-
Notifications
You must be signed in to change notification settings - Fork 2
/
train.py
359 lines (296 loc) · 15.4 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
import os
import time
from datetime import datetime
import random
import argparse
import logging
from collections import defaultdict
import numpy as np
import cv2
import torch
import torch.nn.functional as F
from segment_anything import SamPredictor, sam_model_registry
from segment_anything.utils.transforms import ResizeLongestSide
def clip_gradient(optimizer, grad_clip):
"""
For calibrating misalignment gradient via cliping gradient technique
:param optimizer:
:param grad_clip:
:return:
"""
for group in optimizer.param_groups:
for param in group['params']:
if param.grad is not None:
param.grad.data.clamp_(-grad_clip, grad_clip)
def adjust_lr(optimizer, init_lr, epoch, decay_rate=0.1, decay_epoch=30):
decay = decay_rate ** (epoch // decay_epoch)
for param_group in optimizer.param_groups:
param_group['lr'] *= decay
def structure_loss(pred, mask):
weit = 1 + 5 * torch.abs(F.avg_pool2d(mask, kernel_size=31, stride=1, padding=15) - mask)
wbce = F.binary_cross_entropy_with_logits(pred, mask, reduce='none')
wbce = (weit * wbce).sum(dim=(2, 3)) / weit.sum(dim=(2, 3))
pred = torch.sigmoid(pred)
inter = ((pred * mask) * weit).sum(dim=(2, 3))
union = ((pred + mask) * weit).sum(dim=(2, 3))
wiou = 1 - (inter + 1) / (union - inter + 1)
return (wbce + wiou).mean()
def test(model, path, dataset):
model.eval()
predictor_tuned = SamPredictor(model)
data_path = os.path.join(path, dataset)
image_root = '{}/images/'.format(data_path)
gt_root = '{}/masks/'.format(data_path)
num1 = len(os.listdir(gt_root))
images_path_list = [f for f in os.listdir(image_root) if f.endswith('.jpg') or f.endswith('.png') or f.endswith('.jpeg')]
images_path_list = sorted(images_path_list)
DSC = 0.0
for i in range(num1):
image = cv2.imread(image_root+''+images_path_list[i])
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
mask = cv2.imread(gt_root+''+images_path_list[i], cv2.IMREAD_GRAYSCALE)
mask = mask/255.0
H, W = mask.shape
y_indices, x_indices = np.where(mask > 0)
if(len(x_indices) == 0 or len(y_indices) == 0):
x_min, x_max = 0, W-1
y_min, y_max = 0, H-1
else:
x_min, x_max = np.min(x_indices), np.max(x_indices)
y_min, y_max = np.min(y_indices), np.max(y_indices)
# add perturbation to bounding box coordinates
perturb_h_len = 30 #100
x_min = max(0, x_min - perturb_h_len)
x_max = min(W, x_max + perturb_h_len)
y_min = max(0, y_min - perturb_h_len)
y_max = min(H, y_max + perturb_h_len)
input_bbox = np.array([x_min, y_min, x_max, y_max])
predictor_tuned.set_image(image)
pred, _, _ = predictor_tuned.predict(
point_coords=None,
box=input_bbox,
multimask_output=False,
)
# eval Dice
input = np.where(np.array(pred) >= 0.5, 1, 0)
target = np.where(np.array(mask) > 0.1, 1, 0)
N = mask.shape
smooth = 1
input_flat = np.reshape(input, (-1))
target_flat = np.reshape(target, (-1))
intersection = (input_flat * target_flat)
dice = (2 * intersection.sum() + smooth) / (input.sum() + target.sum() + smooth)
dice = '{:.4f}'.format(dice)
dice = float(dice)
DSC = DSC + dice
return DSC / num1, num1
def train(transformed_data, ground_truth_masks, bbox_coords, model, optimizer, epoch, test_path, model_name = 'SAM'):
model.train()
global best
global total_train_time
time_before_epoch_start = time.time()
size_rates = [1]
keys = transformed_data.keys()
epoch_losses = []
i = 0
for k in keys:
for rate in size_rates:
optimizer.zero_grad()
# ---- data prepare ----
input_image = transformed_data[k]['image'].cuda()
input_size = transformed_data[k]['input_size']
original_image_size = transformed_data[k]['original_image_size']
image_embedding = model.image_encoder(input_image)
prompt_box = bbox_coords[k]
box = transform.apply_boxes(prompt_box, original_image_size)
box_torch = torch.as_tensor(box, dtype=torch.float, device='cuda')
if len(box_torch.shape) == 2:
box_torch = box_torch[:, None, :] # (B, 1, 4)
box_torch = box_torch[None, :]
sparse_embeddings, dense_embeddings = model.prompt_encoder(
points=None,
boxes=box_torch,
masks=None,
)
low_res_masks, iou_predictions = model.mask_decoder(
image_embeddings=image_embedding,
image_pe=model.prompt_encoder.get_dense_pe(),
sparse_prompt_embeddings=sparse_embeddings,
dense_prompt_embeddings=dense_embeddings,
multimask_output=False,
)
gt_mask_resized = torch.from_numpy(np.resize(ground_truth_masks[k], (1, 1, ground_truth_masks[k].shape[0], ground_truth_masks[k].shape[1]))).cuda()
gt_binary_mask = torch.as_tensor(gt_mask_resized > 0.0, dtype=torch.float32)
upscaled_masks = model.postprocess_masks(low_res_masks, input_size, original_image_size).cuda()
loss = structure_loss(upscaled_masks, gt_binary_mask)
loss.backward()
clip_gradient(optimizer, opt.clip)
optimizer.step()
i += 1
if i % 250 == 0 or i == total_step:
print('{} Epoch [{:03d}/{:03d}], Step [{:04d}/{:04d}], '
' loss: {:0.4f}]'.
format(datetime.now(), epoch, opt.epoch, i, total_step,
loss.item()))
epoch_losses.append(loss.item())
print('EPOCH: '+ str(epoch) + ' loss: ' + str(np.mean(epoch_losses)))
time_after_epoch_end = time.time()
total_train_time += (time_after_epoch_end - time_before_epoch_start)
print('total train time till current epoch: '+ str(total_train_time))
logging.info('total train time till current epoch: '+ str(total_train_time))
# save model
save_path = (opt.train_save)
if not os.path.exists(save_path):
os.makedirs(save_path)
torch.save(model.state_dict(), save_path + '' + model_name + '-last.pth')
# choose the best model
global dict_plot
if (epoch + 1) % 1 == 0:
total_dice = 0
total_images = 0
for dataset in ['Kvasir']: #['CVC-300', 'CVC-ClinicDB', 'Kvasir', 'CVC-ColonDB', 'ETIS-LaribPolypDB']:
dataset_dice, n_images = test(model, test_path, dataset)
total_dice += (n_images*dataset_dice)
total_images += n_images
logging.info('epoch: {}, dataset: {}, dice: {}'.format(epoch, dataset, dataset_dice))
print(dataset, ': ', dataset_dice)
dict_plot[dataset].append(dataset_dice)
meandice = total_dice/total_images
dict_plot['test'].append(meandice)
print('Validation dice score: {}'.format(meandice))
logging.info('Validation dice score: {}'.format(meandice))
if meandice > best:
print('##################### Dice score improved from {} to {}'.format(best, meandice))
logging.info('##################### Dice score improved from {} to {}'.format(best, meandice))
best = meandice
torch.save(model.state_dict(), save_path + '' + model_name + '-best.pth')
if __name__ == '__main__':
dict_plot = {'CVC-ClinicDB':[], 'Kvasir':[], 'CVC-300':[], 'CVC-ColonDB':[], 'ETIS-LaribPolypDB':[], 'test':[]} #{'CVC-ClinicDB':[], 'test':[]} #
name = ['CVC-ClinicDB', 'Kvasir', 'CVC-300', 'CVC-ColonDB', 'ETIS-LaribPolypDB', 'test'] #['CVC-ClinicDB', 'test']
n_images = 900
perturb_h_len = 50
perturb_l_len = 0
freeze_image_encoder = 0
freeze_decoder = 1
##################model_name#############################
model_name = 'PolypSAM_freeze_mask_decoder_vit_b_train_p'+str(perturb_l_len)+'_'+str(perturb_h_len)+'_test_p30_Kvasirbest_bs1_random_shot'+str(n_images)+'_e100_Run1'
###############################################
print(model_name)
parser = argparse.ArgumentParser()
parser.add_argument('--epoch', type=int,
default=100, help='epoch number')
parser.add_argument('--lr', type=float,
default=1e-4, help='learning rate')
parser.add_argument('--optimizer', type=str,
default='AdamW', help='choosing optimizer AdamW or SGD')
parser.add_argument('--augmentation',
default=False, help='choose to do random flip rotation')
parser.add_argument('--batchsize', type=int,
default=1, help='training batch size')
parser.add_argument('--img_size', type=int,
default=1024, help='training dataset size')
parser.add_argument('--clip', type=float,
default=0.5, help='gradient clipping margin')
parser.add_argument('--decay_rate', type=float,
default=0.1, help='decay rate of learning rate')
parser.add_argument('--decay_epoch', type=int,
default=300, help='every n epochs decay learning rate')
parser.add_argument('--train_path', type=str,
default='./data/polyp/TrainDatasetKvasir/',
help='path to train dataset')
parser.add_argument('--test_path', type=str,
default='./data/polyp/TestDataset/',
help='path to testing Kvasir dataset')
parser.add_argument('--train_save', type=str,
default='./model_pth/'+model_name+'/')
opt = parser.parse_args()
logging.basicConfig(filename='train_log_'+model_name+'.log',
format='[%(asctime)s-%(filename)s-%(levelname)s:%(message)s]',
level=logging.INFO, filemode='a', datefmt='%Y-%m-%d %I:%M:%S %p')
# ---- build models ----
model_type = 'vit_b'
checkpoint = './pretrained_pth/sam/sam_vit_b_01ec64.pth'#sam_vit_b_01ec64.pth' #sam_vit_l_0b3195.pth
model = sam_model_registry[model_type](checkpoint=checkpoint)
if freeze_image_encoder:
print("Freezing image encoder")
for param in model.image_encoder.parameters():
param.requires_grad = False
if freeze_decoder:
print("Freezing mask decoder")
for param in model.mask_decoder.parameters():
param.requires_grad = False
model.cuda()
# Zero-shot inference of SAM
total_dice, total_images = 0.0, 0.0
for dataset in ['Kvasir']: #['CVC-300', 'CVC-ClinicDB', 'Kvasir', 'CVC-ColonDB', 'ETIS-LaribPolypDB']:
dataset_dice, n_imgs = test(model, opt.test_path, dataset)
total_dice += (n_imgs*dataset_dice)
total_images += n_imgs
logging.info('dataset: {}, dice: {}'.format(dataset, dataset_dice))
print(dataset, ': ', dataset_dice)
meandice = total_dice/total_images
print('Zero-shot validation dice score: {}'.format(meandice))
logging.info('Zero-shot validation dice score: {}'.format(meandice))
best = 0
params = list(model.image_encoder.parameters()) + list(model.prompt_encoder.parameters()) + list(model.mask_decoder.parameters()) #+ list(model.out.parameters()) #+ list(model.pvt_cascade.parameters()) #+ list(model.trans2pvt.parameters()) #+ list(model.pvt_stage2.parameters()) + list(model.pvt_norm2.parameters()) + list(model.pvt_stage3.parameters()) + list(model.pvt_norm3.parameters()) + list(model.pvt_stage4.parameters()) + list(model.pvt_norm4.parameters()) + list(model.decoder.parameters()) #.mask_decoder.
if opt.optimizer == 'AdamW':
optimizer = torch.optim.AdamW(params, opt.lr, weight_decay=1e-4)
#optimizer = torch.optim.Adam(params, opt.lr, weight_decay=0)
else:
optimizer = torch.optim.SGD(params, opt.lr, weight_decay=1e-4, momentum=0.9)
print(optimizer)
image_root = '{}/images/'.format(opt.train_path)
gt_root = '{}/masks/'.format(opt.train_path)
# sort images
images_path_list = sorted([f for f in os.listdir(image_root) if f.endswith('.jpg') or f.endswith('.png')])
# select n images from the dataset
img_idxs = random.sample(range(0, len(images_path_list)), n_images)
print(len(img_idxs), 'image indexes ', img_idxs)
logging.info(str(len(img_idxs)) + 'image indexes '+ str(img_idxs))
bbox_coords = {}
for k in img_idxs:
im = cv2.imread(gt_root+''+images_path_list[k])
gray=cv2.cvtColor(im,cv2.COLOR_BGR2GRAY)
y_indices, x_indices = np.where(gray > 0)
x_min, x_max = np.min(x_indices), np.max(x_indices)
y_min, y_max = np.min(y_indices), np.max(y_indices)
# add perturbation to bounding box coordinates
H, W = gray.shape
###### For fixed perturbations
#x_min = max(0, x_min - perturb_h_len)
#x_max = min(W, x_max + perturb_h_len)
#y_min = max(0, y_min - perturb_h_len)
#y_max = min(H, y_max + perturb_h_len)
###### For variable perturbations
x_min = max(0, x_min - np.random.randint(perturb_l_len, perturb_h_len))
x_max = min(W, x_max + np.random.randint(perturb_l_len, perturb_h_len))
y_min = max(0, y_min - np.random.randint(perturb_l_len, perturb_h_len))
y_max = min(H, y_max + np.random.randint(perturb_l_len, perturb_h_len))
bbox_coords[images_path_list[k]] = np.array([x_min, y_min, x_max, y_max])
#print(bbox_coords)
transformed_data = defaultdict(dict)
masks = defaultdict(dict)
for k in img_idxs:
image = cv2.imread(image_root+''+images_path_list[k])
image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
mask = cv2.imread(gt_root+''+images_path_list[k], cv2.IMREAD_GRAYSCALE)
mask = mask/255.0
transform = ResizeLongestSide(model.image_encoder.img_size)
input_image = transform.apply_image(image)
input_image_torch = torch.as_tensor(input_image, device='cuda')
transformed_image = input_image_torch.permute(2, 0, 1).contiguous()[None, :, :, :]
input_image = model.preprocess(transformed_image)
original_image_size = image.shape[:2]
input_size = tuple(transformed_image.shape[-2:])
transformed_data[images_path_list[k]]['image'] = input_image
transformed_data[images_path_list[k]]['input_size'] = input_size
transformed_data[images_path_list[k]]['original_image_size'] = original_image_size
masks[images_path_list[k]] = mask
total_step = len(transformed_data)
print("#" * 20, "Start Training", "#" * 20)
total_train_time = 0
for epoch in range(1, opt.epoch):
adjust_lr(optimizer, opt.lr, epoch, opt.decay_rate, opt.decay_epoch)
train(transformed_data, masks, bbox_coords, model, optimizer, epoch, opt.test_path, model_name = model_name)
print('avg train time: '+ str(total_train_time/(opt.epoch-1)))
logging.info('avg train time: '+ str(total_train_time/(opt.epoch-1)))