-
Notifications
You must be signed in to change notification settings - Fork 0
/
utils.py
589 lines (467 loc) · 22.2 KB
/
utils.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
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
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
"""
Programmed by Behnaaz Fakhar <fakhar.behnaz@gmail.com>
* 2023-02-12 Initial coding
"""
import cv2, torch, os, skimage, torchvision, random
import json
import numpy as np
import albumentations as A
import segmentation_models_pytorch as smp
import torch.nn as nn
import matplotlib.pyplot as plt
from torch.utils.data import Dataset
from albumentations.pytorch import ToTensorV2
from copy import deepcopy
from torch.utils.data import DataLoader
from torchvision import transforms
from tqdm import tqdm
IMGSZ = 512
class MyDataset(Dataset):
def __init__(self, dataset_dir, transform=None):
self.dataset_dir = dataset_dir
self.transform = transform
self.images = [x for x in os.listdir(os.path.join(dataset_dir, 'images')) if x.endswith(".png")]
def __len__(self):
return len(self.images)
def __getitem__(self, index):
try:
image = cv2.imread(os.path.join(self.dataset_dir, 'images', self.images[index]), cv2.IMREAD_GRAYSCALE)
# image = skimage.io.imread(os.path.join(self.dataset_dir, 'images', self.images[index]))
mask = cv2.imread(os.path.join(self.dataset_dir, 'masks', self.images[index].replace('.png','_mask.png')), cv2.IMREAD_GRAYSCALE)
if self.transform is not None:
transformed = self.transform(image=image, mask=mask)
image, mask = transformed['image'], transformed['mask']
return image, mask, self.images[index]
except:
image = cv2.imread(os.path.join(self.dataset_dir, 'images', self.images[index]), cv2.IMREAD_GRAYSCALE)
# image = skimage.io.imread(os.path.join(self.dataset_dir, 'images', self.images[index]))
if self.transform is not None:
image = self.transform(image=image)['image']
return image, self.images[index]
def get_train_transforms():
return A.Compose(
[
A.Resize(IMGSZ, IMGSZ, interpolation=cv2.INTER_NEAREST, always_apply=False, p=1),
A.HorizontalFlip(p=0.5),
# A.ColorJitter(brightness=0.25, contrast=0.25, saturation=0.2, hue=0.1, always_apply=False, p=0.5),
A.VerticalFlip(p=0.25),
# A.ToFloat(),
# A.Normalize(
# mean = [0.485, 0.456, 0.406],
# std = [0.229, 0.224, 0.225],
# max_pixel_value=255.0,
# ),
ToTensorV2(),
],
)
def get_val_transforms():
return A.Compose(
[
A.Resize(IMGSZ, IMGSZ, interpolation=cv2.INTER_NEAREST, always_apply=False, p=1),
# A.ToFloat(),
# A.Normalize(
# mean = [0.485, 0.456, 0.406],
# std = [0.229, 0.224, 0.225],
# max_pixel_value=255.0,
# ),
ToTensorV2(),
],
)
def build_dataloders(
train_dir,
val_dir,
test_dir = None,
train_bs = 5,
val_bs = 1,
num_workers = 0,
pin_memory = True,
):
train_ds = MyDataset(
dataset_dir = train_dir,
transform = get_train_transforms(),
)
train_loader = DataLoader(
train_ds,
batch_size = train_bs,
num_workers = num_workers,
pin_memory = False,
shuffle = True,
)
val_ds = MyDataset(
dataset_dir = val_dir,
transform = get_val_transforms(),
)
val_loader = DataLoader(
val_ds,
batch_size = val_bs,
num_workers = num_workers,
pin_memory = False,
shuffle = False,
)
if test_dir:
test_ds = MyDataset(
dataset_dir = test_dir,
transform = get_val_transforms(),
)
test_loader = DataLoader(
test_ds,
batch_size = val_bs,
num_workers = num_workers,
pin_memory = False,
shuffle = False,
)
else:
test_loader = None
return train_loader, val_loader, test_loader
def get_random_color(classes):
num = len(classes)
gray_level = [0,255] if num <= 3 else [0,125,255] if num <= 15 else [0,50,100,150,200,250]
colors = {}
color_list = []
while len(color_list) < num:
color = random.choices(gray_level, k = 3)
if color not in color_list:
color_list.append(color)
colors[classes[len(colors)]] = color
return colors
def get_class_colors(classes):
colors = {
'FP': [128, 0, 128], # Purple for class 1
"FPD": [0, 255, 255], # Cyan for class 2
"ASH": [0, 128, 0], # Green for class 3
"bf": [255, 0, 0], # Red for class 4
}
return colors
class Trainer:
def __init__(
self,
dataset_dir,
out_dir,
classes,
patience,
encoder = 'resnet34',
decoder = 'Unet',
optimizer = 'Nadam',
init_lr = 0.0025,
num_epochs = 100,
train_bs = 5,
val_bs = 1,
device ='cuda'
):
self.model = self._get_model(encoder=encoder, decoder=decoder, num_classes=len(classes)+1, device=device)
self.opt = self._get_optimizer(model=self.model, name=optimizer, init_lr=init_lr)
self.train_loader, self.val_loader, self.test_loader = build_dataloders(
train_dir = os.path.join(dataset_dir, 'train'),
val_dir = os.path.join(dataset_dir, 'train'),
test_dir = os.path.join(dataset_dir, 'test'),
train_bs = train_bs,
val_bs = val_bs,
)
# self.loss = nn.BCEWithLogitsLoss()
self.loss = nn.CrossEntropyLoss()
self.scaler = torch.cuda.amp.GradScaler()
self.device = device
self.patience = patience
self.num_epochs = num_epochs
self.classes = classes
self.dataset_dir = dataset_dir
self.colors = get_random_color(classes)
os.makedirs(os.path.join(out_dir, 'dump'), exist_ok=True)
runs = [x for x in os.listdir(os.path.join(out_dir, 'dump')) if '.' not in x]
if len(runs) == 0:
self.out_dir = os.path.join(out_dir, 'dump/run_0')
os.makedirs(os.path.join(out_dir, 'dump/run_0/evaluation'), exist_ok=True)
else:
idx = 0
for name in runs:
idx = max(idx, int(name.split('_')[-1]))
self.out_dir = os.path.join(out_dir, f'dump/run_{idx+1}')
os.makedirs(os.path.join(out_dir, f'dump/run_{idx+1}/evaluation'), exist_ok=True)
@staticmethod
def _get_model(encoder, decoder, num_classes, device, encoder_weights="imagenet"):
if decoder == 'Unet':
model = smp.Unet(
encoder_name = encoder,
encoder_weights = encoder_weights,
in_channels = 1,
classes = num_classes,
)
elif decoder == 'UnetPlusPlus':
model = smp.UnetPlusPlus(
encoder_name = encoder,
encoder_weights = encoder_weights,
in_channels = 1,
classes = num_classes,
)
model.to(device)
return model
@staticmethod
def _get_optimizer(model, name, init_lr):
if name == 'Adam':
opt = torch.optim.Adam(
model.parameters(),
lr = init_lr,
betas = (0.9, 0.999),
eps = 1e-08,
weight_decay = 0,
amsgrad = False,
)
elif name == 'Nadam':
opt = torch.optim.NAdam(
model.parameters(),
lr = init_lr,
betas =(0.9, 0.999),
eps = 1e-08,
weight_decay = 0,
momentum_decay= 0.001,
foreach = None
)
return opt
def train_fn(self, epoch):
loop = tqdm(self.train_loader)
total_loss = 0
for batch_idx, (data, targets, _) in enumerate(loop):
data = data.to(device=self.device).float()
targets = targets.long().to(device=self.device)
# forward
predictions = self.model(data)
loss = self.loss(predictions, targets)
# backward
total_loss += loss.item()
self.opt.zero_grad()
loss.backward()
self.opt.step()
# update tqdm loop
loop.set_postfix(loss=total_loss/(batch_idx+1))
return total_loss/len(self.train_loader)
def _save_checkpoint(self, ckpt):
print(f"=> Saving checkpoint with training loss = {ckpt['train_loss']}, dice_val = {ckpt['DICE_score_val']} and dice_test = {ckpt['DICE_score_test']}")
torch.save(ckpt, os.path.join(self.out_dir, 'training_checkpoint.pth'))
def _save_single_pred(self, preds, name, is_test):
rgb = cv2.imread(os.path.join(self.dataset_dir, 'test/images' if is_test else 'val/images', name[0]), cv2.IMREAD_COLOR)
rgb = A.Compose(
[
A.Resize(IMGSZ, IMGSZ, interpolation=cv2.INTER_NEAREST, always_apply=False, p=1),
],
)(image=rgb)['image']
for ch in range(3):
for idx, cls_name in enumerate(self.classes):
rgb[:,:,ch][preds[idx+1]!=0] = self.colors[cls_name][ch]
skimage.io.imsave(os.path.join(self.out_dir, 'evaluation', f"pred_{name[0].split('.')[0]}.png"), rgb)
def _save_predictions_as_imgs(self, loader, is_test=True, conf = 0.95):
self.model.eval()
if is_test:
for idx, (x, _, name) in enumerate(loader):
x = x.to(device=self.device).float()
with torch.no_grad():
preds = torch.softmax(self.model(x).squeeze(0), dim=0)
preds = (preds > conf).float()
self._save_single_pred(preds.cpu().numpy(), name, is_test)
else:
for idx, (x, y, name) in enumerate(loader):
x = x.to(device=self.device).float()
with torch.no_grad():
preds = torch.sigmoid(self.model(x))
preds = (preds > conf).float()
self._save_single_pred(preds.cpu().numpy(), name, is_test)
self.model.train()
def _trans_mask(self, y):
mask = torch.zeros(len(self.classes)+1, IMGSZ, IMGSZ)
for item in range(len(self.classes)+1):
mask[item,:,:] = y == item
return mask
def _check_accuracy(self):
num_correct = 0
num_pixels = 0
dice_score = 0
self.model.eval()
with torch.inference_mode():
dice_score_c_loop = {'FP':0,'FPD':0,'ASH':0,'bf':0}
num_correct_c = 0
num_pixels_c = 0
for x, y, _ in self.val_loader:
x = x.to(self.device).float()
y = self._trans_mask(y).to(self.device)
preds = torch.softmax(self.model(x).squeeze(0), dim=0)
preds = (preds > 0.5).float()
num_correct += (preds[1:,:,:] == y[1:,:,:]).sum()
num_pixels += torch.numel(preds[1:,:,:])
dice_score += (2 * (preds[1:,:,:] * y[1:,:,:]).sum()) / ((preds[1:,:,:] + y[1:,:,:]).sum() + 1e-8)
# Dice Score for each class
for c in range(1,len(y)):
num_correct_c += (preds[c,:,:] == y[c,:,:]).sum()
num_pixels_c += torch.numel(preds[c,:,:])
# print((2 * (preds[c,:,:] * y[c,:,:]).sum()) / ((preds[c,:,:] + y[c,:,:]).sum() + 1e-8))
dice_score_c_loop[self.classes[c-1]] += (2 * (preds[c,:,:] * y[c,:,:]).sum()) / ((preds[c,:,:] + y[c,:,:]).sum() + 1e-8)
dice_score_c = {key: float(value / len(self.val_loader)) for key, value in dice_score_c_loop.items()}
print(
f"Prefomance: accuracy = {num_correct/num_pixels*100:.2f}, DICE score = {dice_score/len(self.val_loader)}, DICE_score_c = {dice_score_c}"
)
self.model.train()
return dice_score/len(self.val_loader),dice_score_c
def _check_accuracy_test(self):
num_correct = 0
num_pixels = 0
dice_score = 0
self.model.eval()
with torch.inference_mode():
dice_score_c_loop = {'FP':0,'FPD':0,'ASH':0,'bf':0}
num_correct_c = 0
num_pixels_c = 0
for x, y, _ in self.test_loader:
x = x.to(self.device).float()
y = self._trans_mask(y).to(self.device)
preds = torch.softmax(self.model(x).squeeze(0), dim=0)
preds = (preds > 0.5).float()
num_correct += (preds[1:,:,:] == y[1:,:,:]).sum()
num_pixels += torch.numel(preds[1:,:,:])
dice_score += (2 * (preds[1:,:,:] * y[1:,:,:]).sum()) / ((preds[1:,:,:] + y[1:,:,:]).sum() + 1e-8)
# Dice Score for each class
for c in range(1,len(y)):
num_correct_c += (preds[c,:,:] == y[c,:,:]).sum()
num_pixels_c += torch.numel(preds[c,:,:])
dice_score_c_loop[self.classes[c-1]] += (2 * (preds[c,:,:] * y[c,:,:]).sum()) / ((preds[c,:,:] + y[c,:,:]).sum() + 1e-8)
dice_score_c = {key: float(value / len(self.test_loader)) for key, value in dice_score_c_loop.items()}
print(
f"Prefomance: accuracy = {num_correct/num_pixels*100:.2f}, DICE score = {dice_score/len(self.test_loader)}, DICE_score_c = {dice_score_c}"
)
self.model.train()
return dice_score/len(self.test_loader), dice_score_c
def train(self):
SCORE = 0.
counter = 0
with open(os.path.join(self.out_dir, 'training_dice_score_val.json'), 'a') as f_score_val, \
open(os.path.join(self.out_dir, 'training_dice_score_ind_c_val.json'), 'a') as f_score_ind_c_val, \
open(os.path.join(self.out_dir, 'training_dice_score_test.json'), 'a') as f_score_test, \
open(os.path.join(self.out_dir, 'training_dice_score_ind_c_test.json'), 'a') as f_score_ind_c_test, \
open(os.path.join(self.out_dir, 'training_train_loss.json'), 'a') as f_loss:
for epoch in range(self.num_epochs):
counter += 1
print(f"On epoch {epoch+1}/{self.num_epochs}:")
train_loss = self.train_fn(epoch)
# check accuracy
score, score_c = self._check_accuracy()
score_test, score_c_test = self._check_accuracy_test()
# Save dice score for val for each epoch
json.dump(score.detach().cpu().numpy().tolist(), f_score_val)
f_score_val.write('\n') # Add a newline to separate entries in the file
# Save dice score for each class in val for each epoch
json.dump(score_c, f_score_ind_c_val)
f_score_ind_c_val.write('\n') # Add a newline to separate entries in the file
# Save dice score for test for each epoch
json.dump(score_test.detach().cpu().numpy().tolist(), f_score_test)
f_score_test.write('\n')
# Save dice score for each class in test for each epoch
json.dump(score_c_test, f_score_ind_c_test)
f_score_ind_c_test.write('\n')
# Save train loss for each epoch
json.dump(train_loss, f_loss)
f_loss.write('\n')
if score > SCORE:
counter = 0
SCORE = score
checkpoint = {
"state_dict": self.model.state_dict(),
"optimizer": self.opt.state_dict(),
"DICE_score_val": SCORE,
"DICE_score_ind_c_val": score_c,
"DICE_score_test": score_test,
"DICE_score_ind_c_test": score_c_test,
"train_loss": train_loss
}
self._save_predictions_as_imgs(self.test_loader, is_test=True)
self._save_checkpoint(checkpoint)
if counter == self.patience: break
class Evaluator:
def __init__(
self,
model_dir,
classes,
encoder = 'efficientnet-b0',
decoder = 'Unet',
ckpt_name = 'training_checkpoint.pth',
device = 'cuda'
):
self.model = self._load_model(model_dir, encoder, decoder, ckpt_name, len(classes)+1, device)
self.device = device
self.masks = {}
self.classes = classes
self.colors = get_class_colors(classes)
self.model_dir = model_dir
def _load_model(self, path, encoder, decoder, ckpt_name, num_classes, device):
ckpt = torch.load(os.path.join(path, ckpt_name))
if decoder == 'Unet':
model = smp.Unet(
encoder_name = encoder,
encoder_weights = None,
in_channels = 1,
classes = num_classes,
)
elif decoder == 'UnetPlusPlus':
model = smp.UnetPlusPlus(
encoder_name = encoder,
encoder_weights = None,
in_channels = 1,
classes = num_classes,
)
model.to(device)
model.load_state_dict(ckpt["state_dict"])
model.eval()
return model
def _get_loader(self, path):
ds = MyDataset(
dataset_dir = path,
transform = get_val_transforms(),
)
return DataLoader(
ds,
batch_size = 1,
num_workers = 0,
pin_memory = False,
shuffle = False,
)
def evaluate(self, dataset_dir, dataset_name, conf = 0.5, visualize=True):
loader = self._get_loader(os.path.join(dataset_dir, dataset_name))
for batch_idx, (x, _, name) in enumerate(loader):
print(name)
x = x.to(device=self.device)
with torch.no_grad():
preds = torch.softmax(self.model(x.float()).squeeze(0),dim = 0)
preds = (preds>0.5).cpu().numpy()
if visualize:
os.makedirs(os.path.join(dataset_dir,dataset_name,'pred',self.model_dir.split("/dump/")[1]), exist_ok=True)
self._save_single_pred(dataset_dir, dataset_name, preds, name)
def _save_single_pred(self, dataset_dir, dataset_name, preds, name):
rgb = cv2.imread(os.path.join(dataset_dir, dataset_name,'images', name[0]), cv2.IMREAD_COLOR)
rgb = A.Compose(
[
A.Resize(IMGSZ, IMGSZ, interpolation=cv2.INTER_NEAREST, always_apply=True, p=1),
],
)(image=rgb)['image']
for ch in range(3):
for idx, cls_name in enumerate(self.classes):
rgb[:, :, ch][preds[idx + 1] != 0] = self.colors[cls_name][ch]
# Get contours for class FP
contours_fp, _ = cv2.findContours((preds[1] == 1).astype(np.uint8), cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
# Draw arrow from center of FP towards FPD if FPD is present
if np.max(preds[2]) != 0:
# Get center of FP
cords_FP = np.where(preds[1]==1)
cy_fp,cx_fp = int(np.mean(cords_FP[0])),int(np.mean(cords_FP[1]))
# Get center of FPD
cords_FPD = np.where(preds[2]==1)
cy_fpd,cx_fpd = int(np.mean(cords_FPD[0])),int(np.mean(cords_FPD[1]))
# Calculate the vector from center of FP to center of FPD
vector_fp_fpd = np.array([cx_fpd - cx_fp, cy_fpd - cy_fp])
# Scale the vector to make the arrow longer (e.g., multiplying by 1.2)
scaled_vector = 1.1 * vector_fp_fpd
# Calculate the endpoint of the arrow after scaling
endpoint = (cx_fp + scaled_vector[0], cy_fp + scaled_vector[1])
# Draw small circle at the beginning of the arrow
cv2.circle(rgb, (cx_fp, cy_fp), radius=10, color=[0, 0, 255]) # Blue color for the circle
# Draw arrow
arrow_color = [0, 0, 255] # Blue color for the arrow
rgb = cv2.arrowedLine(rgb, (cx_fp, cy_fp), (int(endpoint[0]), int(endpoint[1])), arrow_color, thickness=2)
plt.imshow(rgb)
plt.show()
skimage.io.imsave(os.path.join(dataset_dir, dataset_name,'pred', self.model_dir.split("/dump/")[1], name[0]), rgb)