![\"Open](\"https://colab.research.google.com/assets/colab-badge.svg\"){kind=icon}
"
@@ -38,6 +21,7 @@
},
{
"cell_type": "code",
+ "execution_count": 1,
"metadata": {
"colab": {
"base_uri": "https://localhost:8080/"
@@ -45,21 +29,10 @@
"id": "3ogYpvQAAH4s",
"outputId": "9fff6b83-2551-4644-d59b-bc627535ac97"
},
- "source": [
- "!git clone https://github.com/hila-chefer/Transformer-MM-Explainability\n",
- "\n",
- "import os\n",
- "os.chdir(f'./Transformer-MM-Explainability')\n",
- "\n",
- "!pip install einops\n",
- "!pip install ftfy\n",
- "!pip install captum"
- ],
- "execution_count": 1,
"outputs": [
{
- "output_type": "stream",
"name": "stdout",
+ "output_type": "stream",
"text": [
"Cloning into 'Transformer-MM-Explainability'...\n",
"remote: Enumerating objects: 1463, done.\u001b[K\n",
@@ -97,6 +70,16 @@
"Successfully installed captum-0.5.0\n"
]
}
+ ],
+ "source": [
+ "!git clone https://github.com/hila-chefer/Transformer-MM-Explainability\n",
+ "\n",
+ "import os\n",
+ "os.chdir(f'./Transformer-MM-Explainability')\n",
+ "\n",
+ "!pip install einops\n",
+ "!pip install ftfy\n",
+ "!pip install captum"
]
},
{
@@ -110,9 +93,11 @@
},
{
"cell_type": "code",
+ "execution_count": 2,
"metadata": {
"id": "or8UETbZAYY3"
},
+ "outputs": [],
"source": [
"import torch\n",
"import CLIP.clip as clip\n",
@@ -121,57 +106,47 @@
"import cv2\n",
"import matplotlib.pyplot as plt\n",
"from captum.attr import visualization"
- ],
- "execution_count": 2,
- "outputs": []
+ ]
},
{
"cell_type": "code",
+ "execution_count": 3,
"metadata": {
"cellView": "form",
"id": "DhG24G9cTpHT"
},
+ "outputs": [],
"source": [
"#@title Control context expansion (number of attention layers to consider)\n",
"#@title Number of layers for image Transformer\n",
- "start_layer = -1#@param {type:\"number\"}\n",
+ "START_LAYER = -1#@param {type:\"number\"}\n",
"\n",
"#@title Number of layers for text Transformer\n",
- "start_layer_text = -1#@param {type:\"number\"}"
- ],
- "execution_count": 3,
- "outputs": []
+ "START_LAYER_TEXT = -1#@param {type:\"number\"}"
+ ]
},
{
"cell_type": "code",
+ "execution_count": 4,
"metadata": {
"id": "fWKGyu2YAeSV"
},
+ "outputs": [],
"source": [
- "def interpret(image, texts, model, device, start_layer=start_layer, start_layer_text=start_layer_text):\n",
- " batch_size = texts.shape[0]\n",
- " images = image.repeat(batch_size, 1, 1, 1)\n",
- " logits_per_image, logits_per_text = model(images, texts)\n",
- " probs = logits_per_image.softmax(dim=-1).detach().cpu().numpy()\n",
- " index = [i for i in range(batch_size)]\n",
- " one_hot = np.zeros((logits_per_image.shape[0], logits_per_image.shape[1]), dtype=np.float32)\n",
- " one_hot[torch.arange(logits_per_image.shape[0]), index] = 1\n",
- " one_hot = torch.from_numpy(one_hot).requires_grad_(True)\n",
- " one_hot = torch.sum(one_hot.cuda() * logits_per_image)\n",
- " model.zero_grad()\n",
- "\n",
- " image_attn_blocks = list(dict(model.visual.transformer.resblocks.named_children()).values())\n",
- "\n",
- " if start_layer == -1: \n",
- " # calculate index of last layer \n",
- " start_layer = len(image_attn_blocks) - 1\n",
- " \n",
- " num_tokens = image_attn_blocks[0].attn_probs.shape[-1]\n",
- " R = torch.eye(num_tokens, num_tokens, dtype=image_attn_blocks[0].attn_probs.dtype).to(device)\n",
+ "def r(batch_size, device, attn_blocks, one_hot, start_layer):\n",
+ " \"\"\"\n",
+ " calculate relevance using attention blocks and start layer.\n",
+ " \"\"\"\n",
+ " assert 'attn_probs' in dir(attn_blocks[0]), \"It will not be possible to do a grad cam on this model, as Residual Attention Blocks do not have an attn_probs attribute. Please check that you are using this version of CLIP, or an equivalent: https://github.com/hila-chefer/Transformer-MM-Explainability/tree/main/CLIP\"\n",
+ " if start_layer == -1:\n",
+ " # calculate index of last layer\n",
+ " start_layer = len(attn_blocks) - 1\n",
+ " num_tokens = attn_blocks[0].attn_probs.shape[-1]\n",
+ " R = torch.eye(num_tokens, num_tokens, dtype=attn_blocks[0].attn_probs.dtype).to(device)\n",
" R = R.unsqueeze(0).expand(batch_size, num_tokens, num_tokens)\n",
- " for i, blk in enumerate(image_attn_blocks):\n",
+ " for i, blk in enumerate(attn_blocks):\n",
" if i < start_layer:\n",
- " continue\n",
+ " continue\n",
" grad = torch.autograd.grad(one_hot, [blk.attn_probs], retain_graph=True)[0].detach()\n",
" cam = blk.attn_probs.detach()\n",
" cam = cam.reshape(-1, cam.shape[-1], cam.shape[-1])\n",
@@ -180,39 +155,58 @@
" cam = cam.reshape(batch_size, -1, cam.shape[-1], cam.shape[-1])\n",
" cam = cam.clamp(min=0).mean(dim=1)\n",
" R = R + torch.bmm(cam, R)\n",
- " image_relevance = R[:, 0, 1:]\n",
+ " return R\n",
+ "\n",
+ "\n",
+ "def interpret(input_tensor, texts, model, device, predict, name, start_layer=START_LAYER, start_layer_text=START_LAYER_TEXT):\n",
+ " \"\"\"\n",
+ "\n",
+ " :param input_tensor: image to pass to model\n",
+ " :param texts: The texts to which the tensor will be compared\n",
+ " :param model: CLIP-like model\n",
+ " :param device: cuda or cpu.\n",
+ " :param start_layer: a number choosing where to start backprop for images\n",
+ " :param start_layer_text: a number choosing where to start backprop for text\n",
+ " :return: text_relevance for each proposed text, image_relevance for each proposed text\n",
+ " \"\"\"\n",
+ " batch_size = texts.shape[0]\n",
+ " images = input_tensor.repeat(batch_size, 1, 1, 1)\n",
+ " _, logits_per_image = predict(model, images, name, texts)\n",
+ "\n",
+ " index = [i for i in range(batch_size)]\n",
+ " one_hot = np.zeros((logits_per_image.shape[0], logits_per_image.shape[1]), dtype=np.float32)\n",
+ " one_hot[torch.arange(logits_per_image.shape[0]), index] = 1\n",
+ " one_hot = torch.from_numpy(one_hot).requires_grad_(True)\n",
+ " one_hot = torch.sum(one_hot.cuda() * logits_per_image)\n",
+ " model.zero_grad()\n",
"\n",
- " \n",
+ " image_attn_blocks = list(dict(model.visual.transformer.resblocks.named_children()).values())\n",
" text_attn_blocks = list(dict(model.transformer.resblocks.named_children()).values())\n",
"\n",
- " if start_layer_text == -1: \n",
- " # calculate index of last layer \n",
- " start_layer_text = len(text_attn_blocks) - 1\n",
+ " image_relevance = r(batch_size, device, image_attn_blocks, one_hot, start_layer)[:, 0, 1:]\n",
+ " text_relevance = r(batch_size, device, text_attn_blocks, one_hot, start_layer_text)\n",
"\n",
- " num_tokens = text_attn_blocks[0].attn_probs.shape[-1]\n",
- " R_text = torch.eye(num_tokens, num_tokens, dtype=text_attn_blocks[0].attn_probs.dtype).to(device)\n",
- " R_text = R_text.unsqueeze(0).expand(batch_size, num_tokens, num_tokens)\n",
- " for i, blk in enumerate(text_attn_blocks):\n",
- " if i < start_layer_text:\n",
- " continue\n",
- " grad = torch.autograd.grad(one_hot, [blk.attn_probs], retain_graph=True)[0].detach()\n",
- " cam = blk.attn_probs.detach()\n",
- " cam = cam.reshape(-1, cam.shape[-1], cam.shape[-1])\n",
- " grad = grad.reshape(-1, grad.shape[-1], grad.shape[-1])\n",
- " cam = grad * cam\n",
- " cam = cam.reshape(batch_size, -1, cam.shape[-1], cam.shape[-1])\n",
- " cam = cam.clamp(min=0).mean(dim=1)\n",
- " R_text = R_text + torch.bmm(cam, R_text)\n",
- " text_relevance = R_text\n",
- " \n",
" return text_relevance, image_relevance"
- ],
- "execution_count": 4,
- "outputs": []
+ ]
},
{
"cell_type": "code",
+ "execution_count": 26,
+ "metadata": {
+ "id": "9MJ-Ech7dj6C"
+ },
+ "outputs": [],
"source": [
+ "def mask_from_relevance(image_relevance):\n",
+ " # create heatmap from mask on input_tensor\n",
+ " dim = int(image_relevance.numel() ** 0.5)\n",
+ " image_relevance = image_relevance.reshape(1, 1, dim, dim)\n",
+ " image_relevance = torch.nn.functional.interpolate(image_relevance, size=224, mode='bilinear')\n",
+ " image_relevance = image_relevance.reshape(224, 224).cuda().data.cpu().numpy()\n",
+ " image_relevance = (image_relevance - image_relevance.min()) / (image_relevance.max() - image_relevance.min())\n",
+ " return image_relevance\n",
+ "\n",
+ "\n",
"def show_image_relevance(image_relevance, image, orig_image):\n",
" # create heatmap from mask on image\n",
" def show_cam_on_image(img, mask):\n",
@@ -223,30 +217,26 @@
" return cam\n",
"\n",
" fig, axs = plt.subplots(1, 2)\n",
- " axs[0].imshow(orig_image);\n",
- " axs[0].axis('off');\n",
+ " axs[0].imshow(orig_image)\n",
+ " axs[0].axis('off')\n",
"\n",
- " dim = int(image_relevance.numel() ** 0.5)\n",
- " image_relevance = image_relevance.reshape(1, 1, dim, dim)\n",
- " image_relevance = torch.nn.functional.interpolate(image_relevance, size=224, mode='bilinear')\n",
- " image_relevance = image_relevance.reshape(224, 224).cuda().data.cpu().numpy()\n",
- " image_relevance = (image_relevance - image_relevance.min()) / (image_relevance.max() - image_relevance.min())\n",
+ " mask = mask_from_relevance(image_relevance)\n",
" image = image[0].permute(1, 2, 0).data.cpu().numpy()\n",
" image = (image - image.min()) / (image.max() - image.min())\n",
- " vis = show_cam_on_image(image, image_relevance)\n",
+ " vis = show_cam_on_image(image, mask)\n",
" vis = np.uint8(255 * vis)\n",
" vis = cv2.cvtColor(np.array(vis), cv2.COLOR_RGB2BGR)\n",
- " axs[1].imshow(vis);\n",
- " axs[1].axis('off');"
- ],
- "metadata": {
- "id": "9MJ-Ech7dj6C"
- },
- "execution_count": 26,
- "outputs": []
+ " axs[1].imshow(vis)\n",
+ " axs[1].axis('off')"
+ ]
},
{
"cell_type": "code",
+ "execution_count": 7,
+ "metadata": {
+ "id": "NsNrdWXOxub1"
+ },
+ "outputs": [],
"source": [
"from CLIP.clip.simple_tokenizer import SimpleTokenizer as _Tokenizer\n",
"_tokenizer = _Tokenizer()\n",
@@ -261,58 +251,40 @@
" text_tokens_decoded=[_tokenizer.decode([a]) for a in text_tokens]\n",
" vis_data_records = [visualization.VisualizationDataRecord(text_scores,0,0,0,0,0,text_tokens_decoded,1)]\n",
" visualization.visualize_text(vis_data_records)"
- ],
- "metadata": {
- "id": "NsNrdWXOxub1"
- },
- "execution_count": 7,
- "outputs": []
- },
- {
- "cell_type": "code",
- "metadata": {
- "id": "7YYjztv3Nn9V"
- },
- "source": [
- "clip.clip._MODELS = {\n",
- " \"ViT-B/32\": \"https://openaipublic.azureedge.net/clip/models/40d365715913c9da98579312b702a82c18be219cc2a73407c4526f58eba950af/ViT-B-32.pt\",\n",
- " \"ViT-B/16\": \"https://openaipublic.azureedge.net/clip/models/5806e77cd80f8b59890b7e101eabd078d9fb84e6937f9e85e4ecb61988df416f/ViT-B-16.pt\",\n",
- " \"ViT-L/14\": \"https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836/ViT-L-14.pt\",\n",
- "}"
- ],
- "execution_count": 8,
- "outputs": []
+ ]
},
{
"cell_type": "code",
+ "execution_count": 36,
"metadata": {
- "id": "pns9sG9eAhho",
- "outputId": "180c0886-cb3f-4554-82e2-4a51250f6692",
"colab": {
"base_uri": "https://localhost:8080/"
- }
+ },
+ "id": "pns9sG9eAhho",
+ "outputId": "180c0886-cb3f-4554-82e2-4a51250f6692"
},
- "source": [
- "device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n",
- "model, preprocess = clip.load(\"ViT-B/32\", device=device, jit=False)"
- ],
- "execution_count": 36,
"outputs": [
{
- "output_type": "stream",
"name": "stderr",
+ "output_type": "stream",
"text": [
"/usr/local/lib/python3.7/dist-packages/torchvision/transforms/transforms.py:333: UserWarning: Argument interpolation should be of type InterpolationMode instead of int. Please, use InterpolationMode enum.\n",
" \"Argument interpolation should be of type InterpolationMode instead of int. \"\n"
]
}
+ ],
+ "source": [
+ "device = \"cuda\" if torch.cuda.is_available() else \"cpu\"\n",
+ "model, preprocess = clip.load(\"ViT-B/32\", device=device, jit=False)"
]
},
{
"cell_type": "code",
+ "execution_count": 10,
"metadata": {
"id": "4hIrp94ktMyc"
},
+ "outputs": [],
"source": [
"class color:\n",
" PURPLE = '\\033[95m'\n",
@@ -325,83 +297,57 @@
" BOLD = '\\033[1m'\n",
" UNDERLINE = '\\033[4m'\n",
" END = '\\033[0m'"
- ],
- "execution_count": 10,
- "outputs": []
+ ]
},
{
"cell_type": "code",
+ "execution_count": 37,
"metadata": {
- "id": "WoeuTHtyBq7Q",
"colab": {
"base_uri": "https://localhost:8080/",
"height": 300
},
+ "id": "WoeuTHtyBq7Q",
"outputId": "9d35d3ed-a997-4f23-e5d8-d4110efd4d7d"
},
- "source": [
- "img_path = \"CLIP/glasses.png\"\n",
- "img = preprocess(Image.open(img_path)).unsqueeze(0).to(device)\n",
- "texts = [\"a man with eyeglasses\"]\n",
- "text = clip.tokenize(texts).to(device)\n",
- "\n",
- "R_text, R_image = interpret(model=model, image=img, texts=text, device=device)\n",
- "batch_size = text.shape[0]\n",
- "for i in range(batch_size):\n",
- " show_heatmap_on_text(texts[i], text[i], R_text[i])\n",
- " show_image_relevance(R_image[i], img, orig_image=Image.open(img_path))\n",
- " plt.show()"
- ],
- "execution_count": 37,
"outputs": [
{
- "output_type": "stream",
"name": "stdout",
+ "output_type": "stream",
"text": [
"tensor([0.0274, 0.0645, 0.0048, 0.1697, 0.7339], device='cuda:0',\n",
" dtype=torch.float16)\n"
]
},
{
- "output_type": "display_data",
"data": {
- "text/plain": [
- "| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| a man with eye glasses | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| a man with lipstick | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| a rocket standing on a launch pad | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| a zebra | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| an elephant | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| a lake | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| a zebra | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| an elephant | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| water | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| an elephant | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| a zebra | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| an elephant | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| a zebra | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| a dog | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| a cat | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| a labrador | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| a tabby cat | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| a parrot | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| a dog | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| a portrait of an astronaut | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| a rocket | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| an astronaut suit | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| an image of a dog | ||||
| True Label | Predicted Label | Attribution Label | Attribution Score | Word Importance |
|---|---|---|---|---|
| an image of a dog | ||||