File size: 9,118 Bytes
a6b26e3 254b186 a6b26e3 254b186 a6b26e3 254b186 a6b26e3 254b186 a6b26e3 254b186 a6b26e3 254b186 a6b26e3 254b186 a6b26e3 254b186 a6b26e3 254b186 a6b26e3 254b186 8aa2d09 a6b26e3 254b186 a6b26e3 254b186 a6b26e3 254b186 a6b26e3 254b186 a6b26e3 254b186 a6b26e3 254b186 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 |
import torch
import numpy as np
# from PIL import Image
# import matplotlib.pyplot as plt
import cv2
import regex as re
from .image_utils import show_cam_on_image, show_overlapped_cam
def vit_block_vis(
image,
target_features,
img_encoder,
block,
device,
grad=False,
neg_saliency=False,
img_dim=224,
):
img_encoder.eval()
image_features = img_encoder(image)
image_features_norm = image_features.norm(dim=-1, keepdim=True)
image_features_new = image_features / image_features_norm
target_features_norm = target_features.norm(dim=-1, keepdim=True)
target_features_new = target_features / target_features_norm
similarity = image_features_new[0].dot(target_features_new[0])
image = (image - image.min()) / (image.max() - image.min())
img_encoder.zero_grad()
similarity.backward(retain_graph=True)
image_attn_blocks = list(
dict(img_encoder.transformer.resblocks.named_children()).values()
)
if grad:
cam = image_attn_blocks[block].attn_grad.detach()
else:
cam = image_attn_blocks[block].attn_probs.detach()
cam = cam.mean(dim=0)
image_relevance = cam[0, 1:]
resize_dim = int(np.sqrt(list(image_relevance.shape)[0]))
# image_relevance = image_relevance.reshape(1, 1, 7, 7)
image_relevance = image_relevance.reshape(1, 1, resize_dim, resize_dim)
image_relevance = torch.nn.functional.interpolate(
image_relevance, size=img_dim, mode="bilinear"
)
image_relevance = image_relevance.reshape(img_dim, img_dim)
image_relevance = (image_relevance - image_relevance.min()) / (
image_relevance.max() - image_relevance.min()
)
cam = image_relevance * image
cam = cam / torch.max(cam)
# TODO: maybe we can ignore this...
####
masked_image_features = img_encoder(cam)
masked_image_features_norm = masked_image_features.norm(dim=-1, keepdim=True)
masked_image_features_new = masked_image_features / masked_image_features_norm
new_score = masked_image_features_new[0].dot(target_features_new[0])
####
cam = cam[0].permute(1, 2, 0).data.cpu().numpy()
cam = np.float32(cam)
# plt.imshow(cam)
return new_score
def vit_relevance(
image,
target_features,
img_encoder,
device,
method="last grad",
neg_saliency=False,
img_dim=224,
):
img_encoder.eval()
image_features = img_encoder(image)
image_features_norm = image_features.norm(dim=-1, keepdim=True)
image_features_new = image_features / image_features_norm
target_features_norm = target_features.norm(dim=-1, keepdim=True)
target_features_new = target_features / target_features_norm
similarity = image_features_new[0].dot(target_features_new[0])
if neg_saliency:
objective = 1 - similarity
else:
objective = similarity
img_encoder.zero_grad()
objective.backward(retain_graph=True)
image_attn_blocks = list(
dict(img_encoder.transformer.resblocks.named_children()).values()
)
num_tokens = image_attn_blocks[0].attn_probs.shape[-1]
last_attn = image_attn_blocks[-1].attn_probs.detach()
last_attn = last_attn.reshape(-1, last_attn.shape[-1], last_attn.shape[-1])
last_grad = image_attn_blocks[-1].attn_grad.detach()
last_grad = last_grad.reshape(-1, last_grad.shape[-1], last_grad.shape[-1])
if method == "gradcam":
cam = last_grad * last_attn
cam = cam.clamp(min=0).mean(dim=0)
image_relevance = cam[0, 1:]
else:
R = torch.eye(
num_tokens, num_tokens, dtype=image_attn_blocks[0].attn_probs.dtype
).to(device)
for blk in image_attn_blocks:
cam = blk.attn_probs.detach()
cam = cam.reshape(-1, cam.shape[-1], cam.shape[-1])
if method == "last grad":
grad = last_grad
elif method == "all grads":
grad = blk.attn_grad.detach()
else:
print(
"The available visualization methods are: 'gradcam', 'last grad', 'all grads'."
)
return
cam = grad * cam
cam = cam.clamp(min=0).mean(dim=0)
R += torch.matmul(cam, R)
image_relevance = R[0, 1:]
resize_dim = int(np.sqrt(list(image_relevance.shape)[0]))
# image_relevance = image_relevance.reshape(1, 1, 7, 7)
image_relevance = image_relevance.reshape(1, 1, resize_dim, resize_dim)
image_relevance = torch.nn.functional.interpolate(
image_relevance, size=img_dim, mode="bilinear"
)
image_relevance = image_relevance.reshape(img_dim, img_dim).data.cpu().numpy()
image_relevance = (image_relevance - image_relevance.min()) / (
image_relevance.max() - image_relevance.min()
)
image = image[0].permute(1, 2, 0).data.cpu().numpy()
image = (image - image.min()) / (image.max() - image.min())
return image_relevance, image, image_features, similarity
def interpret_vit(
image,
target_features,
img_encoder,
device,
method="last grad",
neg_saliency=False,
img_dim=224,
):
image_relevance, image, image_features, similarity = vit_relevance(
image,
target_features,
img_encoder,
device,
method=method,
neg_saliency=neg_saliency,
img_dim=img_dim,
)
vis = show_cam_on_image(image, image_relevance, neg_saliency=neg_saliency)
vis = np.uint8(255 * vis)
vis = cv2.cvtColor(np.array(vis), cv2.COLOR_RGB2BGR)
return vis, image_features, similarity
# plt.imshow(vis)
def interpret_vit_overlapped(
image, target_features, img_encoder, device, method="last grad", img_dim=224
):
pos_image_relevance, _, image_features, similarity = vit_relevance(
image,
target_features,
img_encoder,
device,
method=method,
neg_saliency=False,
img_dim=img_dim,
)
neg_image_relevance, image, _, _ = vit_relevance(
image,
target_features,
img_encoder,
device,
method=method,
neg_saliency=True,
img_dim=img_dim,
)
vis = show_overlapped_cam(image, neg_image_relevance, pos_image_relevance)
vis = np.uint8(255 * vis)
vis = cv2.cvtColor(np.array(vis), cv2.COLOR_RGB2BGR)
return vis, image_features, similarity
# plt.imshow(vis)
def vit_perword_relevance(
image_features,
text,
clip_model,
clip_tokenizer,
device,
masked_word="",
use_last_grad=True,
img_dim=224,
):
clip_model.eval()
main_text = clip_tokenizer(text).to(device)
# remove the word for which you want to visualize the saliency
masked_text = re.sub(masked_word, "", text)
masked_text = clip_tokenizer(masked_text).to(device)
main_text_features = clip_model.encode_text(main_text)
masked_text_features = clip_model.encode_text(masked_text)
# image_features = clip_model.encode_image(image)
image_features_norm = image_features.norm(dim=-1, keepdim=True)
image_features_new = image_features / image_features_norm
main_text_features_norm = main_text_features.norm(dim=-1, keepdim=True)
main_text_features_new = main_text_features / main_text_features_norm
masked_text_features_norm = masked_text_features.norm(dim=-1, keepdim=True)
masked_text_features_new = masked_text_features / masked_text_features_norm
objective = image_features_new[0].dot(
main_text_features_new[0] - masked_text_features_new[0]
)
clip_model.visual.zero_grad()
objective.backward(retain_graph=True)
image_attn_blocks = list(
dict(clip_model.visual.transformer.resblocks.named_children()).values()
)
num_tokens = image_attn_blocks[0].attn_probs.shape[-1]
R = torch.eye(
num_tokens, num_tokens, dtype=image_attn_blocks[0].attn_probs.dtype
).to(device)
last_grad = image_attn_blocks[-1].attn_grad.detach()
last_grad = last_grad.reshape(-1, last_grad.shape[-1], last_grad.shape[-1])
for blk in image_attn_blocks:
cam = blk.attn_probs.detach()
cam = cam.reshape(-1, cam.shape[-1], cam.shape[-1])
if use_last_grad:
grad = last_grad
else:
grad = blk.attn_grad.detach()
cam = grad * cam
cam = cam.clamp(min=0).mean(dim=0)
R += torch.matmul(cam, R)
image_relevance = R[0, 1:]
resize_dim = int(np.sqrt(list(image_relevance.shape)[0]))
image_relevance = image_relevance.reshape(1, 1, resize_dim, resize_dim)
image_relevance = torch.nn.functional.interpolate(
image_relevance, size=img_dim, mode="bilinear"
)
image_relevance = image_relevance.reshape(img_dim, img_dim).data.cpu().numpy()
image_relevance = (image_relevance - image_relevance.min()) / (
image_relevance.max() - image_relevance.min()
)
# image = image[0].permute(1, 2, 0).data.cpu().numpy()
# image = (image - image.min()) / (image.max() - image.min())
# return image_relevance, image
return image_relevance
|