
+                }
+                for index, cluster_id in enumerate(
+                    raw_node_data.get("communityIds", [])
+                )
+            ]
+        )
+        return raw_node_data
+
+    async def get_edge(
+        self, source_node_id: str, target_node_id: str
+    ) -> Union[dict, None]:
+        async with self.async_driver.session() as session:
+            result = await session.run(
+                f"MATCH (s:{self.namespace})-[r]->(t:{self.namespace}) "
+                "WHERE s.id = $source_id AND t.id = $target_id "
+                "RETURN properties(r) AS edge_data",
+                source_id=source_node_id,
+                target_id=target_node_id,
+            )
+            record = await result.single()
+            return record["edge_data"] if record else None
+
+    async def get_node_edges(
+        self, source_node_id: str
+    ) -> Union[list[tuple[str, str]], None]:
+        async with self.async_driver.session() as session:
+            result = await session.run(
+                f"MATCH (s:{self.namespace})-[r]->(t:{self.namespace}) WHERE s.id = $source_id "
+                "RETURN s.id AS source, t.id AS target",
+                source_id=source_node_id,
+            )
+            edges = []
+            async for record in result:
+                edges.append((record["source"], record["target"]))
+            return edges
+
+    async def upsert_node(self, node_id: str, node_data: dict[str, str]):
+        node_type = node_data.get("entity_type", "UNKNOWN").strip('"')
+        async with self.async_driver.session() as session:
+            await session.run(
+                f"MERGE (n:{self.namespace}:{node_type} {{id: $node_id}}) "
+                "SET n += $node_data",
+                node_id=node_id,
+                node_data=node_data,
+            )
+
+    async def upsert_edge(
+        self, source_node_id: str, target_node_id: str, edge_data: dict[str, str]
+    ):
+        edge_data.setdefault("weight", 0.0)
+        async with self.async_driver.session() as session:
+            await session.run(
+                f"MATCH (s:{self.namespace}), (t:{self.namespace}) "
+                "WHERE s.id = $source_id AND t.id = $target_id "
+                "MERGE (s)-[r:RELATED]->(t) "  # Added relationship type 'RELATED'
+                "SET r += $edge_data",
+                source_id=source_node_id,
+                target_id=target_node_id,
+                edge_data=edge_data,
+            )
+
+    async def clustering(self, algorithm: str):
+        if algorithm != "leiden":
+            raise ValueError(
+                f"Clustering algorithm {algorithm} not supported in Neo4j implementation"
+            )
+
+        random_seed = self.global_config["graph_cluster_seed"]
+        max_level = self.global_config["max_graph_cluster_size"]
+        async with self.async_driver.session() as session:
+            try:
+                # Project the graph with undirected relationships
+                await session.run(
+                    f"""
+                    CALL gds.graph.project(
+                        'graph_{self.namespace}',
+                        ['{self.namespace}'],
+                        {{
+                            RELATED: {{
+                                orientation: 'UNDIRECTED',
+                                properties: ['weight']
+                            }}
+                        }}
+                    )
+                    """
+                )
+
+                # Run Leiden algorithm
+                result = await session.run(
+                    f"""
+                    CALL gds.leiden.write(
+                        'graph_{self.namespace}',
+                        {{
+                            writeProperty: 'communityIds',
+                            includeIntermediateCommunities: True,
+                            relationshipWeightProperty: "weight",
+                            maxLevels: {max_level},
+                            tolerance: 0.0001,
+                            gamma: 1.0,
+                            theta: 0.01,
+                            randomSeed: {random_seed}
+                        }}
+                    )
+                    YIELD communityCount, modularities;
+                    """
+                )
+                result = await result.single()
+                community_count: int = result["communityCount"]
+                modularities = result["modularities"]
+                logger.info(
+                    f"Performed graph clustering with {community_count} communities and modularities {modularities}"
+                )
+            finally:
+                # Drop the projected graph
+                await session.run(f"CALL gds.graph.drop('graph_{self.namespace}')")
+
+    async def community_schema(self) -> dict[str, SingleCommunitySchema]:
+        results = defaultdict(
+            lambda: dict(
+                level=None,
+                title=None,
+                edges=set(),
+                nodes=set(),
+                chunk_ids=set(),
+                occurrence=0.0,
+                sub_communities=[],
+            )
+        )
+
+        async with self.async_driver.session() as session:
+            # Fetch community data
+            result = await session.run(
+                f"""
+                MATCH (n:{self.namespace})
+                WITH n, n.communityIds AS communityIds, [(n)-[]-(m:{self.namespace}) | m.id] AS connected_nodes
+                RETURN n.id AS node_id, n.source_id AS source_id, 
+                       communityIds AS cluster_key,
+                       connected_nodes
+                """
+            )
+
+            # records = await result.fetch()
+
+            max_num_ids = 0
+            async for record in result:
+                for index, c_id in enumerate(record["cluster_key"]):
+                    node_id = str(record["node_id"])
+                    source_id = record["source_id"]
+                    level = index
+                    cluster_key = str(c_id)
+                    connected_nodes = record["connected_nodes"]
+
+                    results[cluster_key]["level"] = level
+                    results[cluster_key]["title"] = f"Cluster {cluster_key}"
+                    results[cluster_key]["nodes"].add(node_id)
+                    results[cluster_key]["edges"].update(
+                        [
+                            tuple(sorted([node_id, str(connected)]))
+                            for connected in connected_nodes
+                            if connected != node_id
+                        ]
+                    )
+                    chunk_ids = source_id.split(GRAPH_FIELD_SEP)
+                    results[cluster_key]["chunk_ids"].update(chunk_ids)
+                    max_num_ids = max(
+                        max_num_ids, len(results[cluster_key]["chunk_ids"])
+                    )
+
+            # Process results
+            for k, v in results.items():
+                v["edges"] = [list(e) for e in v["edges"]]
+                v["nodes"] = list(v["nodes"])
+                v["chunk_ids"] = list(v["chunk_ids"])
+                v["occurrence"] = len(v["chunk_ids"]) / max_num_ids
+
+            # Compute sub-communities (this is a simplified approach)
+            for cluster in results.values():
+                cluster["sub_communities"] = [
+                    sub_key
+                    for sub_key, sub_cluster in results.items()
+                    if sub_cluster["level"] > cluster["level"]
+                    and set(sub_cluster["nodes"]).issubset(set(cluster["nodes"]))
+                ]
+
+        return dict(results)
+
+    async def index_done_callback(self):
+        await self.async_driver.close()
+
+    async def _debug_delete_all_node_edges(self):
+        async with self.async_driver.session() as session:
+            try:
+                # Delete all relationships in the namespace
+                await session.run(f"MATCH (n:{self.namespace})-[r]-() DELETE r")
+
+                # Delete all nodes in the namespace
+                await session.run(f"MATCH (n:{self.namespace}) DELETE n")
+
+                logger.info(
+                    f"All nodes and edges in namespace '{self.namespace}' have been deleted."
+                )
+            except Exception as e:
+                logger.error(f"Error deleting nodes and edges: {str(e)}")
+                raise

videorag/_storage/gdb_networkx.py

@@ -0,0 +1,238 @@
+import html
+import json
+import os
+from collections import defaultdict
+from dataclasses import dataclass
+from typing import Any, Union, cast
+import networkx as nx
+import numpy as np
+
+from .._utils import logger
+from ..base import (
+    BaseGraphStorage,
+    SingleCommunitySchema,
+)
+from ..prompt import GRAPH_FIELD_SEP
+
+
+@dataclass
+class NetworkXStorage(BaseGraphStorage):
+    @staticmethod
+    def load_nx_graph(file_name) -> nx.Graph:
+        if os.path.exists(file_name):
+            return nx.read_graphml(file_name)
+        return None
+
+    @staticmethod
+    def write_nx_graph(graph: nx.Graph, file_name):
+        logger.info(
+            f"Writing graph with {graph.number_of_nodes()} nodes, {graph.number_of_edges()} edges"
+        )
+        nx.write_graphml(graph, file_name)
+
+    @staticmethod
+    def stable_largest_connected_component(graph: nx.Graph) -> nx.Graph:
+        """Refer to https://github.com/microsoft/graphrag/index/graph/utils/stable_lcc.py
+        Return the largest connected component of the graph, with nodes and edges sorted in a stable way.
+        """
+        from graspologic.utils import largest_connected_component
+
+        graph = graph.copy()
+        graph = cast(nx.Graph, largest_connected_component(graph))
+        node_mapping = {node: html.unescape(node.upper().strip()) for node in graph.nodes()}  # type: ignore
+        graph = nx.relabel_nodes(graph, node_mapping)
+        return NetworkXStorage._stabilize_graph(graph)
+
+    @staticmethod
+    def _stabilize_graph(graph: nx.Graph) -> nx.Graph:
+        """Refer to https://github.com/microsoft/graphrag/index/graph/utils/stable_lcc.py
+        Ensure an undirected graph with the same relationships will always be read the same way.
+        """
+        fixed_graph = nx.DiGraph() if graph.is_directed() else nx.Graph()
+
+        sorted_nodes = graph.nodes(data=True)
+        sorted_nodes = sorted(sorted_nodes, key=lambda x: x[0])
+
+        fixed_graph.add_nodes_from(sorted_nodes)
+        edges = list(graph.edges(data=True))
+
+        if not graph.is_directed():
+
+            def _sort_source_target(edge):
+                source, target, edge_data = edge
+                if source > target:
+                    temp = source
+                    source = target
+                    target = temp
+                return source, target, edge_data
+
+            edges = [_sort_source_target(edge) for edge in edges]
+
+        def _get_edge_key(source: Any, target: Any) -> str:
+            return f"{source} -> {target}"
+
+        edges = sorted(edges, key=lambda x: _get_edge_key(x[0], x[1]))
+
+        fixed_graph.add_edges_from(edges)
+        return fixed_graph
+
+    def __post_init__(self):
+        self._graphml_xml_file = os.path.join(
+            self.global_config["working_dir"], f"graph_{self.namespace}.graphml"
+        )
+        preloaded_graph = NetworkXStorage.load_nx_graph(self._graphml_xml_file)
+        if preloaded_graph is not None:
+            logger.info(
+                f"Loaded graph from {self._graphml_xml_file} with {preloaded_graph.number_of_nodes()} nodes, {preloaded_graph.number_of_edges()} edges"
+            )
+        self._graph = preloaded_graph or nx.Graph()
+        self._clustering_algorithms = {
+            "leiden": self._leiden_clustering,
+        }
+        self._node_embed_algorithms = {
+            "node2vec": self._node2vec_embed,
+        }
+
+    async def index_done_callback(self):
+        NetworkXStorage.write_nx_graph(self._graph, self._graphml_xml_file)
+
+    async def has_node(self, node_id: str) -> bool:
+        return self._graph.has_node(node_id)
+
+    async def has_edge(self, source_node_id: str, target_node_id: str) -> bool:
+        return self._graph.has_edge(source_node_id, target_node_id)
+
+    async def get_node(self, node_id: str) -> Union[dict, None]:
+        return self._graph.nodes.get(node_id)
+
+    async def node_degree(self, node_id: str) -> int:
+        # [numberchiffre]: node_id not part of graph returns `DegreeView({})` instead of 0
+        return self._graph.degree(node_id) if self._graph.has_node(node_id) else 0
+
+    async def edge_degree(self, src_id: str, tgt_id: str) -> int:
+        return (self._graph.degree(src_id) if self._graph.has_node(src_id) else 0) + (
+            self._graph.degree(tgt_id) if self._graph.has_node(tgt_id) else 0
+        )
+
+    async def get_edge(
+        self, source_node_id: str, target_node_id: str
+    ) -> Union[dict, None]:
+        return self._graph.edges.get((source_node_id, target_node_id))
+
+    async def get_node_edges(self, source_node_id: str):
+        if self._graph.has_node(source_node_id):
+            return list(self._graph.edges(source_node_id))
+        return None
+
+    async def upsert_node(self, node_id: str, node_data: dict[str, str]):
+        self._graph.add_node(node_id, **node_data)
+
+    async def upsert_edge(
+        self, source_node_id: str, target_node_id: str, edge_data: dict[str, str]
+    ):
+        self._graph.add_edge(source_node_id, target_node_id, **edge_data)
+
+    async def clustering(self, algorithm: str):
+        if algorithm not in self._clustering_algorithms:
+            raise ValueError(f"Clustering algorithm {algorithm} not supported")
+        await self._clustering_algorithms[algorithm]()
+
+    async def community_schema(self) -> dict[str, SingleCommunitySchema]:
+        results = defaultdict(
+            lambda: dict(
+                level=None,
+                title=None,
+                edges=set(),
+                nodes=set(),
+                chunk_ids=set(),
+                occurrence=0.0,
+                sub_communities=[],
+            )
+        )
+        max_num_ids = 0
+        levels = defaultdict(set)
+        for node_id, node_data in self._graph.nodes(data=True):
+            if "clusters" not in node_data:
+                continue
+            clusters = json.loads(node_data["clusters"])
+            this_node_edges = self._graph.edges(node_id)
+
+            for cluster in clusters:
+                level = cluster["level"]
+                cluster_key = str(cluster["cluster"])
+                levels[level].add(cluster_key)
+                results[cluster_key]["level"] = level
+                results[cluster_key]["title"] = f"Cluster {cluster_key}"
+                results[cluster_key]["nodes"].add(node_id)
+                results[cluster_key]["edges"].update(
+                    [tuple(sorted(e)) for e in this_node_edges]
+                )
+                results[cluster_key]["chunk_ids"].update(
+                    node_data["source_id"].split(GRAPH_FIELD_SEP)
+                )
+                max_num_ids = max(max_num_ids, len(results[cluster_key]["chunk_ids"]))
+
+        ordered_levels = sorted(levels.keys())
+        for i, curr_level in enumerate(ordered_levels[:-1]):
+            next_level = ordered_levels[i + 1]
+            this_level_comms = levels[curr_level]
+            next_level_comms = levels[next_level]
+            # compute the sub-communities by nodes intersection
+            for comm in this_level_comms:
+                results[comm]["sub_communities"] = [
+                    c
+                    for c in next_level_comms
+                    if results[c]["nodes"].issubset(results[comm]["nodes"])
+                ]
+
+        for k, v in results.items():
+            v["edges"] = list(v["edges"])
+            v["edges"] = [list(e) for e in v["edges"]]
+            v["nodes"] = list(v["nodes"])
+            v["chunk_ids"] = list(v["chunk_ids"])
+            v["occurrence"] = len(v["chunk_ids"]) / max_num_ids
+        return dict(results)
+
+    def _cluster_data_to_subgraphs(self, cluster_data: dict[str, list[dict[str, str]]]):
+        for node_id, clusters in cluster_data.items():
+            self._graph.nodes[node_id]["clusters"] = json.dumps(clusters)
+
+    async def _leiden_clustering(self):
+        from graspologic.partition import hierarchical_leiden
+
+        graph = NetworkXStorage.stable_largest_connected_component(self._graph)
+        community_mapping = hierarchical_leiden(
+            graph,
+            max_cluster_size=self.global_config["max_graph_cluster_size"],
+            random_seed=self.global_config["graph_cluster_seed"],
+        )
+
+        node_communities: dict[str, list[dict[str, str]]] = defaultdict(list)
+        __levels = defaultdict(set)
+        for partition in community_mapping:
+            level_key = partition.level
+            cluster_id = partition.cluster
+            node_communities[partition.node].append(
+                {"level": level_key, "cluster": cluster_id}
+            )
+            __levels[level_key].add(cluster_id)
+        node_communities = dict(node_communities)
+        __levels = {k: len(v) for k, v in __levels.items()}
+        logger.info(f"Each level has communities: {dict(__levels)}")
+        self._cluster_data_to_subgraphs(node_communities)
+
+    async def embed_nodes(self, algorithm: str) -> tuple[np.ndarray, list[str]]:
+        if algorithm not in self._node_embed_algorithms:
+            raise ValueError(f"Node embedding algorithm {algorithm} not supported")
+        return await self._node_embed_algorithms[algorithm]()
+
+    async def _node2vec_embed(self):
+        from graspologic import embed
+
+        embeddings, nodes = embed.node2vec_embed(
+            self._graph,
+            **self.global_config["node2vec_params"],
+        )
+
+        nodes_ids = [self._graph.nodes[node_id]["id"] for node_id in nodes]
+        return embeddings, nodes_ids

videorag/_storage/kv_json.py

@@ -0,0 +1,46 @@
+import os
+from dataclasses import dataclass
+
+from .._utils import load_json, logger, write_json
+from ..base import (
+    BaseKVStorage,
+)
+
+
+@dataclass
+class JsonKVStorage(BaseKVStorage):
+    def __post_init__(self):
+        working_dir = self.global_config["working_dir"]
+        self._file_name = os.path.join(working_dir, f"kv_store_{self.namespace}.json")
+        self._data = load_json(self._file_name) or {}
+        logger.info(f"Load KV {self.namespace} with {len(self._data)} data")
+
+    async def all_keys(self) -> list[str]:
+        return list(self._data.keys())
+
+    async def index_done_callback(self):
+        write_json(self._data, self._file_name)
+
+    async def get_by_id(self, id):
+        return self._data.get(id, None)
+
+    async def get_by_ids(self, ids, fields=None):
+        if fields is None:
+            return [self._data.get(id, None) for id in ids]
+        return [
+            (
+                {k: v for k, v in self._data[id].items() if k in fields}
+                if self._data.get(id, None)
+                else None
+            )
+            for id in ids
+        ]
+
+    async def filter_keys(self, data: list[str]) -> set[str]:
+        return set([s for s in data if s not in self._data])
+
+    async def upsert(self, data: dict[str, dict]):
+        self._data.update(data)
+
+    async def drop(self):
+        self._data = {}

videorag/_storage/vdb_hnswlib.py

@@ -0,0 +1,141 @@
+import asyncio
+import os
+from dataclasses import dataclass, field
+from typing import Any
+import pickle
+import hnswlib
+import numpy as np
+import xxhash
+
+from .._utils import logger
+from ..base import BaseVectorStorage
+
+
+@dataclass
+class HNSWVectorStorage(BaseVectorStorage):
+    ef_construction: int = 100
+    M: int = 16
+    max_elements: int = 1000000
+    ef_search: int = 50
+    num_threads: int = -1
+    _index: Any = field(init=False)
+    _metadata: dict[str, dict] = field(default_factory=dict)
+    _current_elements: int = 0
+
+    def __post_init__(self):
+        self._index_file_name = os.path.join(
+            self.global_config["working_dir"], f"{self.namespace}_hnsw.index"
+        )
+        self._metadata_file_name = os.path.join(
+            self.global_config["working_dir"], f"{self.namespace}_hnsw_metadata.pkl"
+        )
+        self._embedding_batch_num = self.global_config.get("embedding_batch_num", 100)
+
+        hnsw_params = self.global_config.get("vector_db_storage_cls_kwargs", {})
+        self.ef_construction = hnsw_params.get("ef_construction", self.ef_construction)
+        self.M = hnsw_params.get("M", self.M)
+        self.max_elements = hnsw_params.get("max_elements", self.max_elements)
+        self.ef_search = hnsw_params.get("ef_search", self.ef_search)
+        self.num_threads = hnsw_params.get("num_threads", self.num_threads)
+        self._index = hnswlib.Index(
+            space="cosine", dim=self.embedding_func.embedding_dim
+        )
+
+        if os.path.exists(self._index_file_name) and os.path.exists(
+            self._metadata_file_name
+        ):
+            self._index.load_index(
+                self._index_file_name, max_elements=self.max_elements
+            )
+            with open(self._metadata_file_name, "rb") as f:
+                self._metadata, self._current_elements = pickle.load(f)
+            logger.info(
+                f"Loaded existing index for {self.namespace} with {self._current_elements} elements"
+            )
+        else:
+            self._index.init_index(
+                max_elements=self.max_elements,
+                ef_construction=self.ef_construction,
+                M=self.M,
+            )
+            self._index.set_ef(self.ef_search)
+            self._metadata = {}
+            self._current_elements = 0
+            logger.info(f"Created new index for {self.namespace}")
+
+    async def upsert(self, data: dict[str, dict]) -> np.ndarray:
+        logger.info(f"Inserting {len(data)} vectors to {self.namespace}")
+        if not data:
+            logger.warning("You insert an empty data to vector DB")
+            return []
+
+        if self._current_elements + len(data) > self.max_elements:
+            raise ValueError(
+                f"Cannot insert {len(data)} elements. Current: {self._current_elements}, Max: {self.max_elements}"
+            )
+
+        list_data = [
+            {
+                "id": k,
+                **{k1: v1 for k1, v1 in v.items() if k1 in self.meta_fields},
+            }
+            for k, v in data.items()
+        ]
+        contents = [v["content"] for v in data.values()]
+        batch_size = min(self._embedding_batch_num, len(contents))
+        embeddings = np.concatenate(
+            await asyncio.gather(
+                *[
+                    self.embedding_func(contents[i : i + batch_size])
+                    for i in range(0, len(contents), batch_size)
+                ]
+            )
+        )
+
+        ids = np.fromiter(
+            (xxhash.xxh32_intdigest(d["id"].encode()) for d in list_data),
+            dtype=np.uint32,
+            count=len(list_data),
+        )
+        self._metadata.update(
+            {
+                id_int: {
+                    k: v for k, v in d.items() if k in self.meta_fields or k == "id"
+                }
+                for id_int, d in zip(ids, list_data)
+            }
+        )
+        self._index.add_items(data=embeddings, ids=ids, num_threads=self.num_threads)
+        self._current_elements = self._index.get_current_count()
+        return ids
+
+    async def query(self, query: str, top_k: int = 5) -> list[dict]:
+        if self._current_elements == 0:
+            return []
+
+        top_k = min(top_k, self._current_elements)
+
+        if top_k > self.ef_search:
+            logger.warning(
+                f"Setting ef_search to {top_k} because top_k is larger than ef_search"
+            )
+            self._index.set_ef(top_k)
+
+        embedding = await self.embedding_func([query])
+        labels, distances = self._index.knn_query(
+            data=embedding[0], k=top_k, num_threads=self.num_threads
+        )
+
+        return [
+            {
+                **self._metadata.get(label, {}),
+                "distance": distance,
+                "similarity": 1 - distance,
+            }
+            for label, distance in zip(labels[0], distances[0])
+        ]
+
+    async def index_done_callback(self):
+        self._index.save_index(self._index_file_name)
+        with open(self._metadata_file_name, "wb") as f:
+            pickle.dump((self._metadata, self._current_elements), f)

videorag/_storage/vdb_nanovectordb.py

@@ -0,0 +1,145 @@
+import asyncio
+import os
+import torch
+from dataclasses import dataclass
+import numpy as np
+from nano_vectordb import NanoVectorDB
+from tqdm import tqdm
+from imagebind.models import imagebind_model
+
+from .._utils import logger
+from ..base import BaseVectorStorage
+from .._videoutil import encode_video_segments, encode_string_query
+
+
+@dataclass
+class NanoVectorDBStorage(BaseVectorStorage):
+    cosine_better_than_threshold: float = 0.2
+    
+    def __post_init__(self):
+
+        self._client_file_name = os.path.join(
+            self.global_config["working_dir"], f"vdb_{self.namespace}.json"
+        )
+        self._max_batch_size = self.global_config["embedding_batch_num"]
+        self._client = NanoVectorDB(
+            self.embedding_func.embedding_dim, storage_file=self._client_file_name
+        )
+        self.cosine_better_than_threshold = self.global_config.get(
+            "query_better_than_threshold", self.cosine_better_than_threshold
+        )
+
+    async def upsert(self, data: dict[str, dict]):
+        logger.info(f"Inserting {len(data)} vectors to {self.namespace}")
+        if not len(data):
+            logger.warning("You insert an empty data to vector DB")
+            return []
+        list_data = [
+            {
+                "__id__": k,
+                **{k1: v1 for k1, v1 in v.items() if k1 in self.meta_fields},
+            }
+            for k, v in data.items()
+        ]
+        contents = [v["content"] for v in data.values()]
+        batches = [
+            contents[i : i + self._max_batch_size]
+            for i in range(0, len(contents), self._max_batch_size)
+        ]
+        embeddings_list = await asyncio.gather(
+            *[self.embedding_func(batch) for batch in batches]
+        )
+        embeddings = np.concatenate(embeddings_list)
+        for i, d in enumerate(list_data):
+            d["__vector__"] = embeddings[i]
+        results = self._client.upsert(datas=list_data)
+        return results
+
+    async def query(self, query: str, top_k=5):
+        embedding = await self.embedding_func([query])
+        embedding = embedding[0]
+        results = self._client.query(
+            query=embedding,
+            top_k=top_k,
+            better_than_threshold=self.cosine_better_than_threshold,
+        )
+        results = [
+            {**dp, "id": dp["__id__"], "distance": dp["__metrics__"]} for dp in results
+        ]
+        return results
+
+    async def index_done_callback(self):
+        self._client.save()
+
+
+@dataclass
+class NanoVectorDBVideoSegmentStorage(BaseVectorStorage):
+    embedding_func = None
+    segment_retrieval_top_k: float = 2
+    
+    def __post_init__(self):
+        
+        self._client_file_name = os.path.join(
+            self.global_config["working_dir"], f"vdb_{self.namespace}.json"
+        )
+        self._max_batch_size = self.global_config["video_embedding_batch_num"]
+        self._client = NanoVectorDB(
+            self.global_config["video_embedding_dim"], storage_file=self._client_file_name
+        )
+        self.top_k = self.global_config.get(
+            "segment_retrieval_top_k", self.segment_retrieval_top_k
+        )
+    
+    async def upsert(self, video_name, segment_index2name, video_output_format):
+        embedder = imagebind_model.imagebind_huge(pretrained=True).cuda()
+        embedder.eval()
+        
+        logger.info(f"Inserting {len(segment_index2name)} segments to {self.namespace}")
+        if not len(segment_index2name):
+            logger.warning("You insert an empty data to vector DB")
+            return []
+        list_data, video_paths = [], []
+        cache_path = os.path.join(self.global_config["working_dir"], '_cache', video_name)
+        index_list = list(segment_index2name.keys())
+        for index in index_list:
+            list_data.append({
+                "__id__": f"{video_name}_{index}",
+                "__video_name__": video_name,
+                "__index__": index,
+            })
+            segment_name = segment_index2name[index]
+            video_file = os.path.join(cache_path, f"{segment_name}.{video_output_format}")
+            video_paths.append(video_file)
+        batches = [
+            video_paths[i: i + self._max_batch_size]
+            for i in range(0, len(video_paths), self._max_batch_size)
+        ]
+        embeddings = []
+        for _batch in tqdm(batches, desc=f"Encoding Video Segments {video_name}"):
+            batch_embeddings = encode_video_segments(_batch, embedder)
+            embeddings.append(batch_embeddings)
+        embeddings = torch.concat(embeddings, dim=0)
+        embeddings = embeddings.numpy()
+        for i, d in enumerate(list_data):
+            d["__vector__"] = embeddings[i]
+        results = self._client.upsert(datas=list_data)
+        return results
+    
+    async def query(self, query: str):
+        embedder = imagebind_model.imagebind_huge(pretrained=True).cuda()
+        embedder.eval()
+        
+        embedding = encode_string_query(query, embedder)
+        embedding = embedding[0]
+        results = self._client.query(
+            query=embedding,
+            top_k=self.top_k,
+            better_than_threshold=-1,
+        )
+        results = [
+            {**dp, "id": dp["__id__"], "distance": dp["__metrics__"]} for dp in results
+        ]
+        return results
+    
+    async def index_done_callback(self):
+        self._client.save()

videorag/_utils.py

@@ -0,0 +1,193 @@
+import asyncio
+import html
+import json
+import logging
+import os
+import re
+import numbers
+from dataclasses import dataclass
+from functools import wraps
+from hashlib import md5
+from typing import Any, Union
+
+import numpy as np
+import tiktoken
+
+logger = logging.getLogger("nano-graphrag")
+ENCODER = None
+
+
+def always_get_an_event_loop() -> asyncio.AbstractEventLoop:
+    try:
+        # If there is already an event loop, use it.
+        loop = asyncio.get_event_loop()
+    except RuntimeError:
+        # If in a sub-thread, create a new event loop.
+        logger.info("Creating a new event loop in a sub-thread.")
+        loop = asyncio.new_event_loop()
+        asyncio.set_event_loop(loop)
+    return loop
+
+
+def locate_json_string_body_from_string(content: str) -> Union[str, None]:
+    """Locate the JSON string body from a string"""
+    maybe_json_str = re.search(r"{.*}", content, re.DOTALL)
+    if maybe_json_str is not None:
+        return maybe_json_str.group(0)
+    else:
+        return None
+
+
+def convert_response_to_json(response: str) -> dict:
+    json_str = locate_json_string_body_from_string(response)
+    assert json_str is not None, f"Unable to parse JSON from response: {response}"
+    try:
+        data = json.loads(json_str)
+        return data
+    except json.JSONDecodeError as e:
+        logger.error(f"Failed to parse JSON: {json_str}")
+        raise e from None
+
+
+def encode_string_by_tiktoken(content: str, model_name: str = "gpt-4o"):
+    global ENCODER
+    if ENCODER is None:
+        ENCODER = tiktoken.encoding_for_model(model_name)
+    tokens = ENCODER.encode(content)
+    return tokens
+
+
+def decode_tokens_by_tiktoken(tokens: list[int], model_name: str = "gpt-4o"):
+    global ENCODER
+    if ENCODER is None:
+        ENCODER = tiktoken.encoding_for_model(model_name)
+    content = ENCODER.decode(tokens)
+    return content
+
+
+def truncate_list_by_token_size(list_data: list, key: callable, max_token_size: int):
+    """Truncate a list of data by token size"""
+    if max_token_size <= 0:
+        return []
+    tokens = 0
+    for i, data in enumerate(list_data):
+        tokens += len(encode_string_by_tiktoken(key(data)))
+        if tokens > max_token_size:
+            return list_data[:i]
+    return list_data
+
+
+def compute_mdhash_id(content, prefix: str = ""):
+    return prefix + md5(content.encode()).hexdigest()
+
+
+def write_json(json_obj, file_name):
+    with open(file_name, "w", encoding="utf-8") as f:
+        json.dump(json_obj, f, indent=2, ensure_ascii=False)
+
+
+def load_json(file_name):
+    if not os.path.exists(file_name):
+        return None
+    with open(file_name, encoding="utf-8") as f:
+        return json.load(f)
+
+
+# it's dirty to type, so it's a good way to have fun
+def pack_user_ass_to_openai_messages(*args: str):
+    roles = ["user", "assistant"]
+    return [
+        {"role": roles[i % 2], "content": content} for i, content in enumerate(args)
+    ]
+
+
+def is_float_regex(value):
+    return bool(re.match(r"^[-+]?[0-9]*\.?[0-9]+$", value))
+
+
+def compute_args_hash(*args):
+    return md5(str(args).encode()).hexdigest()
+
+
+def split_string_by_multi_markers(content: str, markers: list[str]) -> list[str]:
+    """Split a string by multiple markers"""
+    if not markers:
+        return [content]
+    results = re.split("|".join(re.escape(marker) for marker in markers), content)
+    return [r.strip() for r in results if r.strip()]
+
+
+def enclose_string_with_quotes(content: Any) -> str:
+    """Enclose a string with quotes"""
+    if isinstance(content, numbers.Number):
+        return str(content)
+    content = str(content)
+    content = content.strip().strip("'").strip('"')
+    return f'"{content}"'
+
+
+def list_of_list_to_csv(data: list[list]):
+    return "\n".join(
+        [
+            ",\t".join([f"{enclose_string_with_quotes(data_dd)}" for data_dd in data_d])
+            for data_d in data
+        ]
+    )
+
+
+# -----------------------------------------------------------------------------------
+# Refer the utils functions of the official GraphRAG implementation:
+# https://github.com/microsoft/graphrag
+def clean_str(input: Any) -> str:
+    """Clean an input string by removing HTML escapes, control characters, and other unwanted characters."""
+    # If we get non-string input, just give it back
+    if not isinstance(input, str):
+        return input
+
+    result = html.unescape(input.strip())
+    # https://stackoverflow.com/questions/4324790/removing-control-characters-from-a-string-in-python
+    return re.sub(r"[\x00-\x1f\x7f-\x9f]", "", result)
+
+
+# Utils types -----------------------------------------------------------------------
+@dataclass
+class EmbeddingFunc:
+    embedding_dim: int
+    max_token_size: int
+    func: callable
+
+    async def __call__(self, *args, **kwargs) -> np.ndarray:
+        return await self.func(*args, **kwargs)
+
+
+# Decorators ------------------------------------------------------------------------
+def limit_async_func_call(max_size: int, waitting_time: float = 0.0001):
+    """Add restriction of maximum async calling times for a async func"""
+
+    def final_decro(func):
+        """Not using async.Semaphore to aovid use nest-asyncio"""
+        __current_size = 0
+
+        @wraps(func)
+        async def wait_func(*args, **kwargs):
+            nonlocal __current_size
+            while __current_size >= max_size:
+                await asyncio.sleep(waitting_time)
+            __current_size += 1
+            result = await func(*args, **kwargs)
+            __current_size -= 1
+            return result
+
+        return wait_func
+
+    return final_decro
+
+
+def wrap_embedding_func_with_attrs(**kwargs):
+    """Wrap a function with attributes"""
+
+    def final_decro(func) -> EmbeddingFunc:
+        new_func = EmbeddingFunc(**kwargs, func=func)
+        return new_func
+
+    return final_decro

videorag/_videoutil/__init__.py

@@ -0,0 +1,4 @@
+from .split import split_video, saving_video_segments
+from .asr import speech_to_text
+from .caption import segment_caption, merge_segment_information, retrieved_segment_caption
+from .feature import encode_video_segments, encode_string_query

videorag/_videoutil/asr.py

@@ -0,0 +1,24 @@
+import os
+import torch
+import logging
+from tqdm import tqdm
+from faster_whisper import WhisperModel
+from transformers import AutoModelForSpeechSeq2Seq, AutoProcessor, pipeline
+
+def speech_to_text(video_name, working_dir, segment_index2name, audio_output_format):
+    model = WhisperModel("./faster-distil-whisper-large-v3")
+    model.logger.setLevel(logging.WARNING)
+    
+    cache_path = os.path.join(working_dir, '_cache', video_name)
+    
+    transcripts = {}
+    for index in tqdm(segment_index2name, desc=f"Speech Recognition {video_name}"):
+        segment_name = segment_index2name[index]
+        audio_file = os.path.join(cache_path, f"{segment_name}.{audio_output_format}")
+        segments, info = model.transcribe(audio_file)
+        result = ""
+        for segment in segments:
+            result += "[%.2fs -> %.2fs] %s\n" % (segment.start, segment.end, segment.text)
+        transcripts[index] = result
+    
+    return transcripts

videorag/_videoutil/caption.py

@@ -0,0 +1,88 @@
+import os
+import torch
+import numpy as np
+from PIL import Image
+from tqdm import tqdm
+from transformers import AutoModel, AutoTokenizer
+from moviepy.video.io.VideoFileClip import VideoFileClip
+
+def encode_video(video, frame_times):
+    frames = []
+    for t in frame_times:
+        frames.append(video.get_frame(t))
+    frames = np.stack(frames, axis=0)
+    frames = [Image.fromarray(v.astype('uint8')).resize((1280, 720)) for v in frames]
+    return frames
+    
+def segment_caption(video_name, video_path, segment_index2name, transcripts, segment_times_info, caption_result, error_queue):
+    try:
+        model = AutoModel.from_pretrained('./MiniCPM-V-2_6-int4', trust_remote_code=True)
+        tokenizer = AutoTokenizer.from_pretrained('./MiniCPM-V-2_6-int4', trust_remote_code=True)
+        model.eval()
+        
+        with VideoFileClip(video_path) as video:
+            for index in tqdm(segment_index2name, desc=f"Captioning Video {video_name}"):
+                frame_times = segment_times_info[index]["frame_times"]
+                video_frames = encode_video(video, frame_times)
+                segment_transcript = transcripts[index]
+                query = f"The transcript of the current video:\n{segment_transcript}.\nNow provide a description (caption) of the video in English."
+                msgs = [{'role': 'user', 'content': video_frames + [query]}]
+                params = {}
+                params["use_image_id"] = False
+                params["max_slice_nums"] = 2
+                segment_caption = model.chat(
+                    image=None,
+                    msgs=msgs,
+                    tokenizer=tokenizer,
+                    **params
+                )
+                caption_result[index] = segment_caption.replace("\n", "").replace("<|endoftext|>", "")
+                torch.cuda.empty_cache()
+    except Exception as e:
+        error_queue.put(f"Error in segment_caption:\n {str(e)}")
+        raise RuntimeError
+
+def merge_segment_information(segment_index2name, segment_times_info, transcripts, captions):
+    inserting_segments = {}
+    for index in segment_index2name:
+        inserting_segments[index] = {"content": None, "time": None}
+        segment_name = segment_index2name[index]
+        inserting_segments[index]["time"] = '-'.join(segment_name.split('-')[-2:])
+        inserting_segments[index]["content"] = f"Caption:\n{captions[index]}\nTranscript:\n{transcripts[index]}\n\n"
+        inserting_segments[index]["transcript"] = transcripts[index]
+        inserting_segments[index]["frame_times"] = segment_times_info[index]["frame_times"].tolist()
+    return inserting_segments
+        
+def retrieved_segment_caption(caption_model, caption_tokenizer, refine_knowledge, retrieved_segments, video_path_db, video_segments, num_sampled_frames):
+    # model = AutoModel.from_pretrained('./MiniCPM-V-2_6-int4', trust_remote_code=True)
+    # tokenizer = AutoTokenizer.from_pretrained('./MiniCPM-V-2_6-int4', trust_remote_code=True)
+    # model.eval()
+    
+    caption_result = {}
+    for this_segment in tqdm(retrieved_segments, desc='Captioning Segments for Given Query'):
+        video_name = '_'.join(this_segment.split('_')[:-1])
+        index = this_segment.split('_')[-1]
+        video_path = video_path_db._data[video_name]
+        timestamp = video_segments._data[video_name][index]["time"].split('-')
+        start, end = eval(timestamp[0]), eval(timestamp[1])
+        video = VideoFileClip(video_path)
+        frame_times = np.linspace(start, end, num_sampled_frames, endpoint=False)
+        video_frames = encode_video(video, frame_times)
+        segment_transcript = video_segments._data[video_name][index]["transcript"]
+        # query = f"The transcript of the current video:\n{segment_transcript}.\nGiven a question: {query}, you have to extract relevant information from the video and transcript for answering the question."
+        query = f"The transcript of the current video:\n{segment_transcript}.\nNow provide a very detailed description (caption) of the video in English and extract relevant information about: {refine_knowledge}'"
+        msgs = [{'role': 'user', 'content': video_frames + [query]}]
+        params = {}
+        params["use_image_id"] = False
+        params["max_slice_nums"] = 2
+        segment_caption = caption_model.chat(
+            image=None,
+            msgs=msgs,
+            tokenizer=caption_tokenizer,
+            **params
+        )
+        this_caption = segment_caption.replace("\n", "").replace("<|endoftext|>", "")
+        caption_result[this_segment] = f"Caption:\n{this_caption}\nTranscript:\n{segment_transcript}\n\n"
+        torch.cuda.empty_cache()
+    
+    return caption_result

videorag/_videoutil/feature.py

@@ -0,0 +1,28 @@
+import os
+import torch
+import pickle
+from tqdm import tqdm
+from imagebind import data
+from imagebind.models import imagebind_model
+from imagebind.models.imagebind_model import ImageBindModel, ModalityType
+
+
+def encode_video_segments(video_paths, embedder: ImageBindModel):
+    device = next(embedder.parameters()).device
+    inputs = {
+        ModalityType.VISION: data.load_and_transform_video_data(video_paths, device),
+    }
+    with torch.no_grad():
+        embeddings = embedder(inputs)[ModalityType.VISION]
+    embeddings = embeddings.cpu()
+    return embeddings
+
+def encode_string_query(query:str, embedder: ImageBindModel):
+    device = next(embedder.parameters()).device
+    inputs = {
+        ModalityType.TEXT: data.load_and_transform_text([query], device),
+    }
+    with torch.no_grad():
+        embeddings = embedder(inputs)[ModalityType.TEXT]
+    embeddings = embeddings.cpu()
+    return embeddings

videorag/_videoutil/split.py

@@ -0,0 +1,76 @@
+import os
+import time
+import shutil
+import numpy as np
+from tqdm import tqdm
+from moviepy.video import fx as vfx
+from moviepy.video.io.VideoFileClip import VideoFileClip
+
+def split_video(
+    video_path,
+    working_dir,
+    segment_length,
+    num_frames_per_segment,
+    audio_output_format='mp3',
+):  
+    unique_timestamp = str(int(time.time() * 1000))
+    video_name = os.path.basename(video_path).split('.')[0]
+    video_segment_cache_path = os.path.join(working_dir, '_cache', video_name)
+    if os.path.exists(video_segment_cache_path):
+        shutil.rmtree(video_segment_cache_path)
+    os.makedirs(video_segment_cache_path, exist_ok=False)
+    
+    segment_index = 0
+    segment_index2name, segment_times_info = {}, {}
+    with VideoFileClip(video_path) as video:
+    
+        total_video_length = int(video.duration)
+        start_times = list(range(0, total_video_length, segment_length))
+        # if the last segment is shorter than 5 seconds, we merged it to the last segment
+        if len(start_times) > 1 and (total_video_length - start_times[-1]) < 5:
+            start_times = start_times[:-1]
+        
+        for start in tqdm(start_times, desc=f"Spliting Video {video_name}"):
+            if start != start_times[-1]:
+                end = min(start + segment_length, total_video_length)
+            else:
+                end = total_video_length
+            
+            subvideo = video.subclip(start, end)
+            subvideo_length = subvideo.duration
+            frame_times = np.linspace(0, subvideo_length, num_frames_per_segment, endpoint=False)
+            frame_times += start
+            
+            segment_index2name[f"{segment_index}"] = f"{unique_timestamp}-{segment_index}-{start}-{end}"
+            segment_times_info[f"{segment_index}"] = {"frame_times": frame_times, "timestamp": (start, end)}
+            
+            # save audio
+            audio_file_base_name = segment_index2name[f"{segment_index}"]
+            audio_file = f'{audio_file_base_name}.{audio_output_format}'
+            subaudio = subvideo.audio
+            subaudio.write_audiofile(os.path.join(video_segment_cache_path, audio_file), codec='mp3', verbose=False, logger=None)
+            
+            segment_index += 1
+
+    return segment_index2name, segment_times_info
+
+def saving_video_segments(
+    video_name,
+    video_path,
+    working_dir,
+    segment_index2name,
+    segment_times_info,
+    error_queue,
+    video_output_format='mp4',
+):
+    try:
+        with VideoFileClip(video_path) as video:
+            video_segment_cache_path = os.path.join(working_dir, '_cache', video_name)
+            for index in tqdm(segment_index2name, desc=f"Saving Video Segments {video_name}"):
+                start, end = segment_times_info[index]["timestamp"][0], segment_times_info[index]["timestamp"][1]
+                video_file = f'{segment_index2name[index]}.{video_output_format}'
+                subvideo = video.subclip(start, end)
+                subvideo.write_videofile(os.path.join(video_segment_cache_path, video_file), codec='libx264', verbose=False, logger=None)
+    except Exception as e:
+        error_queue.put(f"Error in saving_video_segments:\n {str(e)}")
+        raise RuntimeError

videorag/base.py

@@ -0,0 +1,149 @@
+from dataclasses import dataclass, field
+from typing import TypedDict, Union, Literal, Generic, TypeVar
+
+import numpy as np
+
+from ._utils import EmbeddingFunc
+
+
+@dataclass
+class QueryParam:
+    mode: Literal["local", "global", "naive"] = "global"
+    only_need_context: bool = False
+    response_type: str = "Multiple Paragraphs"
+    level: int = 2
+    top_k: int = 20
+    # naive search
+    naive_max_token_for_text_unit = 12000
+    # videorag search
+    only_need_context: bool = False
+
+
+TextChunkSchema = TypedDict(
+    "TextChunkSchema",
+    {"tokens": int, "content": str, "video_segment_id": str, "chunk_order_index": int},
+)
+
+SingleCommunitySchema = TypedDict(
+    "SingleCommunitySchema",
+    {
+        "level": int,
+        "title": str,
+        "edges": list[list[str, str]],
+        "nodes": list[str],
+        "chunk_ids": list[str],
+        "occurrence": float,
+        "sub_communities": list[str],
+    },
+)
+
+
+class CommunitySchema(SingleCommunitySchema):
+    report_string: str
+    report_json: dict
+
+
+T = TypeVar("T")
+
+
+@dataclass
+class StorageNameSpace:
+    namespace: str
+    global_config: dict
+
+    async def index_start_callback(self):
+        """commit the storage operations after indexing"""
+        pass
+
+    async def index_done_callback(self):
+        """commit the storage operations after indexing"""
+        pass
+
+    async def query_done_callback(self):
+        """commit the storage operations after querying"""
+        pass
+
+
+@dataclass
+class BaseVectorStorage(StorageNameSpace):
+    embedding_func: EmbeddingFunc
+    meta_fields: set = field(default_factory=set)
+
+    async def query(self, query: str, top_k: int) -> list[dict]:
+        raise NotImplementedError
+
+    async def upsert(self, data: dict[str, dict]):
+        """Use 'content' field from value for embedding, use key as id.
+        If embedding_func is None, use 'embedding' field from value
+        """
+        raise NotImplementedError
+
+
+@dataclass
+class BaseKVStorage(Generic[T], StorageNameSpace):
+    async def all_keys(self) -> list[str]:
+        raise NotImplementedError
+
+    async def get_by_id(self, id: str) -> Union[T, None]:
+        raise NotImplementedError
+
+    async def get_by_ids(
+        self, ids: list[str], fields: Union[set[str], None] = None
+    ) -> list[Union[T, None]]:
+        raise NotImplementedError
+
+    async def filter_keys(self, data: list[str]) -> set[str]:
+        """return un-exist keys"""
+        raise NotImplementedError
+
+    async def upsert(self, data: dict[str, T]):
+        raise NotImplementedError
+
+    async def drop(self):
+        raise NotImplementedError
+
+
+@dataclass
+class BaseGraphStorage(StorageNameSpace):
+    async def has_node(self, node_id: str) -> bool:
+        raise NotImplementedError
+
+    async def has_edge(self, source_node_id: str, target_node_id: str) -> bool:
+        raise NotImplementedError
+
+    async def node_degree(self, node_id: str) -> int:
+        raise NotImplementedError
+
+    async def edge_degree(self, src_id: str, tgt_id: str) -> int:
+        raise NotImplementedError
+
+    async def get_node(self, node_id: str) -> Union[dict, None]:
+        raise NotImplementedError
+
+    async def get_edge(
+        self, source_node_id: str, target_node_id: str
+    ) -> Union[dict, None]:
+        raise NotImplementedError
+
+    async def get_node_edges(
+        self, source_node_id: str
+    ) -> Union[list[tuple[str, str]], None]:
+        raise NotImplementedError
+
+    async def upsert_node(self, node_id: str, node_data: dict[str, str]):
+        raise NotImplementedError
+
+    async def upsert_edge(
+        self, source_node_id: str, target_node_id: str, edge_data: dict[str, str]
+    ):
+        raise NotImplementedError
+
+    async def clustering(self, algorithm: str):
+        raise NotImplementedError
+
+    async def community_schema(self) -> dict[str, SingleCommunitySchema]:
+        """Return the community representation with report and nodes"""
+        raise NotImplementedError
+
+    async def embed_nodes(self, algorithm: str) -> tuple[np.ndarray, list[str]]:
+        raise NotImplementedError("Node embedding is not used in nano-graphrag.")

videorag/prompt.py

@@ -0,0 +1,403 @@
+"""
+Reference:
+ - Prompts are from [graphrag](https://github.com/microsoft/graphrag)
+"""
+
+GRAPH_FIELD_SEP = "<SEP>"
+PROMPTS = {}
+
+PROMPTS[
+    "entity_extraction"
+] = """-Goal-
+Given a text document that is potentially relevant to this activity and a list of entity types, identify all entities of those types from the text and all relationships among the identified entities.
+
+-Steps-
+1. Identify all entities. For each identified entity, extract the following information:
+- entity_name: Name of the entity, capitalized
+- entity_type: One of the following types: [{entity_types}]
+- entity_description: Comprehensive description of the entity's attributes and activities
+Format each entity as ("entity"{tuple_delimiter}<entity_name>{tuple_delimiter}<entity_type>{tuple_delimiter}<entity_description>
+
+2. From the entities identified in step 1, identify all pairs of (source_entity, target_entity) that are *clearly related* to each other.
+For each pair of related entities, extract the following information:
+- source_entity: name of the source entity, as identified in step 1
+- target_entity: name of the target entity, as identified in step 1
+- relationship_description: explanation as to why you think the source entity and the target entity are related to each other
+- relationship_strength: a numeric score indicating strength of the relationship between the source entity and target entity
+ Format each relationship as ("relationship"{tuple_delimiter}<source_entity>{tuple_delimiter}<target_entity>{tuple_delimiter}<relationship_description>{tuple_delimiter}<relationship_strength>)
+
+3. Return output in English as a single list of all the entities and relationships identified in steps 1 and 2. Use **{record_delimiter}** as the list delimiter.
+
+4. When finished, output {completion_delimiter}
+
+######################
+-Examples-
+######################
+Example 1:
+
+Entity_types: [person, technology, mission, organization, location]
+Text:
+while Alex clenched his jaw, the buzz of frustration dull against the backdrop of Taylor's authoritarian certainty. It was this competitive undercurrent that kept him alert, the sense that his and Jordan's shared commitment to discovery was an unspoken rebellion against Cruz's narrowing vision of control and order.
+
+Then Taylor did something unexpected. They paused beside Jordan and, for a moment, observed the device with something akin to reverence. "If this tech can be understood..." Taylor said, their voice quieter, "It could change the game for us. For all of us."
+
+The underlying dismissal earlier seemed to falter, replaced by a glimpse of reluctant respect for the gravity of what lay in their hands. Jordan looked up, and for a fleeting heartbeat, their eyes locked with Taylor's, a wordless clash of wills softening into an uneasy truce.
+
+It was a small transformation, barely perceptible, but one that Alex noted with an inward nod. They had all been brought here by different paths
+################
+Output:
+("entity"{tuple_delimiter}"Alex"{tuple_delimiter}"person"{tuple_delimiter}"Alex is a character who experiences frustration and is observant of the dynamics among other characters."){record_delimiter}
+("entity"{tuple_delimiter}"Taylor"{tuple_delimiter}"person"{tuple_delimiter}"Taylor is portrayed with authoritarian certainty and shows a moment of reverence towards a device, indicating a change in perspective."){record_delimiter}
+("entity"{tuple_delimiter}"Jordan"{tuple_delimiter}"person"{tuple_delimiter}"Jordan shares a commitment to discovery and has a significant interaction with Taylor regarding a device."){record_delimiter}
+("entity"{tuple_delimiter}"Cruz"{tuple_delimiter}"person"{tuple_delimiter}"Cruz is associated with a vision of control and order, influencing the dynamics among other characters."){record_delimiter}
+("entity"{tuple_delimiter}"The Device"{tuple_delimiter}"technology"{tuple_delimiter}"The Device is central to the story, with potential game-changing implications, and is revered by Taylor."){record_delimiter}
+("relationship"{tuple_delimiter}"Alex"{tuple_delimiter}"Taylor"{tuple_delimiter}"Alex is affected by Taylor's authoritarian certainty and observes changes in Taylor's attitude towards the device."{tuple_delimiter}7){record_delimiter}
+("relationship"{tuple_delimiter}"Alex"{tuple_delimiter}"Jordan"{tuple_delimiter}"Alex and Jordan share a commitment to discovery, which contrasts with Cruz's vision."{tuple_delimiter}6){record_delimiter}
+("relationship"{tuple_delimiter}"Taylor"{tuple_delimiter}"Jordan"{tuple_delimiter}"Taylor and Jordan interact directly regarding the device, leading to a moment of mutual respect and an uneasy truce."{tuple_delimiter}8){record_delimiter}
+("relationship"{tuple_delimiter}"Jordan"{tuple_delimiter}"Cruz"{tuple_delimiter}"Jordan's commitment to discovery is in rebellion against Cruz's vision of control and order."{tuple_delimiter}5){record_delimiter}
+("relationship"{tuple_delimiter}"Taylor"{tuple_delimiter}"The Device"{tuple_delimiter}"Taylor shows reverence towards the device, indicating its importance and potential impact."{tuple_delimiter}9){completion_delimiter}
+#############################
+Example 2:
+
+Entity_types: [person, technology, mission, organization, location]
+Text:
+They were no longer mere operatives; they had become guardians of a threshold, keepers of a message from a realm beyond stars and stripes. This elevation in their mission could not be shackled by regulations and established protocols—it demanded a new perspective, a new resolve.
+
+Tension threaded through the dialogue of beeps and static as communications with Washington buzzed in the background. The team stood, a portentous air enveloping them. It was clear that the decisions they made in the ensuing hours could redefine humanity's place in the cosmos or condemn them to ignorance and potential peril.
+
+Their connection to the stars solidified, the group moved to address the crystallizing warning, shifting from passive recipients to active participants. Mercer's latter instincts gained precedence— the team's mandate had evolved, no longer solely to observe and report but to interact and prepare. A metamorphosis had begun, and Operation: Dulce hummed with the newfound frequency of their daring, a tone set not by the earthly
+#############
+Output:
+("entity"{tuple_delimiter}"Washington"{tuple_delimiter}"location"{tuple_delimiter}"Washington is a location where communications are being received, indicating its importance in the decision-making process."){record_delimiter}
+("entity"{tuple_delimiter}"Operation: Dulce"{tuple_delimiter}"mission"{tuple_delimiter}"Operation: Dulce is described as a mission that has evolved to interact and prepare, indicating a significant shift in objectives and activities."){record_delimiter}
+("entity"{tuple_delimiter}"The team"{tuple_delimiter}"organization"{tuple_delimiter}"The team is portrayed as a group of individuals who have transitioned from passive observers to active participants in a mission, showing a dynamic change in their role."){record_delimiter}
+("relationship"{tuple_delimiter}"The team"{tuple_delimiter}"Washington"{tuple_delimiter}"The team receives communications from Washington, which influences their decision-making process."{tuple_delimiter}7){record_delimiter}
+("relationship"{tuple_delimiter}"The team"{tuple_delimiter}"Operation: Dulce"{tuple_delimiter}"The team is directly involved in Operation: Dulce, executing its evolved objectives and activities."{tuple_delimiter}9){completion_delimiter}
+#############################
+Example 3:
+
+Entity_types: [person, role, technology, organization, event, location, concept]
+Text:
+their voice slicing through the buzz of activity. "Control may be an illusion when facing an intelligence that literally writes its own rules," they stated stoically, casting a watchful eye over the flurry of data.
+
+"It's like it's learning to communicate," offered Sam Rivera from a nearby interface, their youthful energy boding a mix of awe and anxiety. "This gives talking to strangers' a whole new meaning."
+
+Alex surveyed his team—each face a study in concentration, determination, and not a small measure of trepidation. "This might well be our first contact," he acknowledged, "And we need to be ready for whatever answers back."
+
+Together, they stood on the edge of the unknown, forging humanity's response to a message from the heavens. The ensuing silence was palpable—a collective introspection about their role in this grand cosmic play, one that could rewrite human history.
+
+The encrypted dialogue continued to unfold, its intricate patterns showing an almost uncanny anticipation
+#############
+Output:
+("entity"{tuple_delimiter}"Sam Rivera"{tuple_delimiter}"person"{tuple_delimiter}"Sam Rivera is a member of a team working on communicating with an unknown intelligence, showing a mix of awe and anxiety."){record_delimiter}
+("entity"{tuple_delimiter}"Alex"{tuple_delimiter}"person"{tuple_delimiter}"Alex is the leader of a team attempting first contact with an unknown intelligence, acknowledging the significance of their task."){record_delimiter}
+("entity"{tuple_delimiter}"Control"{tuple_delimiter}"concept"{tuple_delimiter}"Control refers to the ability to manage or govern, which is challenged by an intelligence that writes its own rules."){record_delimiter}
+("entity"{tuple_delimiter}"Intelligence"{tuple_delimiter}"concept"{tuple_delimiter}"Intelligence here refers to an unknown entity capable of writing its own rules and learning to communicate."){record_delimiter}
+("entity"{tuple_delimiter}"First Contact"{tuple_delimiter}"event"{tuple_delimiter}"First Contact is the potential initial communication between humanity and an unknown intelligence."){record_delimiter}
+("entity"{tuple_delimiter}"Humanity's Response"{tuple_delimiter}"event"{tuple_delimiter}"Humanity's Response is the collective action taken by Alex's team in response to a message from an unknown intelligence."){record_delimiter}
+("relationship"{tuple_delimiter}"Sam Rivera"{tuple_delimiter}"Intelligence"{tuple_delimiter}"Sam Rivera is directly involved in the process of learning to communicate with the unknown intelligence."{tuple_delimiter}9){record_delimiter}
+("relationship"{tuple_delimiter}"Alex"{tuple_delimiter}"First Contact"{tuple_delimiter}"Alex leads the team that might be making the First Contact with the unknown intelligence."{tuple_delimiter}10){record_delimiter}
+("relationship"{tuple_delimiter}"Alex"{tuple_delimiter}"Humanity's Response"{tuple_delimiter}"Alex and his team are the key figures in Humanity's Response to the unknown intelligence."{tuple_delimiter}8){record_delimiter}
+("relationship"{tuple_delimiter}"Control"{tuple_delimiter}"Intelligence"{tuple_delimiter}"The concept of Control is challenged by the Intelligence that writes its own rules."{tuple_delimiter}7){completion_delimiter}
+#############################
+-Real Data-
+######################
+Entity_types: {entity_types}
+Text: {input_text}
+######################
+Output:
+"""
+
+PROMPTS[
+    "summarize_entity_descriptions"
+] = """You are a helpful assistant responsible for generating a comprehensive summary of the data provided below.
+Given one or two entities, and a list of descriptions, all related to the same entity or group of entities.
+Please concatenate all of these into a single, comprehensive description. Make sure to include information collected from all the descriptions.
+If the provided descriptions are contradictory, please resolve the contradictions and provide a single, coherent summary.
+Make sure it is written in third person, and include the entity names so we the have full context.
+
+#######
+-Data-
+Entities: {entity_name}
+Description List: {description_list}
+#######
+Output:
+"""
+
+PROMPTS[
+    "entiti_continue_extraction"
+] = """MANY entities were missed in the last extraction.  Add them below using the same format:
+"""
+
+PROMPTS[
+    "entiti_if_loop_extraction"
+] = """It appears some entities may have still been missed.  Answer YES | NO if there are still entities that need to be added.
+"""
+
+PROMPTS["DEFAULT_ENTITY_TYPES"] = ["organization", "person", "geo", "event"]
+PROMPTS["DEFAULT_TUPLE_DELIMITER"] = "<|>"
+PROMPTS["DEFAULT_RECORD_DELIMITER"] = "##"
+PROMPTS["DEFAULT_COMPLETION_DELIMITER"] = "<|COMPLETE|>"
+PROMPTS["fail_response"] = "Sorry, I'm not able to provide an answer to that question."
+PROMPTS["process_tickers"] = ["⠋", "⠙", "⠹", "⠸", "⠼", "⠴", "⠦", "⠧", "⠇", "⠏"]
+PROMPTS["default_text_separator"] = [
+    # Paragraph separators
+    "\n\n",
+    "\r\n\r\n",
+    # Line breaks
+    "\n",
+    "\r\n",
+    # Sentence ending punctuation
+    "。",  # Chinese period
+    "．",  # Full-width dot
+    ".",  # English period
+    "！",  # Chinese exclamation mark
+    "!",  # English exclamation mark
+    "？",  # Chinese question mark
+    "?",  # English question mark
+    # Whitespace characters
+    " ",  # Space
+    "\t",  # Tab
+    "\u3000",  # Full-width space
+    # Special characters
+    "\u200b",  # Zero-width space (used in some Asian languages)
+]
+
+
+PROMPTS[
+    "naive_rag_response"
+] = """---Role---
+
+You are a helpful assistant responding to a query with retrieved knowledge.
+
+---Goal---
+
+Generate a response of the target length and format that responds to the user's question with relevant general knowledge.
+Summarize useful and relevant information from the input data tables, suitable for the specified response length and format.
+If you don't know the answer or if the provided knowledge do not contain sufficient information to provide an answer, just say so. Do not make anything up.
+Do not include information where the supporting evidence for it is not provided.
+
+---Target response length and format---
+
+{response_type}
+
+---Data tables---
+
+{content_data}
+
+---Goal---
+
+Generate a response of the target length and format that responds to the user's question with relevant general knowledge.
+Summarize useful and relevant information from the input data tables appropriate for the response length and format.
+If you don't know the answer or if the provided knowledge do not contain sufficient information to provide an answer, just say so. Do not make anything up.
+Do not include information where the supporting evidence for it is not provided.
+
+---Notice---
+Please add sections and commentary as appropriate for the length and format if necessary. Format the response in Markdown.
+"""
+
+
+PROMPTS[
+    "query_rewrite_for_entity_retrieval"
+] = """-Goal-
+For a given query, generate a declarative sentence to serve as a query for retrieving relevant knowledge.
+
+######################
+-Examples-
+######################
+
+Question: What are the main characters? \n(A) Alice\n(B) Bob\n(C) Charlie\n(D) Dana
+################
+Output:
+The main characters. (Maybe Alice, Bob, Charlie or Dana)
+
+Question: What locations are shown in the video?
+################
+Output:
+The locations shown in the video.
+
+Question: Which animals appear in the wildlife footage? \n(A) Lions\n(B) Elephants\n(C) Zebras
+################
+Output:
+The animals that appear in the wildlife footage. (Maybe Lions, Elephants or Zebras)
+
+#############################
+-Real Data-
+######################
+Question: {input_text}
+######################
+Output:
+"""
+
+
+
+PROMPTS[
+    "query_rewrite_for_visual_retrieval"
+] = """-Goal-
+Given a question that may include scene-related information, generate a declarative sentence to serve as a query for retrieving relevant video segments.
+
+######################
+-Examples-
+######################
+
+Question: Which animal does the protagonist encounter in the forest scene?
+################
+Output:
+The protagonist encounters an animal in the forest.
+
+Question: In the movie, what color is the car that chases the main character through the city?
+################
+Output:
+A city chase scene where the main character is pursued by a car.
+
+Question: What is the weather like during the opening scene of the film?\n(A) Sunny\n(B) Rainy\n(C) Snowy\n(D) Windy
+################
+Output:
+The opening scene of the film featuring specific weather conditions. (Maybe Sunny, Rainy, Snowy or Windy)
+
+#############################
+-Real Data-
+######################
+Question: {input_text}
+######################
+Output:
+"""
+
+
+
+PROMPTS[
+    "keywords_extraction"
+] = """- Goal -
+Given a query, extract the relevant keywords that can help answer the query. Please list the keywords separated by commas.
+
+######################
+- Examples -
+######################
+
+Question: Which animal does the protagonist encounter in the forest scene?
+################
+Output:
+animal, protagonist, forest, scene
+
+Question: In the movie, what color is the car that chases the main character through the city?
+################
+Output:
+color, car, chases, main character, city
+
+Question: What is the weather like during the opening scene of the film?\n(A) Sunny\n(B) Rainy\n(C) Snowy\n(D) Windy
+################
+Output:
+weather, opening scene, film, Sunny, Rainy, Snowy, Windy
+
+#############################
+- Real Data -
+######################
+Question: {input_text}
+######################
+Output:
+"""
+
+
+
+PROMPTS[
+    "filtering_segment"
+] = """---Role---
+
+You are a helpful assistant to determine whether the video may contain information relevant to the knowledge based on its rough caption.
+Please note that this is a rough caption of the video segments, which means it may not directly contain the answer but may indicate that the video segment is likely to contain information relevant to answering the question. 
+
+---Video Caption---
+
+{caption}
+
+---Knowledge We Need---
+
+{knowledge}
+
+---Answer---
+Please provide an answer that begins with "yes" or "no," followed by a brief step-by-step explanation.
+Answer:
+"""
+
+
+
+PROMPTS[
+    "videorag_response"
+] = """---Role---
+
+You are a helpful assistant responding to a query with retrieved knowledge.
+
+---Goal---
+
+Generate a response of the target length and format that responds to the user's question with relevant general knowledge.
+Summarize useful and relevant information from the retrieved text chunks and the information retrieved from videos, suitable for the specified response length and format.
+If you don't know the answer or if the input data tables do not contain sufficient information to provide an answer, just say so. Do not make anything up.
+Do not include information where the supporting evidence for it is not provided.
+
+---Target response length and format---
+
+{response_type}
+
+---Retrieved Information From Videos---
+
+{video_data}
+
+---Retrieved Text Chunks---
+
+{chunk_data}
+
+---Goal---
+
+Generate a response of the target length and format that responds to the user's question with relevant general knowledge.
+Summarize useful and relevant information from the retrieved text chunks and the information retrieved from videos, suitable for the specified response length and format.
+If you don't know the answer or if the input data tables do not contain sufficient information to provide an answer, just say so. Do not make anything up.
+Do not include information where the supporting evidence for it is not provided.
+Reference relevant video segments within the answers, specifying the video name and start & end timestamps. Use the following reference format:
+
+---Example of Reference---
+
+In one segment, the film highlights the devastating effects of deforestation on wildlife habitats [1]. Another part illustrates successful conservation efforts that have helped endangered species recover [2].
+
+#### Reference:
+[1] video_name_1, 05:30, 08:00  
+[2] video_name_2, 25:00, 28:00 
+
+---Notice---
+Please add sections and commentary as appropriate for the length and format if necessary. Format the response in Markdown.
+"""
+
+PROMPTS[
+    "videorag_response_wo_reference"
+] = """---Role---
+
+You are a helpful assistant responding to a query with retrieved knowledge.
+
+---Goal---
+
+Generate a response of the target length and format that responds to the user's question with relevant general knowledge.
+Summarize useful and relevant information from the retrieved text chunks and the information retrieved from videos, suitable for the specified response length and format.
+If you don't know the answer or if the input data tables do not contain sufficient information to provide an answer, just say so. Do not make anything up.
+Do not include information where the supporting evidence for it is not provided.
+
+---Target response length and format---
+
+{response_type}
+
+---Retrieved Information From Videos---
+
+{video_data}
+
+---Retrieved Text Chunks---
+
+{chunk_data}
+
+---Goal---
+
+Generate a response of the target length and format that responds to the user's question with relevant general knowledge.
+Summarize useful and relevant information from the retrieved text chunks and the information retrieved from videos, suitable for the specified response length and format.
+If you don't know the answer or if the input data tables do not contain sufficient information to provide an answer, just say so. Do not make anything up.
+Do not include information where the supporting evidence for it is not provided.
+
+---Notice---
+Please add sections and commentary as appropriate for the length and format if necessary. Format the response in Markdown.
+"""

videorag/videorag.py

@@ -0,0 +1,441 @@
+import os
+import sys
+import json
+import shutil
+import asyncio
+import multiprocessing
+from dataclasses import asdict, dataclass, field
+from datetime import datetime
+from functools import partial
+from typing import Callable, Dict, List, Optional, Type, Union, cast
+from transformers import AutoModel, AutoTokenizer
+import tiktoken
+
+
+from ._llm import (
+    gpt_4o_complete,
+    gpt_4o_mini_complete,
+    openai_embedding,
+    azure_gpt_4o_complete,
+    azure_openai_embedding,
+    azure_gpt_4o_mini_complete,
+)
+from ._op import (
+    chunking_by_video_segments,
+    extract_entities,
+    get_chunks,
+    videorag_query,
+    naive_query,
+    videorag_query_wo_graph,
+    videorag_query_wo_vision,
+)
+from ._storage import (
+    JsonKVStorage,
+    NanoVectorDBStorage,
+    NanoVectorDBVideoSegmentStorage,
+    NetworkXStorage,
+)
+from ._utils import (
+    EmbeddingFunc,
+    compute_mdhash_id,
+    limit_async_func_call,
+    convert_response_to_json,
+    always_get_an_event_loop,
+    logger,
+)
+from .base import (
+    BaseGraphStorage,
+    BaseKVStorage,
+    BaseVectorStorage,
+    StorageNameSpace,
+    QueryParam,
+)
+from ._videoutil import(
+    split_video,
+    speech_to_text,
+    segment_caption,
+    merge_segment_information,
+    saving_video_segments,
+)
+
+
+@dataclass
+class VideoRAG:
+    working_dir: str = field(
+        default_factory=lambda: f"./videorag_cache_{datetime.now().strftime('%Y-%m-%d-%H:%M:%S')}"
+    )
+    
+    # video
+    threads_for_split: int = 10
+    video_segment_length: int = 30 # seconds
+    rough_num_frames_per_segment: int = 5 # frames
+    fine_num_frames_per_segment: int = 15 # frames
+    video_output_format: str = "mp4"
+    audio_output_format: str = "mp3"
+    video_embedding_batch_num: int = 2
+    segment_retrieval_top_k: int = 4
+    video_embedding_dim: int = 1024
+    
+    # query
+    retrieval_topk_chunks: int = 2
+    
+    # graph mode
+    enable_local: bool = True
+    enable_naive_rag: bool = True
+
+    # text chunking
+    chunk_func: Callable[
+        [
+            list[list[int]],
+            List[str],
+            tiktoken.Encoding,
+            Optional[int],
+        ],
+        List[Dict[str, Union[str, int]]],
+    ] = chunking_by_video_segments
+    chunk_token_size: int = 1200
+    # chunk_overlap_token_size: int = 100
+    tiktoken_model_name: str = "gpt-4o"
+
+    # entity extraction
+    entity_extract_max_gleaning: int = 1
+    entity_summary_to_max_tokens: int = 500
+
+    # text embedding
+    embedding_func: EmbeddingFunc = field(default_factory=lambda: openai_embedding)
+    embedding_batch_num: int = 32
+    embedding_func_max_async: int = 16
+    query_better_than_threshold: float = 0.2
+
+    # LLM
+    using_azure_openai: bool = False
+    best_model_func: callable = gpt_4o_mini_complete
+    best_model_max_token_size: int = 32768
+    best_model_max_async: int = 16
+    cheap_model_func: callable = gpt_4o_mini_complete
+    cheap_model_max_token_size: int = 32768
+    cheap_model_max_async: int = 16
+
+    # entity extraction
+    entity_extraction_func: callable = extract_entities
+    
+    # storage
+    key_string_value_json_storage_cls: Type[BaseKVStorage] = JsonKVStorage
+    vector_db_storage_cls: Type[BaseVectorStorage] = NanoVectorDBStorage
+    vs_vector_db_storage_cls: Type[BaseVectorStorage] = NanoVectorDBVideoSegmentStorage
+    vector_db_storage_cls_kwargs: dict = field(default_factory=dict)
+    graph_storage_cls: Type[BaseGraphStorage] = NetworkXStorage
+    enable_llm_cache: bool = True
+
+    # extension
+    always_create_working_dir: bool = True
+    addon_params: dict = field(default_factory=dict)
+    convert_response_to_json_func: callable = convert_response_to_json
+
+    def load_caption_model(self, debug=False):
+        # caption model
+        if not debug:
+            self.caption_model = AutoModel.from_pretrained('./MiniCPM-V-2_6-int4', trust_remote_code=True)
+            self.caption_tokenizer = AutoTokenizer.from_pretrained('./MiniCPM-V-2_6-int4', trust_remote_code=True)
+            self.caption_model.eval()
+        else:
+            self.caption_model = None
+            self.caption_tokenizer = None
+    
+    def __post_init__(self):
+        _print_config = ",\n  ".join([f"{k} = {v}" for k, v in asdict(self).items()])
+        logger.debug(f"VideoRAG init with param:\n\n  {_print_config}\n")
+        
+        if self.using_azure_openai:
+            # If there's no OpenAI API key, use Azure OpenAI
+            if self.best_model_func == gpt_4o_complete:
+                self.best_model_func = azure_gpt_4o_complete
+            if self.cheap_model_func == gpt_4o_mini_complete:
+                self.cheap_model_func = azure_gpt_4o_mini_complete
+            if self.embedding_func == openai_embedding:
+                self.embedding_func = azure_openai_embedding
+            logger.info(
+                "Switched the default openai funcs to Azure OpenAI if you didn't set any of it"
+            )
+
+        if not os.path.exists(self.working_dir) and self.always_create_working_dir:
+            logger.info(f"Creating working directory {self.working_dir}")
+            os.makedirs(self.working_dir)
+
+        self.video_path_db = self.key_string_value_json_storage_cls(
+            namespace="video_path", global_config=asdict(self)
+        )
+        
+        self.video_segments = self.key_string_value_json_storage_cls(
+            namespace="video_segments", global_config=asdict(self)
+        )
+
+        self.text_chunks = self.key_string_value_json_storage_cls(
+            namespace="text_chunks", global_config=asdict(self)
+        )
+
+        self.llm_response_cache = (
+            self.key_string_value_json_storage_cls(
+                namespace="llm_response_cache", global_config=asdict(self)
+            )
+            if self.enable_llm_cache
+            else None
+        )
+
+        self.chunk_entity_relation_graph = self.graph_storage_cls(
+            namespace="chunk_entity_relation", global_config=asdict(self)
+        )
+
+        self.embedding_func = limit_async_func_call(self.embedding_func_max_async)(
+            self.embedding_func
+        )
+        self.entities_vdb = (
+            self.vector_db_storage_cls(
+                namespace="entities",
+                global_config=asdict(self),
+                embedding_func=self.embedding_func,
+                meta_fields={"entity_name"},
+            )
+            if self.enable_local
+            else None
+        )
+        self.chunks_vdb = (
+            self.vector_db_storage_cls(
+                namespace="chunks",
+                global_config=asdict(self),
+                embedding_func=self.embedding_func,
+            )
+            if self.enable_naive_rag
+            else None
+        )
+        
+        self.video_segment_feature_vdb = (
+            self.vs_vector_db_storage_cls(
+                namespace="video_segment_feature",
+                global_config=asdict(self),
+                embedding_func=None, # we code the embedding process inside the insert() function.
+            )
+        )
+        
+        self.best_model_func = limit_async_func_call(self.best_model_max_async)(
+            partial(self.best_model_func, hashing_kv=self.llm_response_cache)
+        )
+        self.cheap_model_func = limit_async_func_call(self.cheap_model_max_async)(
+            partial(self.cheap_model_func, hashing_kv=self.llm_response_cache)
+        )
+
+    def insert_video(self, video_path_list=None):
+        loop = always_get_an_event_loop()
+        for video_path in video_path_list:
+            # Step0: check the existence
+            video_name = os.path.basename(video_path).split('.')[0]
+            if video_name in self.video_segments._data:
+                logger.info(f"Find the video named {os.path.basename(video_path)} in storage and skip it.")
+                continue
+            loop.run_until_complete(self.video_path_db.upsert(
+                {video_name: video_path}
+            ))
+            
+            # Step1: split the videos
+            segment_index2name, segment_times_info = split_video(
+                video_path, 
+                self.working_dir, 
+                self.video_segment_length,
+                self.rough_num_frames_per_segment,
+                self.audio_output_format,
+            )
+            
+            # Step2: obtain transcript with whisper
+            transcripts = speech_to_text(
+                video_name, 
+                self.working_dir, 
+                segment_index2name,
+                self.audio_output_format
+            )
+            
+            # Step3: saving video segments **as well as** obtain caption with vision language model
+            manager = multiprocessing.Manager()
+            captions = manager.dict()
+            error_queue = manager.Queue()
+            
+            process_saving_video_segments = multiprocessing.Process(
+                target=saving_video_segments,
+                args=(
+                    video_name,
+                    video_path,
+                    self.working_dir,
+                    segment_index2name,
+                    segment_times_info,
+                    error_queue,
+                    self.video_output_format,
+                )
+            )
+            
+            process_segment_caption = multiprocessing.Process(
+                target=segment_caption,
+                args=(
+                    video_name,
+                    video_path,
+                    segment_index2name,
+                    transcripts,
+                    segment_times_info,
+                    captions,
+                    error_queue,
+                )
+            )
+            
+            process_saving_video_segments.start()
+            process_segment_caption.start()
+            process_saving_video_segments.join()
+            process_segment_caption.join()
+            
+            # if raise error in this two, stop the processing
+            while not error_queue.empty():
+                error_message = error_queue.get()
+                with open('error_log_videorag.txt', 'a', encoding='utf-8') as log_file:
+                    log_file.write(f"Video Name:{video_name} Error processing:\n{error_message}\n\n")
+                raise RuntimeError(error_message)
+            
+            # Step4: insert video segments information
+            segments_information = merge_segment_information(
+                segment_index2name,
+                segment_times_info,
+                transcripts,
+                captions,
+            )
+            manager.shutdown()
+            loop.run_until_complete(self.video_segments.upsert(
+                {video_name: segments_information}
+            ))
+            
+            # Step5: encode video segment features
+            loop.run_until_complete(self.video_segment_feature_vdb.upsert(
+                video_name,
+                segment_index2name,
+                self.video_output_format,
+            ))
+            
+            # Step6: delete the cache file
+            video_segment_cache_path = os.path.join(self.working_dir, '_cache', video_name)
+            if os.path.exists(video_segment_cache_path):
+                shutil.rmtree(video_segment_cache_path)
+            
+            # Step 7: saving current video information
+            loop.run_until_complete(self._save_video_segments())
+        
+        loop.run_until_complete(self.ainsert(self.video_segments._data))
+
+    def query(self, query: str, param: QueryParam = QueryParam()):
+        loop = always_get_an_event_loop()
+        return loop.run_until_complete(self.aquery(query, param))
+
+    async def aquery(self, query: str, param: QueryParam = QueryParam()):
+        if param.mode == "videorag":
+            response = await videorag_query(
+                query,
+                self.entities_vdb,
+                self.text_chunks,
+                self.chunks_vdb,
+                self.video_path_db,
+                self.video_segments,
+                self.video_segment_feature_vdb,
+                self.chunk_entity_relation_graph,
+                self.caption_model, 
+                self.caption_tokenizer,
+                param,
+                asdict(self),
+            )
+        else:
+            raise ValueError(f"Unknown mode {param.mode}")
+        await self._query_done()
+        return response
+
+    async def ainsert(self, new_video_segment):
+        await self._insert_start()
+        try:
+            # ---------- chunking
+            inserting_chunks = get_chunks(
+                new_videos=new_video_segment,
+                chunk_func=self.chunk_func,
+                max_token_size=self.chunk_token_size,
+            )
+            _add_chunk_keys = await self.text_chunks.filter_keys(
+                list(inserting_chunks.keys())
+            )
+            inserting_chunks = {
+                k: v for k, v in inserting_chunks.items() if k in _add_chunk_keys
+            }
+            if not len(inserting_chunks):
+                logger.warning(f"All chunks are already in the storage")
+                return
+            logger.info(f"[New Chunks] inserting {len(inserting_chunks)} chunks")
+            if self.enable_naive_rag:
+                logger.info("Insert chunks for naive RAG")
+                await self.chunks_vdb.upsert(inserting_chunks)
+
+            # TODO: no incremental update for communities now, so just drop all
+            # await self.community_reports.drop()
+
+            # ---------- extract/summary entity and upsert to graph
+            logger.info("[Entity Extraction]...")
+            maybe_new_kg, _, _ = await self.entity_extraction_func(
+                inserting_chunks,
+                knowledge_graph_inst=self.chunk_entity_relation_graph,
+                entity_vdb=self.entities_vdb,
+                global_config=asdict(self),
+            )
+            if maybe_new_kg is None:
+                logger.warning("No new entities found")
+                return
+            self.chunk_entity_relation_graph = maybe_new_kg
+            # ---------- commit upsertings and indexing
+            await self.text_chunks.upsert(inserting_chunks)
+        finally:
+            await self._insert_done()
+
+    async def _insert_start(self):
+        tasks = []
+        for storage_inst in [
+            self.chunk_entity_relation_graph,
+        ]:
+            if storage_inst is None:
+                continue
+            tasks.append(cast(StorageNameSpace, storage_inst).index_start_callback())
+        await asyncio.gather(*tasks)
+
+    async def _save_video_segments(self):
+        tasks = []
+        for storage_inst in [
+            self.video_segment_feature_vdb,
+            self.video_segments,
+            self.video_path_db,
+        ]:
+            if storage_inst is None:
+                continue
+            tasks.append(cast(StorageNameSpace, storage_inst).index_done_callback())
+        await asyncio.gather(*tasks)
+    
+    async def _insert_done(self):
+        tasks = []
+        for storage_inst in [
+            self.text_chunks,
+            self.llm_response_cache,
+            self.entities_vdb,
+            self.chunks_vdb,
+            self.chunk_entity_relation_graph,
+            self.video_segment_feature_vdb,
+            self.video_segments,
+            self.video_path_db,
+        ]:
+            if storage_inst is None:
+                continue
+            tasks.append(cast(StorageNameSpace, storage_inst).index_done_callback())
+        await asyncio.gather(*tasks)
+
+    async def _query_done(self):
+        tasks = []
+        for storage_inst in [self.llm_response_cache]:
+            if storage_inst is None:
+                continue
+            tasks.append(cast(StorageNameSpace, storage_inst).index_done_callback())
+        await asyncio.gather(*tasks)