-
Notifications
You must be signed in to change notification settings - Fork 92
/
train.py
414 lines (355 loc) · 18 KB
/
train.py
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
import argparse
import torch
from torch.nn import functional as F
from torch import nn
from torchvision import transforms
from torch.utils.data import Dataset, DataLoader
import cv2
import numpy as np
import os
import glob
import shutil
from PIL import Image
from sklearn.metrics import roc_auc_score
from torch import nn
import pytorch_lightning as pl
from sklearn.metrics import confusion_matrix
import pickle
from sampling_methods.kcenter_greedy import kCenterGreedy
from sklearn.random_projection import SparseRandomProjection
from sklearn.neighbors import NearestNeighbors
from scipy.ndimage import gaussian_filter
def copy_files(src, dst, ignores=[]):
src_files = os.listdir(src)
for file_name in src_files:
ignore_check = [True for i in ignores if i in file_name]
if ignore_check:
continue
full_file_name = os.path.join(src, file_name)
if os.path.isfile(full_file_name):
shutil.copy(full_file_name, os.path.join(dst,file_name))
if os.path.isdir(full_file_name):
os.makedirs(os.path.join(dst, file_name), exist_ok=True)
copy_files(full_file_name, os.path.join(dst, file_name), ignores)
def prep_dirs(root):
# make embeddings dir
# embeddings_path = os.path.join(root, 'embeddings')
embeddings_path = os.path.join('./', 'embeddings', args.category)
os.makedirs(embeddings_path, exist_ok=True)
# make sample dir
sample_path = os.path.join(root, 'sample')
os.makedirs(sample_path, exist_ok=True)
# make source code record dir & copy
source_code_save_path = os.path.join(root, 'src')
os.makedirs(source_code_save_path, exist_ok=True)
copy_files('./', source_code_save_path, ['.git','.vscode','__pycache__','logs','README','samples','LICENSE']) # copy source code
return embeddings_path, sample_path, source_code_save_path
def auto_select_weights_file(weights_file_version):
print()
version_list = glob.glob(os.path.join(args.project_root_path, args.category) + '/lightning_logs/version_*')
version_list.sort(reverse=True, key=lambda x: os.path.getmtime(x))
if weights_file_version != None:
version_list = [os.path.join(args.project_root_path, args.category) + '/lightning_logs/' + weights_file_version] + version_list
for i in range(len(version_list)):
# if os.path.exists(os.path.join(version_list[i],'checkpoints')):
weights_file_path = glob.glob(os.path.join(version_list[i],'checkpoints')+'/*')
if len(weights_file_path) == 0:
if weights_file_version != None and i == 0:
print(f'Checkpoint of {weights_file_version} not found')
continue
else:
weights_file_path = weights_file_path[0]
if weights_file_path.split('.')[-1] != 'ckpt':
continue
print('Checkpoint found : ', weights_file_path)
print()
return weights_file_path
print('Checkpoint not found')
print()
return None
def embedding_concat(x, y):
# from https://github.com/xiahaifeng1995/PaDiM-Anomaly-Detection-Localization-master
B, C1, H1, W1 = x.size()
_, C2, H2, W2 = y.size()
s = int(H1 / H2)
x = F.unfold(x, kernel_size=s, dilation=1, stride=s)
x = x.view(B, C1, -1, H2, W2)
z = torch.zeros(B, C1 + C2, x.size(2), H2, W2)
for i in range(x.size(2)):
z[:, :, i, :, :] = torch.cat((x[:, :, i, :, :], y), 1)
z = z.view(B, -1, H2 * W2)
z = F.fold(z, kernel_size=s, output_size=(H1, W1), stride=s)
return z
def reshape_embedding(embedding):
embedding_list = []
for k in range(embedding.shape[0]):
for i in range(embedding.shape[2]):
for j in range(embedding.shape[3]):
embedding_list.append(embedding[k, :, i, j])
return embedding_list
#imagenet
mean_train = [0.485, 0.456, 0.406]
std_train = [0.229, 0.224, 0.225]
class MVTecDataset(Dataset):
def __init__(self, root, transform, input_size, phase):
if phase=='train':
self.img_path = os.path.join(root, 'train')
else:
self.img_path = os.path.join(root, 'test')
self.gt_path = os.path.join(root, 'ground_truth')
self.transform = transform
self.input_size = input_size
# load dataset
self.img_paths, self.gt_paths, self.labels, self.types = self.load_dataset() # self.labels => good : 0, anomaly : 1
def load_dataset(self):
img_tot_paths = []
gt_tot_paths = []
tot_labels = []
tot_types = []
defect_types = os.listdir(self.img_path)
for defect_type in defect_types:
if defect_type == 'good':
img_paths = glob.glob(os.path.join(self.img_path, defect_type) + "/*.png")
img_tot_paths.extend(img_paths)
gt_tot_paths.extend([0]*len(img_paths))
tot_labels.extend([0]*len(img_paths))
tot_types.extend(['good']*len(img_paths))
else:
img_paths = glob.glob(os.path.join(self.img_path, defect_type) + "/*.png")
gt_paths = glob.glob(os.path.join(self.gt_path, defect_type) + "/*.png")
img_paths.sort()
gt_paths.sort()
img_tot_paths.extend(img_paths)
gt_tot_paths.extend(gt_paths)
tot_labels.extend([1]*len(img_paths))
tot_types.extend([defect_type]*len(img_paths))
assert len(img_tot_paths) == len(gt_tot_paths), "Something wrong with test and ground truth pair!"
return img_tot_paths, gt_tot_paths, tot_labels, tot_types
def __len__(self):
return len(self.img_paths)
def __getitem__(self, idx):
img_path, gt, label, img_type = self.img_paths[idx], self.gt_paths[idx], self.labels[idx], self.types[idx]
img = Image.open(img_path).convert('RGB')
img = self.transform(img)
if gt == 0:
gt = torch.zeros([1, self.input_size, self.input_size])
else:
gt = Image.open(gt)
gt = gt.resize((self.input_size, self.input_size)) # intend to avoid ground-truth manipulation
gt = transforms.ToTensor()(gt)
return img, gt, label, os.path.basename(img_path[:-4]), img_type
def cvt2heatmap(gray):
heatmap = cv2.applyColorMap(np.uint8(gray), cv2.COLORMAP_JET)
return heatmap
def heatmap_on_image(heatmap, image):
if heatmap.shape != image.shape:
heatmap = cv2.resize(heatmap, (image.shape[0], image.shape[1]))
out = np.float32(heatmap)/255 + np.float32(image)/255
out = out / np.max(out)
return np.uint8(255 * out)
def min_max_norm(image):
a_min, a_max = image.min(), image.max()
return (image-a_min)/(a_max - a_min)
def init_weights(m):
if type(m) == nn.Conv2d:
torch.nn.init.xavier_uniform(m.weight)
def cal_confusion_matrix(y_true, y_pred_no_thresh, thresh, img_path_list):
pred_thresh = []
false_n = []
false_p = []
for i in range(len(y_pred_no_thresh)):
if y_pred_no_thresh[i] > thresh:
pred_thresh.append(1)
if y_true[i] == 0:
false_p.append(img_path_list[i])
else:
pred_thresh.append(0)
if y_true[i] == 1:
false_n.append(img_path_list[i])
cm = confusion_matrix(y_true, pred_thresh)
print(cm)
print('false positive')
print(false_p)
print('false negative')
print(false_n)
class STPM(pl.LightningModule):
def __init__(self, hparams):
super(STPM, self).__init__()
self.save_hyperparameters(hparams)
self.init_features()
def hook_t(module, input, output):
self.features.append(output)
self.model = torch.hub.load('pytorch/vision:v0.9.0', 'wide_resnet50_2', pretrained=True)
for param in self.model.parameters():
param.requires_grad = False
self.model.layer2[-1].register_forward_hook(hook_t)
self.model.layer3[-1].register_forward_hook(hook_t)
self.criterion = torch.nn.MSELoss(reduction='sum')
self.init_results_list()
self.data_transforms = transforms.Compose([
transforms.Resize((args.load_size, args.load_size), Image.ANTIALIAS),
transforms.ToTensor(),
transforms.CenterCrop(args.input_size),
transforms.Normalize(mean=mean_train,
std=std_train)])
self.inv_normalize = transforms.Normalize(mean=[-0.485/0.229, -0.456/0.224, -0.406/0.255], std=[1/0.229, 1/0.224, 1/0.255])
def init_results_list(self):
self.gt_list_px_lvl = []
self.pred_list_px_lvl = []
self.gt_list_img_lvl = []
self.pred_list_img_lvl = []
self.img_path_list = []
def init_features(self):
self.features = []
def forward(self, x_t):
self.init_features()
_ = self.model(x_t)
return self.features
def save_anomaly_map(self, anomaly_map, input_img, gt_img, file_name, x_type):
if anomaly_map.shape != input_img.shape:
anomaly_map = cv2.resize(anomaly_map, (input_img.shape[0], input_img.shape[1]))
anomaly_map_norm = min_max_norm(anomaly_map)
anomaly_map_norm_hm = cvt2heatmap(anomaly_map_norm*255)
# anomaly map on image
heatmap = cvt2heatmap(anomaly_map_norm*255)
hm_on_img = heatmap_on_image(heatmap, input_img)
# save images
cv2.imwrite(os.path.join(self.sample_path, f'{x_type}_{file_name}.jpg'), input_img)
cv2.imwrite(os.path.join(self.sample_path, f'{x_type}_{file_name}_amap.jpg'), anomaly_map_norm_hm)
cv2.imwrite(os.path.join(self.sample_path, f'{x_type}_{file_name}_amap_on_img.jpg'), hm_on_img)
cv2.imwrite(os.path.join(self.sample_path, f'{x_type}_{file_name}_gt.jpg'), gt_img)
def train_dataloader(self):
image_datasets = MVTecDataset(root=os.path.join(args.dataset_path,args.category), transform=self.data_transforms, input_size=args.input_size, phase='train')
train_loader = DataLoader(image_datasets, batch_size=args.batch_size, shuffle=True, num_workers=0) #, pin_memory=True)
return train_loader
def test_dataloader(self):
test_datasets = MVTecDataset(root=os.path.join(args.dataset_path,args.category), transform=self.data_transforms, input_size=args.input_size, phase='test')
test_loader = DataLoader(test_datasets, batch_size=1, shuffle=False, num_workers=0) #, pin_memory=True) # only work on batch_size=1, now.
return test_loader
def configure_optimizers(self):
return None
def on_train_start(self):
self.model.eval() # to stop running_var move (maybe not critical)
self.embedding_dir_path, self.sample_path, self.source_code_save_path = prep_dirs(self.logger.log_dir)
self.embedding_list = []
def on_test_start(self):
self.init_results_list()
self.embedding_dir_path, self.sample_path, self.source_code_save_path = prep_dirs(self.logger.log_dir)
def training_step(self, batch, batch_idx): # save locally aware patch features
x, _, _, file_name, _ = batch
features = self(x)
embeddings = []
for feature in features:
m = torch.nn.AdaptiveAvgPool2d(feature[0].shape[-2:])
embeddings.append(m(feature))
embedding = embedding_concat(embeddings[0], embeddings[1])
self.embedding_list.extend(reshape_embedding(np.array(embedding)))
def training_epoch_end(self, outputs):
total_embeddings = np.array(self.embedding_list)
# Random projection
self.randomprojector = SparseRandomProjection(n_components='auto', eps=0.9) # 'auto' => Johnson-Lindenstrauss lemma
# embedding_small = self.randomprojector.fit_transform(total_embeddings)
self.randomprojector.fit(total_embeddings)
# Coreset Subsampling
# selector = kCenterGreedy(embedding_small,0,0)
# selected_idx = selector.select_batch(model=None, already_selected=[], N=int(embedding_small.shape[0]*args.coreset_sampling_ratio))
# self.embedding_coreset = embedding_small[selected_idx]
selector = kCenterGreedy(total_embeddings,0,0)
selected_idx = selector.select_batch(model=self.randomprojector, already_selected=[], N=int(total_embeddings.shape[0]*args.coreset_sampling_ratio))
self.embedding_coreset = total_embeddings[selected_idx]
print('initial embedding size : ', total_embeddings.shape)
print('final embedding size : ', self.embedding_coreset.shape)
with open(os.path.join(self.embedding_dir_path, 'embedding.pickle'), 'wb') as f:
pickle.dump(self.embedding_coreset, f)
# with open(os.path.join(self.embedding_dir_path, 'randomprojector.pickle'), 'wb') as f:
# pickle.dump(self.randomprojector, f)
def test_step(self, batch, batch_idx): # Nearest Neighbour Search
self.embedding_coreset = pickle.load(open(os.path.join(self.embedding_dir_path, 'embedding.pickle'), 'rb'))
# self.randomprojector = pickle.load(open(os.path.join(self.embedding_dir_path, 'randomprojector.pickle'), 'rb'))
x, gt, label, file_name, x_type = batch
# extract embedding
features = self(x)
embeddings = []
for feature in features:
m = torch.nn.AdaptiveAvgPool2d(feature[0].shape[-2:])
embeddings.append(m(feature))
embedding_ = embedding_concat(embeddings[0], embeddings[1])
embedding_test = np.array(reshape_embedding(np.array(embedding_)))
# Random projection
# embedding_small_test = self.randomprojector.transform(embedding_test)
# NN
nbrs = NearestNeighbors(n_neighbors=args.n_neighbors, algorithm='ball_tree', metric='minkowski', p=2).fit(self.embedding_coreset)
# score_patches, _ = nbrs.kneighbors(embedding_small_test)
score_patches, _ = nbrs.kneighbors(embedding_test)
anomaly_map = score_patches[:,0].reshape((28,28))
N_b = score_patches[np.argmax(score_patches[:,0])]
w = (1 - (np.max(np.exp(N_b))/np.sum(np.exp(N_b))))
score = w*max(score_patches[:,0]) # Image-level score
gt_resized = transforms.Compose([transforms.Resize(args.load_size), transforms.CenterCrop(args.input_size)])(gt)
gt_np = gt_resized.cpu().numpy()[0][0].astype(int)
anomaly_map_resized = cv2.resize(anomaly_map, (args.input_size, args.input_size))
anomaly_map_resized_blur = gaussian_filter(anomaly_map_resized, sigma=4) # todo
self.gt_list_px_lvl.extend(gt_np.ravel())
self.pred_list_px_lvl.extend(anomaly_map_resized_blur.ravel())
self.gt_list_img_lvl.append(label.cpu().numpy()[0])
self.pred_list_img_lvl.append(score)
self.img_path_list.extend(file_name)
# save images
x = self.inv_normalize(x)
input_x = cv2.cvtColor(x.permute(0,2,3,1).cpu().numpy()[0]*255, cv2.COLOR_BGR2RGB)
self.save_anomaly_map(anomaly_map_resized_blur, input_x, gt_np*255, file_name[0], x_type[0])
def test_epoch_end(self, outputs):
print("Total pixel-level auc-roc score :")
pixel_auc = roc_auc_score(self.gt_list_px_lvl, self.pred_list_px_lvl)
print(pixel_auc)
print("Total image-level auc-roc score :")
img_auc = roc_auc_score(self.gt_list_img_lvl, self.pred_list_img_lvl)
print(img_auc)
print('test_epoch_end')
values = {'pixel_auc': pixel_auc, 'img_auc': img_auc}
self.log_dict(values)
# anomaly_list = []
# normal_list = []
# for i in range(len(self.gt_list_img_lvl)):
# if self.gt_list_img_lvl[i] == 1:
# anomaly_list.append(self.pred_list_img_lvl[i])
# else:
# normal_list.append(self.pred_list_img_lvl[i])
# # thresholding
# # cal_confusion_matrix(self.gt_list_img_lvl, self.pred_list_img_lvl, img_path_list = self.img_path_list, thresh = 0.00097)
# # print()
# with open(args.project_root_path + r'/results.txt', 'a') as f:
# f.write(args.category + ' : ' + str(values) + '\n')
def get_args():
parser = argparse.ArgumentParser(description='ANOMALYDETECTION')
parser.add_argument('--phase', choices=['train','test'], default='train')
parser.add_argument('--dataset_path', default=r'D:\Dataset\mvtec_anomaly_detection')#'/home/changwoo/hdd/datasets/mvtec_anomaly_detection')
parser.add_argument('--category', default='zipper')
parser.add_argument('--num_epochs', default=1)
parser.add_argument('--batch_size', default=32)
parser.add_argument('--load_size', default=256) # 256
parser.add_argument('--input_size', default=224)
parser.add_argument('--coreset_sampling_ratio', default=0.01)
parser.add_argument('--project_root_path', default=r'D:\Project_Train_Results\mvtec_anomaly_detection\210622\test') #'/home/changwoo/hdd/project_results/patchcore/test')
parser.add_argument('--save_src_code', default=True)
parser.add_argument('--save_anomaly_map', default=True)
parser.add_argument('--n_neighbors', type=int, default=9)
parser.add_argument('--weights_file_version', type=str, default=None) #'version_1'
args = parser.parse_args()
return args
if __name__ == '__main__':
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
args = get_args()
trainer = pl.Trainer.from_argparse_args(args, default_root_dir=os.path.join(args.project_root_path, args.category), max_epochs=args.num_epochs, gpus=[1]) #, check_val_every_n_epoch=args.val_freq, num_sanity_val_steps=0) # ,fast_dev_run=True)
if args.phase == 'train':
model = STPM(hparams=args)
trainer.fit(model)
trainer.test(model)
elif args.phase == 'test':
# selet weights file.
weights_file_path = auto_select_weights_file(args.weights_file_version) # auto select if args.weights_file_version == None
if weights_file_path != None:
model = STPM(hparams=args).load_from_checkpoint(weights_file_path)
trainer.test(model)
else:
print('Weights file is not found!')