Romain Beaumont
800104efa7
Fix publish yml
2022-04-09 14:21:44 -07:00
Ross Wightman
b7ce3897a2
Release 0.2.1
...
Update README.md
2022-04-08 09:32:11 -07:00
Romain Beaumont
dd980df0b8
Add github action to automatically push to pypi on Release x.y.z commit
2022-04-08 08:46:13 -07:00
Gabriel Ilharco Magalhães
f512ba39ac
Update README.md
2022-04-07 08:58:22 -07:00
Ross Wightman
2856dad5b6
Add support for locked image towers and pretrained image tower weight load (timm models only)
2022-04-06 16:07:11 -07:00
Romain Beaumont
8f24b79092
Add basic continuous integration.
2022-04-06 14:56:33 -07:00
Ross Wightman
5592f38a05
Merge pull request #54 from rom1504/readme_colab_diagram
...
Add colab, example in readme and architecture diagram
2022-04-06 14:19:41 -07:00
Romain Beaumont
d45a3c7c39
Add colab, example in readme and architecture diagram
2022-04-06 23:14:08 +02:00
Ross Wightman
db0429cf25
Merge pull request #51 from mlfoundations/trace_jit_support
...
torch.jit tracing and scripting support w/ attn_mask as input arg
2022-04-06 14:04:45 -07:00
Ross Wightman
ae6e721786
Move timm model wrapper to own file, add improved OpenAI attribution to relevant files
2022-04-06 10:07:27 -07:00
Ross Wightman
41159c154a
Make wandb dependency optional
2022-04-06 09:32:54 -07:00
Ross Wightman
aec3dca802
Make tensorboard dependency optional
2022-04-06 09:23:44 -07:00
Ross Wightman
3ea395e2e9
Remove dead attn_mask code from model
2022-04-06 09:23:28 -07:00
Ludwig Schmidt
9681faac96
Update README.md
2022-04-04 21:27:12 -07:00
Ludwig Schmidt
d26aed6feb
Update README.md
2022-04-04 21:24:28 -07:00
Ross Wightman
28fbf4a86c
open_clip pypi name conflicts with existing 'openclip', pypi name now open_clip_torch
2022-04-04 14:27:24 -07:00
Ross Wightman
4ae357f47a
version 0.2.1
2022-04-04 13:10:07 -07:00
Ross Wightman
2ece4ef047
Minor zero_shot tweak including batch_size for tqdm so it/sec == samp/sec
2022-04-04 13:09:22 -07:00
Ross Wightman
f447bd144c
Add model jit tracing and torchscript support via flags and model changes
...
* attn_mask moved to input arg and non-persistent buffer
* ddp static_graph support added
2022-04-04 12:50:00 -07:00
Ross Wightman
bc8b3a23ce
Update README.md
2022-04-03 22:14:15 -07:00
Ross Wightman
3055617f23
Update README.md
2022-04-03 22:00:58 -07:00
Ross Wightman
98fc29079e
Update README.md
2022-04-03 21:50:28 -07:00
Ross Wightman
1c8647f14f
Merge pull request #50 from rwightman/refactor_cleanup
...
Cleanup and refactoring for model release
2022-04-03 21:30:01 -07:00
Zasder3
b1af00906a
Update README.md
2022-04-04 06:21:35 +02:00
Ross Wightman
fc45b31051
Complete the rest of last commit, missed quickgelu json and updated list_models()
2022-04-01 15:16:27 -07:00
Ross Wightman
c1149e55c7
Add pretrained weights, adjust model config to scan at startup, allow adding extra configs. Add 'quickgelu' config variants to avoid requiring --force-quick-gelu flag for all current weights.
2022-04-01 15:09:34 -07:00
Ross Wightman
adef9b9e2a
add create_model w/o transforms, fix openai jit mode for recent pytorch versions
2022-03-30 22:34:23 -07:00
Ross Wightman
1cd9fc48cc
wandb / tb loggers use non .avg versions of metered metrics
2022-03-30 16:39:16 -07:00
Ross Wightman
a3b4c26df7
Move src/clip -> src/open_clip, create setup.py + MANIFEST.in and test package install as open_clip
2022-03-30 16:32:51 -07:00
Ross Wightman
db8f6dfcef
clip/openai_clip.py -> clip/openai.py
2022-03-30 11:30:37 -07:00
Ross Wightman
2e8f3a48c3
Move ClipLoss to clip package, enable label caching for train
2022-03-30 11:25:38 -07:00
Ross Wightman
d8084a6d1f
Fix per worker batch counting for wds data pipeline. Improve train loop logging.
2022-03-30 10:41:05 -07:00
Ross Wightman
9362b5d32d
Add the new clip package files that were missed in previous commit.
2022-03-29 22:33:26 -07:00
Ross Wightman
337e84442d
Refine timm model adapter for improved handling of global pooling
2022-03-29 22:14:56 -07:00
Ross Wightman
4a246ea7d6
More reorganization and refactor
...
* 'clip' package more stand-alone, with model factory and configs
* add string based pretain tags for openai and future pretrain on different datasets
2022-03-29 22:12:03 -07:00
Ross Wightman
4304870b17
Adjust train vs eval wds num_batches calc. with_epoch before loader.
2022-03-29 16:54:49 -07:00
Ross Wightman
d023b3a682
Webdataset 0.2.x update
...
* add initial wds 0.2.x support (needs testing)
* add flags for specifing dataset samples manually
* AMP used for zero-shot / eval if enabled
* eval accumulates on CPU (only slightly better..)
2022-03-29 08:01:27 -07:00
Ross Wightman
bc39ec984e
Model cleanup
...
* remove some old typing
* fix quick GELU toggle
* give transforms fn public name
* grad scaler checkpoint save / restore
2022-03-29 08:00:32 -07:00
Gabriel Ilharco Magalhães
db7504f070
Merge pull request #49 from mlfoundations/mitchellnw-patch-1
...
Update README.md
2022-03-28 12:13:54 -07:00
Mitchell Wortsman
fdf513e9d7
Update README.md
2022-03-28 11:59:51 -07:00
Gabriel Ilharco Magalhães
a9f472f4fa
Update README.md
2022-03-28 08:49:58 -07:00
Gabriel Ilharco Magalhães
9bfd0f3b09
Merge pull request #48 from rwightman/cleanup_remove_dp
...
Remove DataParallel support. Minor cleanup.
2022-03-28 08:35:36 -07:00
Ross Wightman
c576b04d17
Remove broken DataParallel support. Slower than DDP on single node, causes issues with jit, and likely to be deprecated in future PyTorch.
2022-03-25 14:33:44 -07:00
Ross Wightman
e486c74579
Fix B-16 and L-14 config json
2022-03-25 14:31:18 -07:00
Mitchell Wortsman
7d03adb5e0
Merge pull request #47 from rwightman/fix_openai
...
Use new cfg interface for OpenAI model loading. Fix #46 .
2022-03-25 14:07:24 -07:00
Ross Wightman
87489decd0
Add ViT-L-14 config for train
2022-03-25 14:04:58 -07:00
Ross Wightman
421b73e6b1
OpenAI weights should be used with QuickGELU
2022-03-25 14:04:38 -07:00
Ross Wightman
c9bf60977d
Fix pretrained weight keys to change path/uri unfriendly / to -
2022-03-25 14:03:07 -07:00
Ross Wightman
fcad1bf648
Use new cfg interface for OpenAI model loading. Fix #46 .
2022-03-25 13:07:45 -07:00
Vaishaal
6e6c2473c0
Merge pull request #44 from mlfoundations/rw-refactor
...
Cade/Ross' changes (after Gabe's comments)
2022-03-15 10:02:07 -07:00
