File size: 18,771 Bytes
edcf5ee |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 |
"""
Attention Visulization Script ver: Oct 23rd 18:00
use rgb format input
"""
import torch
import torch.nn as nn
import numpy as np
import matplotlib.pyplot as plt
import os
from PIL import Image
from torchvision.transforms import ToPILImage
def softmax(x):
"""Compute the softmax in a numerically stable way."""
sof = nn.Softmax()
return sof(x)
def imshow(inp, title=None): # Imshow for Tensor
"""Imshow for Tensor."""
inp = inp.numpy().transpose((1, 2, 0))
'''
# if required: Alter the transform
# because transforms.Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225])
mean = np.array([0.485, 0.456, 0.406])
std = np.array([0.229, 0.224, 0.225])
inp = std * inp + mean
inp = np.clip(inp, 0, 1)
'''
plt.imshow(inp)
if title is not None:
plt.title(title)
plt.pause(0.001) # pause a bit so that plots are updated
def Draw_tri_fig(Ori_img, Puz_img, Rec_img, picpath):
plt.figure()
ax = plt.subplot(1, 3, 1)
ax.axis('off')
ax.set_title('Original')
plt.imshow(Ori_img)
ax = plt.subplot(1, 3, 2)
ax.axis('off')
ax.set_title('Puzzle')
plt.imshow(Puz_img)
ax = plt.subplot(1, 3, 3)
ax.axis('off')
ax.set_title('Restored')
plt.imshow(Rec_img)
plt.savefig(picpath, dpi=400)
plt.show()
plt.cla()
plt.close("all")
# Grad CAM part:Visualize of CNN+Transformer attention area
def cls_token_s12_transform(tensor, height=12, width=12): # based on pytorch_grad_cam
result = tensor[:, 1:, :].reshape(tensor.size(0), height, width, tensor.size(2))
# Bring the channels to the first dimension,
# like in CNNs.
result = result.transpose(2, 3).transpose(1, 2)
return result
def cls_token_s14_transform(tensor, height=14, width=14): # based on pytorch_grad_cam
result = tensor[:, 1:, :].reshape(tensor.size(0), height, width, tensor.size(2))
# Bring the channels to the first dimension,
# like in CNNs.
result = result.transpose(2, 3).transpose(1, 2)
return result
def cls_token_s16_transform(tensor, height=16, width=16): # based on pytorch_grad_cam
result = tensor[:, 1:, :].reshape(tensor.size(0), height, width, tensor.size(2))
# Bring the channels to the first dimension,
# like in CNNs.
result = result.transpose(2, 3).transpose(1, 2)
return result
def cls_token_s24_transform(tensor, height=24, width=24): # based on pytorch_grad_cam
result = tensor[:, 1:, :].reshape(tensor.size(0), height, width, tensor.size(2))
# Bring the channels to the first dimension,
# like in CNNs.
result = result.transpose(2, 3).transpose(1, 2)
return result
def no_cls_token_s12_transform(tensor, height=12, width=12): # based on pytorch_grad_cam
result = tensor.reshape(tensor.size(0), height, width, tensor.size(2))
# Bring the channels to the first dimension,
# like in CNNs.
result = result.transpose(2, 3).transpose(1, 2)
return result
def swinT_transform_224(tensor, height=7, width=7): # 224 7
result = tensor.reshape(tensor.size(0), height, width, tensor.size(2))
# Bring the channels to the first dimension,
# like in CNNs.
result = result.transpose(2, 3).transpose(1, 2)
return result
def swinT_transform_384(tensor, height=12, width=12): # 384 12
result = tensor.reshape(tensor.size(0), height, width, tensor.size(2))
# Bring the channels to the first dimension,
# like in CNNs.
result = result.transpose(2, 3).transpose(1, 2)
return result
def choose_cam_by_model(model, model_idx, edge_size, use_cuda=True, model_type='CLS'):
"""
:param model: model object
:param model_idx: model idx for the getting pre-setted layer and size
:param edge_size: image size for the getting pre-setted layer and size
:param use_cuda: use cuda to speed up imaging
:param model_type: default 'CLS' for model, 'MIL' for model backbone
"""
from pytorch_grad_cam import GradCAM
# reshape_transform todo conformer 224!!
# check class: target_category = None
# If None, returns the map for the highest scoring category.
# Otherwise, targets the requested category.
if model_idx[0:3] == 'ViT' or model_idx[0:4] == 'deit':
# We should chose any layer before the final attention block,
# check: https://github.com/jacobgil/pytorch-grad-cam/blob/master/tutorials/vision_transformers.md
if model_type == 'CLS':
target_layers = [model.blocks[-1].norm1]
else: # MIL-SI
target_layers = [model.backbone.blocks[-1].norm1]
if model_idx[0:5] == 'ViT_h':
if edge_size == 224:
grad_cam = GradCAM(model, target_layers=target_layers, use_cuda=use_cuda,
reshape_transform=cls_token_s16_transform)
else:
print('ERRO in ViT_huge edge size')
return -1
else:
if edge_size == 384:
grad_cam = GradCAM(model, target_layers=target_layers, use_cuda=use_cuda,
reshape_transform=cls_token_s24_transform)
elif edge_size == 224:
grad_cam = GradCAM(model, target_layers=target_layers, use_cuda=use_cuda,
reshape_transform=cls_token_s14_transform)
else:
print('ERRO in ViT/DeiT edge size')
return -1
elif model_idx[0:3] == 'vgg':
if model_type == 'CLS':
target_layers = [model.features[-1]]
else:
target_layers = [model.backbone.features[-1]]
grad_cam = GradCAM(model, target_layers=target_layers, use_cuda=use_cuda, reshape_transform=None)
elif model_idx[0:6] == 'swin_b':
if model_type == 'CLS':
target_layers = [model.layers[-1].blocks[-1].norm1]
else:
target_layers = [model.backbone.layers[-1].blocks[-1].norm1]
if edge_size == 384:
grad_cam = GradCAM(model, target_layers=target_layers, use_cuda=use_cuda,
reshape_transform=swinT_transform_384)
elif edge_size == 224:
grad_cam = GradCAM(model, target_layers=target_layers, use_cuda=use_cuda,
reshape_transform=swinT_transform_224)
else:
print('ERRO in Swin Transformer edge size')
return -1
elif model_idx[0:6] == 'ResNet':
if model_type == 'CLS':
target_layers = [model.layer4[-1]]
else:
target_layers = [model.backbone.layer4[-1]]
grad_cam = GradCAM(model, target_layers=target_layers, use_cuda=use_cuda, reshape_transform=None) # CNN: None
elif model_idx[0:7] == 'Hybrid1' and edge_size == 384:
target_layers = [model.blocks[-1].norm1]
grad_cam = GradCAM(model, target_layers=target_layers, use_cuda=use_cuda,
reshape_transform=cls_token_s12_transform)
elif model_idx[0:7] == 'Hybrid2' and edge_size == 384:
target_layers = [model.dec4.norm1]
if 'CLS' in model_idx.split('_') and 'No' in model_idx.split('_'):
grad_cam = GradCAM(model, target_layers=target_layers, use_cuda=use_cuda,
reshape_transform=no_cls_token_s12_transform)
else:
grad_cam = GradCAM(model, target_layers=target_layers, use_cuda=use_cuda,
reshape_transform=cls_token_s12_transform)
elif model_idx[0:7] == 'Hybrid3' and edge_size == 384:
target_layers = [model.dec3.norm1]
grad_cam = GradCAM(model, target_layers=target_layers, use_cuda=use_cuda,
reshape_transform=cls_token_s24_transform)
elif model_idx[0:9] == 'mobilenet':
if model_type == 'CLS':
target_layers = [model.blocks[-1]]
else:
target_layers = [model.backbone.blocks[-1]]
grad_cam = GradCAM(model, target_layers=target_layers, use_cuda=use_cuda, reshape_transform=None) # CNN: None
elif model_idx[0:10] == 'ResN50_ViT' and edge_size == 384:
if model_type == 'CLS':
target_layers = [model.blocks[-1].norm1]
else:
target_layers = [model.backbone.blocks[-1].norm1]
if edge_size == 384:
grad_cam = GradCAM(model, target_layers=target_layers, use_cuda=use_cuda,
reshape_transform=cls_token_s24_transform)
elif edge_size == 224:
grad_cam = GradCAM(model, target_layers=target_layers, use_cuda=use_cuda,
reshape_transform=cls_token_s14_transform)
else:
print('ERRO in ResN50_ViT edge size')
return -1
elif model_idx[0:12] == 'efficientnet':
target_layers = [model.conv_head]
grad_cam = GradCAM(model, target_layers=target_layers, use_cuda=use_cuda, reshape_transform=None) # CNN: None
else:
print('ERRO in model_idx')
return -1
return grad_cam
def check_SAA(inputs, labels, model, model_idx, edge_size, class_names, model_type='CLS', num_images=-1,
pic_name='test',
draw_path='../imaging_results', check_all=True, unknown_GT=False, writer=None):
"""
check num_images of images and visual the models's attention area
output a pic with 2 column and rows of num_images
:param inputs: inputs of data
:param labels: labels or the K+1 soft label of data
:param model: model object
:param model_idx: model idx for the getting pre-setted layer and size
:param edge_size: image size for the getting pre-setted layer and size
:param class_names: The name of classes for painting
:param model_type: default 'CLS' for model, 'MIL' for model backbone
:param num_images: how many image u want to check, should SMALLER THAN the batchsize
:param pic_name: name of the output pic
:param draw_path: path folder for output pic
:param check_all: choose the type of checking CAM : by default False to be only on the predicted type'
True to be on all types
:param unknown_GT: cam on unknown GT
:param writer: attach the pic to the tensorboard backend
:return: None
"""
from pytorch_grad_cam.utils import show_cam_on_image
# choose checking type: false to be only on the predicted type'; true to be on all types
if check_all:
checking_type = ['ori', ]
checking_type.extend([cls for cls in range(len(class_names))])
else:
checking_type = ['ori', 'tar']
# test model
was_training = model.training
model.eval()
outputs = model(inputs)
_, preds = torch.max(outputs, 1)
grad_cam = choose_cam_by_model(model, model_idx, edge_size, model_type=model_type) # choose model
if num_images == -1: # auto detect a batch
num_images = int(inputs.shape[0])
images_so_far = 0
plt.figure()
for j in range(num_images):
for type in checking_type:
images_so_far += 1
if type == 'ori':
ax = plt.subplot(num_images, len(checking_type), images_so_far)
ax.axis('off')
if unknown_GT and not len(labels) == 1: # Ground Truth of the K+1 soft label
soft_label = labels.cpu().numpy()[j] # K+1 soft label
title = 'A' + str(round(soft_label[0], 0))
for i in range(1, len(soft_label)):
title += class_names[i - 1][0] # use the first character only
title += str(round(soft_label[i], 0)) # use int (float 0)
title += ' '
ax.set_title(title)
else:
ax.set_title('Ground Truth:{}'.format(class_names[int(labels[j])]))
imshow(inputs.cpu().data[j])
plt.pause(0.001) # pause a bit so that plots are updated
else:
ax = plt.subplot(num_images, len(checking_type), images_so_far)
ax.axis('off')
if type == 'tar': # target categories
ax.set_title('Predict: {}'.format(class_names[preds[j]]))
# focus on the specific target class to create grayscale_cam
# grayscale_cam is generate on batch
grayscale_cam = grad_cam(inputs, target_category=None, eigen_smooth=False, aug_smooth=False)
else:
# pseudo confidence by softmax
ax.set_title('{:.1%} {}'.format(softmax(outputs[j])[int(type)], class_names[int(type)]))
# focus on the specific target class to create grayscale_cam
# grayscale_cam is generate on batch
grayscale_cam = grad_cam(inputs, target_category=int(type), eigen_smooth=False, aug_smooth=False)
# get a cv2 encoding image from dataloder by inputs[j].cpu().numpy().transpose((1, 2, 0))
cam_img = show_cam_on_image(inputs[j].cpu().numpy().transpose((1, 2, 0)), grayscale_cam[j],
use_rgb=True) # Fixme: use rgb format input (already fixed)
plt.imshow(cam_img)
plt.pause(0.001) # pause a bit so that plots are updated
if images_so_far == num_images * len(checking_type): # complete when the pics is enough
picpath = os.path.join(draw_path, pic_name + '.jpg')
if not os.path.exists(draw_path):
os.makedirs(draw_path)
plt.savefig(picpath, dpi=1000)
plt.show()
model.train(mode=was_training)
if writer is not None: # attach the pic to the tensorboard backend if avilable
image_PIL = Image.open(picpath)
img = np.array(image_PIL)
writer.add_image(pic_name, img, 1, dataformats='HWC')
plt.cla()
plt.close("all")
return
model.train(mode=was_training)
def visualize_check(inputs, labels, model, class_names, num_images=-1, pic_name='test',
draw_path='/home/ZTY/imaging_results', writer=None): # visual check
"""
check num_images of images and visual them
output a pic with 3 column and rows of num_images//3
:param inputs: inputs of data
:param labels: labels of data
:param model: model object
:param class_names: The name of classes for painting
:param num_images: how many image u want to check, should SMALLER THAN the batchsize
:param pic_name: name of the output pic
:param draw_path: path folder for output pic
:param writer: attach the pic to the tensorboard backend
:return: None
"""
was_training = model.training
model.eval()
images_so_far = 0
plt.figure()
with torch.no_grad():
outputs = model(inputs)
_, preds = torch.max(outputs, 1)
if num_images == -1: # auto detect a batch
num_images = int(inputs.shape[0])
if num_images % 5 == 0:
line_imgs_num = 5
elif num_images % 4 == 0:
line_imgs_num = 4
elif num_images % 3 == 0:
line_imgs_num = 3
elif num_images % 2 == 0:
line_imgs_num = 2
else:
line_imgs_num = int(num_images)
rows_imgs_num = int(num_images // line_imgs_num)
num_images = line_imgs_num * rows_imgs_num
for j in range(num_images): # each batch input idx: j
images_so_far += 1
ax = plt.subplot(rows_imgs_num, line_imgs_num, images_so_far)
ax.axis('off')
ax.set_title('Pred: {} True: {}'.format(class_names[preds[j]], class_names[int(labels[j])]))
imshow(inputs.cpu().data[j])
if images_so_far == num_images:
picpath = os.path.join(draw_path, pic_name + '.jpg')
if not os.path.exists(draw_path):
os.makedirs(draw_path)
'''
myfig = plt.gcf() # get current image
myfig.savefig(picpath, dpi=1000)
'''
plt.savefig(picpath, dpi=1000)
plt.show()
model.train(mode=was_training)
if writer is not None: # attach the pic to the tensorboard backend if avilable
image_PIL = Image.open(picpath)
img = np.array(image_PIL)
writer.add_image(pic_name, img, 1, dataformats='HWC')
plt.cla()
plt.close("all")
return
model.train(mode=was_training)
def unpatchify(pred, patch_size=16):
"""
Decoding embeded patch tokens
input:
x: (B, num_patches, patch_size**2 *3) AKA [B, num_patches, flatten_dim]
patch_size:
output:
imgs: (B, 3, H, W)
"""
# squre root of num_patches (without CLS token is required)
h = w = int(pred.shape[1] ** .5)
# assert num_patches is with out CLS token
assert h * w == pred.shape[1]
# ReArrange dimensions [B, num_patches, flatten_dim] -> [B, h_p, w_p, patch_size, patch_size, C]
pred = pred.reshape(shape=(pred.shape[0], h, w, patch_size, patch_size, 3))
# ReArrange dimensions [B, h_p, w_p, patch_size, patch_size, C] -> [B, C, h_p, patch_size, w_p, patch_size]
pred = torch.einsum('nhwpqc->nchpwq', pred)
# use reshape to compose patch [B, C, h_p, patch_size, w_p, patch_size] -> [B, C, H, W]
imgs = pred.reshape(shape=(pred.shape[0], 3, h * patch_size, h * patch_size))
return imgs
def patchify(imgs, patch_size=16):
"""
Break image to patch tokens
input:
imgs: (B, 3, H, W)
output:
x: (B, num_patches, patch_size**2 *3) AKA [B, num_patches, flatten_dim]
"""
# assert H == W and image shape is dividedable by patch
assert imgs.shape[2] == imgs.shape[3] and imgs.shape[2] % patch_size == 0
# patch num in rol or column
h = w = imgs.shape[2] // patch_size
# use reshape to split patch [B, C, H, W] -> [B, C, h_p, patch_size, w_p, patch_size]
imgs = imgs.reshape(shape=(imgs.shape[0], 3, h, patch_size, w, patch_size))
# ReArrange dimensions [B, C, h_p, patch_size, w_p, patch_size] -> [B, h_p, w_p, patch_size, patch_size, C]
imgs = torch.einsum('nchpwq->nhwpqc', imgs)
# ReArrange dimensions [B, h_p, w_p, patch_size, patch_size, C] -> [B, num_patches, flatten_dim]
imgs = imgs.reshape(shape=(imgs.shape[0], h * w, patch_size ** 2 * 3))
return imgs
def anti_tensor_norm(batch_tensor):
pass # TODO 总之想一下
|