Files
deep-learning-with-python-n…/chapter09_part03_interpreting-what-convnets-learn.ipynb
2021-06-05 20:28:07 -07:00

785 lines
19 KiB
Plaintext

{
"cells": [
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"This is a companion notebook for the book [Deep Learning with Python, Second Edition](https://www.manning.com/books/deep-learning-with-python-second-edition?a_aid=keras&a_bid=76564dff). For readability, it only contains runnable code blocks and section titles, and omits everything else in the book: text paragraphs, figures, and pseudocode.\n\n**If you want to be able to follow what's going on, I recommend reading the notebook side by side with your copy of the book.**\n\nThis notebook was generated for TensorFlow 2.6."
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"## Interpreting what convnets learn"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"### Visualizing intermediate activations"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"# You can use this to load the file \"convnet_from_scratch_with_augmentation.keras\"\n",
"# you obtained in the last chapter.\n",
"from google.colab import files\n",
"files.upload()"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"from tensorflow import keras\n",
"model = keras.models.load_model(\"convnet_from_scratch_with_augmentation.keras\")\n",
"model.summary()"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Preprocessing a single image**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"from tensorflow import keras\n",
"import numpy as np\n",
"\n",
"img_path = keras.utils.get_file(\n",
" fname=\"cat.jpg\",\n",
" origin=\"https://img-datasets.s3.amazonaws.com/cat.jpg\")\n",
"\n",
"def get_img_array(img_path, target_size):\n",
" img = keras.preprocessing.image.load_img(\n",
" img_path, target_size=target_size)\n",
" array = keras.preprocessing.image.img_to_array(img)\n",
" array = np.expand_dims(array, axis=0)\n",
" return array\n",
"\n",
"img_tensor = get_img_array(img_path, target_size=(180, 180))"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Displaying the test picture**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"import matplotlib.pyplot as plt\n",
"plt.axis(\"off\")\n",
"plt.imshow(img_tensor[0].astype(\"uint8\"))\n",
"plt.show()"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Instantiating a model that returns layer activations**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"from tensorflow.keras import layers\n",
"\n",
"layer_outputs = []\n",
"layer_names = []\n",
"for layer in model.layers:\n",
" if isinstance(layer, (layers.Conv2D, layers.MaxPooling2D)):\n",
" layer_outputs.append(layer.output)\n",
" layer_names.append(layer.name)\n",
"activation_model = keras.Model(inputs=model.input, outputs=layer_outputs)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Using the model to compute layer activations**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"activations = activation_model.predict(img_tensor)"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"first_layer_activation = activations[0]\n",
"print(first_layer_activation.shape)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Visualizing the fifth channel**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"import matplotlib.pyplot as plt\n",
"plt.matshow(first_layer_activation[0, :, :, 5], cmap=\"viridis\")"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Visualizing every channel in every intermediate activation**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"images_per_row = 16\n",
"for layer_name, layer_activation in zip(layer_names, activations):\n",
" n_features = layer_activation.shape[-1]\n",
" size = layer_activation.shape[1]\n",
" n_cols = n_features // images_per_row\n",
" display_grid = np.zeros(((size + 1) * n_cols - 1,\n",
" images_per_row * (size + 1) - 1))\n",
" for col in range(n_cols):\n",
" for row in range(images_per_row):\n",
" channel_index = col * images_per_row + row\n",
" channel_image = layer_activation[0, :, :, channel_index].copy()\n",
" if channel_image.sum() != 0:\n",
" channel_image -= channel_image.mean()\n",
" channel_image /= channel_image.std()\n",
" channel_image *= 64\n",
" channel_image += 128\n",
" channel_image = np.clip(channel_image, 0, 255).astype(\"uint8\")\n",
" display_grid[\n",
" col * (size + 1): (col + 1) * size + col,\n",
" row * (size + 1) : (row + 1) * size + row] = channel_image\n",
" scale = 1. / size\n",
" plt.figure(figsize=(scale * display_grid.shape[1],\n",
" scale * display_grid.shape[0]))\n",
" plt.title(layer_name)\n",
" plt.grid(False)\n",
" plt.axis(\"off\")\n",
" plt.imshow(display_grid, aspect=\"auto\", cmap=\"viridis\")"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"### Visualizing convnet filters"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Instantiating the Xception convolutional base**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"model = keras.applications.xception.Xception(\n",
" weights=\"imagenet\",\n",
" include_top=False)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Printing the names of all convolutional layers in Xception**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"for layer in model.layers:\n",
" if isinstance(layer, (keras.layers.Conv2D, keras.layers.SeparableConv2D)):\n",
" print(layer.name)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Creating a \"feature extractor\" model that returns the output of a specific layer**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"layer_name = \"block3_sepconv1\"\n",
"layer = model.get_layer(name=layer_name)\n",
"feature_extractor = keras.Model(inputs=model.input, outputs=layer.output)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Using the feature extractor**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"activation = feature_extractor(\n",
" keras.applications.xception.preprocess_input(img_tensor)\n",
")"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"import tensorflow as tf\n",
"\n",
"def compute_loss(image, filter_index):\n",
" activation = feature_extractor(image)\n",
" filter_activation = activation[:, 2:-2, 2:-2, filter_index]\n",
" return tf.reduce_mean(filter_activation)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Loss maximization via stochastic gradient ascent**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"@tf.function\n",
"def gradient_ascent_step(image, filter_index, learning_rate):\n",
" with tf.GradientTape() as tape:\n",
" tape.watch(image)\n",
" loss = compute_loss(image, filter_index)\n",
" grads = tape.gradient(loss, image)\n",
" grads = tf.math.l2_normalize(grads)\n",
" image += learning_rate * grads\n",
" return image"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Function to generate filter visualizations**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"img_width = 200\n",
"img_height = 200\n",
"\n",
"def generate_filter_pattern(filter_index):\n",
" iterations = 30\n",
" learning_rate = 10.\n",
" image = tf.random.uniform(\n",
" minval=0.4,\n",
" maxval=0.6,\n",
" shape=(1, img_width, img_height, 3))\n",
" for i in range(iterations):\n",
" image = gradient_ascent_step(image, filter_index, learning_rate)\n",
" return image[0].numpy()"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Utility function to convert a tensor into a valid image**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"def deprocess_image(image):\n",
" image -= image.mean()\n",
" image /= image.std()\n",
" image *= 64\n",
" image += 128\n",
" image = np.clip(image, 0, 255).astype(\"uint8\")\n",
" image = image[25:-25, 25:-25, :]\n",
" return image"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"plt.axis(\"off\")\n",
"plt.imshow(deprocess_image(generate_filter_pattern(filter_index=2)))"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Generating a grid of all filter response patterns in a layer**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"all_images = []\n",
"for filter_index in range(64):\n",
" print(f\"Processing filter {filter_index}\")\n",
" image = deprocess_image(\n",
" generate_filter_pattern(filter_index)\n",
" )\n",
" all_images.append(image)\n",
"\n",
"margin = 5\n",
"n = 8\n",
"cropped_width = img_width - 25 * 2\n",
"cropped_height = img_height - 25 * 2\n",
"width = n * cropped_width + (n - 1) * margin\n",
"height = n * cropped_height + (n - 1) * margin\n",
"stitched_filters = np.zeros((width, height, 3))\n",
"\n",
"for i in range(n):\n",
" for j in range(n):\n",
" image = all_images[i * n + j]\n",
" stitched_filters[\n",
" (cropped_width + margin) * i : (cropped_width + margin) * i + cropped_width,\n",
" (cropped_height + margin) * j : (cropped_height + margin) * j\n",
" + cropped_height,\n",
" :,\n",
" ] = image\n",
"\n",
"keras.preprocessing.image.save_img(\n",
" f\"filters_for_layer_{layer_name}.png\", stitched_filters)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"### Visualizing heatmaps of class activation"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Loading the Xception network with pretrained weights**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"model = keras.applications.xception.Xception(weights=\"imagenet\")"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Preprocessing an input image for Xception**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"img_path = keras.utils.get_file(\n",
" fname=\"elephant.jpg\",\n",
" origin=\"https://img-datasets.s3.amazonaws.com/elephant.jpg\")\n",
"\n",
"def get_img_array(img_path, target_size):\n",
" img = keras.preprocessing.image.load_img(img_path, target_size=target_size)\n",
" array = keras.preprocessing.image.img_to_array(img)\n",
" array = np.expand_dims(array, axis=0)\n",
" array = keras.applications.xception.preprocess_input(array)\n",
" return array\n",
"\n",
"img_array = get_img_array(img_path, target_size=(299, 299))"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"preds = model.predict(img_array)\n",
"print(keras.applications.xception.decode_predictions(preds, top=3)[0])"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"np.argmax(preds[0])"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Setting up a model that returns the last convolutional output**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"last_conv_layer_name = \"block14_sepconv2_act\"\n",
"classifier_layer_names = [\n",
" \"avg_pool\",\n",
" \"predictions\",\n",
"]\n",
"last_conv_layer = model.get_layer(last_conv_layer_name)\n",
"last_conv_layer_model = keras.Model(model.inputs, last_conv_layer.output)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Setting up a model that goes from the last convolutional output to the final predictions**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"classifier_input = keras.Input(shape=last_conv_layer.output.shape[1:])\n",
"x = classifier_input\n",
"for layer_name in classifier_layer_names:\n",
" x = model.get_layer(layer_name)(x)\n",
"classifier_model = keras.Model(classifier_input, x)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Retrieving the gradients of the top predicted class with regard to the last convolutional output**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"import tensorflow as tf\n",
"\n",
"with tf.GradientTape() as tape:\n",
" last_conv_layer_output = last_conv_layer_model(img_array)\n",
" tape.watch(last_conv_layer_output)\n",
" preds = classifier_model(last_conv_layer_output)\n",
" top_pred_index = tf.argmax(preds[0])\n",
" top_class_channel = preds[:, top_pred_index]\n",
"\n",
"grads = tape.gradient(top_class_channel, last_conv_layer_output)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Gradient pooling and channel importance weighting**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"pooled_grads = tf.reduce_mean(grads, axis=(0, 1, 2)).numpy()\n",
"last_conv_layer_output = last_conv_layer_output.numpy()[0]\n",
"for i in range(pooled_grads.shape[-1]):\n",
" last_conv_layer_output[:, :, i] *= pooled_grads[i]\n",
"heatmap = np.mean(last_conv_layer_output, axis=-1)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Heatmap post-processing**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"heatmap = np.maximum(heatmap, 0)\n",
"heatmap /= np.max(heatmap)\n",
"plt.matshow(heatmap)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"**Superimposing the heatmap with the original picture**"
]
},
{
"cell_type": "code",
"execution_count": 0,
"metadata": {
"colab_type": "code"
},
"outputs": [],
"source": [
"import matplotlib.cm as cm\n",
"\n",
"img = keras.preprocessing.image.load_img(img_path)\n",
"img = keras.preprocessing.image.img_to_array(img)\n",
"\n",
"heatmap = np.uint8(255 * heatmap)\n",
"\n",
"jet = cm.get_cmap(\"jet\")\n",
"jet_colors = jet(np.arange(256))[:, :3]\n",
"jet_heatmap = jet_colors[heatmap]\n",
"\n",
"jet_heatmap = keras.preprocessing.image.array_to_img(jet_heatmap)\n",
"jet_heatmap = jet_heatmap.resize((img.shape[1], img.shape[0]))\n",
"jet_heatmap = keras.preprocessing.image.img_to_array(jet_heatmap)\n",
"\n",
"superimposed_img = jet_heatmap * 0.4 + img\n",
"superimposed_img = keras.preprocessing.image.array_to_img(superimposed_img)\n",
"\n",
"save_path = \"elephant_cam.jpg\"\n",
"superimposed_img.save(save_path)"
]
},
{
"cell_type": "markdown",
"metadata": {
"colab_type": "text"
},
"source": [
"## Chapter summary"
]
}
],
"metadata": {
"colab": {
"collapsed_sections": [],
"name": "chapter09_part03_interpreting-what-convnets-learn.i",
"private_outputs": false,
"provenance": [],
"toc_visible": true
},
"kernelspec": {
"display_name": "Python 3",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.7.0"
}
},
"nbformat": 4,
"nbformat_minor": 0
}