/
train.py
382 lines (310 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
import os, time, cv2
import torch
import torch.optim as optim
import torch.nn as nn
from torch.utils.data import DataLoader
from torch.optim.lr_scheduler import StepLR, CyclicLR, LambdaLR
import numpy as np
from tqdm import tqdm
from tensorboardX import SummaryWriter
from Backbones.vit import VisionTransformer
from Backbones.models_denseformer import DenseFormer
from Backbones.discriminator import discriminator
from utils import *
from config import config
from dataset.MFIRST import G1G2Dataset
# from dataset.SIRST import G1G2Dataset
def get_customized_schedule_with_warmup(optimizer, num_warmup_steps, d_model=1.0, last_epoch=-1):
def lr_lambda(current_step):
current_step += 1
arg1 = current_step ** -0.5
arg2 = current_step * (num_warmup_steps ** -1.5)
return (d_model ** -0.5) * min(arg1, arg2)
return LambdaLR(optimizer, lr_lambda, last_epoch)
def train(g1_path_checkpoint=None, g2_path_checkpoint=None, dis_path_checkpoint=None, RESUME=False):
assert RESUME is False or g1_path_checkpoint is not None, 'if RESUME, checkpoint must be specified!'
# output_dir
root_result_dir = os.path.join('pytorch_outputs')
os.makedirs(root_result_dir, exist_ok=True)
model_result_dir = os.path.join(root_result_dir, 'models')
os.makedirs(model_result_dir, exist_ok=True)
images_dir = os.path.join(root_result_dir, 'images')
os.makedirs(images_dir, exist_ok=True)
# log
log_dir = os.path.join(root_result_dir, 'logs')
os.makedirs(log_dir, exist_ok=True)
summary_writer = SummaryWriter(log_dir)
# dataset
trainsplit = G1G2Dataset(mode='train')
trainset = DataLoader(trainsplit, batch_size=config.mini_batch_size, pin_memory=True,
num_workers=4, shuffle=True, drop_last=True)
testsplit = G1G2Dataset(mode='test.py')
testset = DataLoader(testsplit, batch_size=1, pin_memory=True,
num_workers=4, shuffle=False, drop_last=True)
# Model
g1 = DenseFormer()
g2 = VisionTransformer()
dis = discriminator()
if g1_path_checkpoint is not None:
g1_checkpoint = torch.load(g1_path_checkpoint)
g1.load_state_dict(g1_checkpoint['model_state'])
g2_checkpoint = torch.load(g2_path_checkpoint)
g2.load_state_dict(g2_checkpoint['model_state'])
dis_checkpoint = torch.load(dis_path_checkpoint)
dis.load_state_dict(dis_checkpoint['model_state'])
print('Load .pth successfully....')
g1.cuda()
g2.cuda()
dis.cuda()
# optimizer
optimizer_g1 = torch.optim.Adam(g1.parameters(), lr=1e-2)
optimizer_g2 = torch.optim.Adam(g2.parameters(), lr=1e-2)
optimizer_dis = torch.optim.Adam(dis.parameters(), lr=1e-4, betas=(0.5, 0.999), weight_decay=5e-4)
if RESUME:
optimizer_g1.load_state_dict(g1_checkpoint['optimizer_state'])
optimizer_g2.load_state_dict(g2_checkpoint['optimizer_state'])
optimizer_dis.load_state_dict(dis_checkpoint['optimizer_state'])
print('Resume last training....')
# scheduler_g1 = StepLR(optimizer_g1, step_size=config.step_size, gamma=config.gamma)
# scheduler_g2 = StepLR(optimizer_g2, step_size=config.step_size, gamma=config.gamma)
# scheduler_dis = StepLR(optimizer_dis, step_size=config.step_size, gamma=config.gamma)
scheduler_g1 = get_customized_schedule_with_warmup(optimizer_g1, num_warmup_steps=200, d_model=728)
scheduler_g2 = get_customized_schedule_with_warmup(optimizer_g2, num_warmup_steps=200, d_model=728)
scheduler_dis = StepLR(optimizer_dis, step_size=config.step_size, gamma=config.gamma)
# loss
loss1 = nn.BCEWithLogitsLoss()
it = 0
start_epoch = 0
if RESUME:
start_epoch = g1_checkpoint['epoch']
it = g1_checkpoint['it']
for epoch in range(start_epoch + 0, start_epoch + config.max_epoch_num):
# epoch
total_it_per_epoch = len(trainset)
for bt_idx, data in enumerate(tqdm(trainset)):
# batch
torch.cuda.empty_cache()
it = it + 1
summary_writer.add_scalar('lr/g1', float(optimizer_g1.param_groups[0]['lr']), it)
summary_writer.add_scalar('lr/g2', float(optimizer_g2.param_groups[0]['lr']), it)
summary_writer.add_scalar('lr/dis', float(optimizer_dis.param_groups[0]['lr']), it)
# dis
dis.train()
g1.eval()
g2.eval()
optimizer_g1.zero_grad()
optimizer_g2.zero_grad()
optimizer_dis.zero_grad()
# cuda
input_images, output_images = data['input_images'], data['output_images'] # [B, 1, 128, 128]
input_images = input_images.cuda(non_blocking=True).float()
output_images = output_images.cuda(non_blocking=True).float()
with torch.no_grad():
g1_out = g1(input_images) # [B, 1, 128, 128]
g1_out = torch.clamp(g1_out, 0.0, 1.0)
g2_out = g2(input_images) # [B, 1, 128, 128]
g2_out = torch.clamp(g2_out, 0.0, 1.0)
pos1 = torch.cat([input_images, 2 * output_images - 1], dim=1) # [B, 2, 128, 128]
neg1 = torch.cat([input_images, 2 * g1_out - 1], dim=1) # [B, 2, 128, 128]
neg2 = torch.cat([input_images, 2 * g2_out - 1], dim=1) # [B, 2, 128, 128]
disc_input = torch.cat([pos1, neg1, neg2], dim=0) # [3*B, 2, 128, 128]
logits_real, logits_fake1, logits_fake2, Lgc = dis(disc_input) # [B, 3] [B, 3] [B, 3] [B, 1]
const1 = torch.ones(config.mini_batch_size, 1).cuda(non_blocking=True).float()
const0 = torch.zeros(config.mini_batch_size, 1).cuda(non_blocking=True).float()
gen_gt = torch.cat([const1, const0, const0], dim=1)
gen_gt1 = torch.cat([const0, const1, const0], dim=1)
gen_gt2 = torch.cat([const0, const0, const1], dim=1)
ES0 = torch.mean(loss1(logits_real, gen_gt))
ES1 = torch.mean(loss1(logits_fake1, gen_gt1))
ES2 = torch.mean(loss1(logits_fake2, gen_gt2))
disc_loss = ES0 + ES1 + ES2
summary_writer.add_scalar('loss/dis', disc_loss, it)
# logger.info(" discriminator loss is {}".format(disc_loss))
disc_loss.backward()
optimizer_dis.step()
# g1
dis.eval()
g1.train()
g2.eval()
optimizer_g1.zero_grad()
optimizer_g2.zero_grad()
optimizer_dis.zero_grad()
g1_out = g1(input_images) # [B, 1, 128, 128]
g1_out = torch.clamp(g1_out, 0.0, 1.0)
# MD1 = torch.mean(torch.mul(torch.pow(g1_out - output_images, 2), output_images))
# FA1 = torch.mean(torch.mul(torch.pow(g1_out - output_images, 2), 1 - output_images))
# MF_loss1 = config.lambda1 * MD1 + FA1
axes = tuple(range(2, len(output_images.shape)))
precision = torch.mean(torch.sum(g1_out * output_images, axes) / (torch.sum(g1_out, axes) + 1e-6))
recall = torch.mean(torch.sum(g1_out * output_images, axes) / (torch.sum(output_images, axes) + 1e-6))
MF_loss1 = -50 * torch.pow((1 - precision), 3) * precision.log() - torch.pow((1 - recall), 3) * recall.log()
# F1 = 2 * recall * precision / (recall + precision + 1e-6)
# MF_loss1 = -10 * F1.log()
with torch.no_grad():
g2_out = g2(input_images) # [B, 1, 128, 128]
g2_out = torch.clamp(g2_out, 0.0, 1.0)
pos1 = torch.cat([input_images, 2 * output_images - 1], dim=1) # [B, 2, 128, 128]
neg1 = torch.cat([input_images, 2 * g1_out - 1], dim=1) # [B, 2, 128, 128]
neg2 = torch.cat([input_images, 2 * g2_out - 1], dim=1) # [B, 2, 128, 128]
disc_input = torch.cat([pos1, neg1, neg2], dim=0) # [3*B, 2, 128, 128]
with torch.no_grad():
logits_real, logits_fake1, logits_fake2, Lgc = dis(disc_input) # [B, 3] [B, 3] [B, 3] [B, 1]
const1 = torch.ones(config.mini_batch_size, 1).cuda(non_blocking=True).float()
const0 = torch.zeros(config.mini_batch_size, 1).cuda(non_blocking=True).float()
gen_gt = torch.cat([const1, const0, const0], dim=1)
gen_gt1 = torch.cat([const0, const1, const0], dim=1)
gen_gt2 = torch.cat([const0, const0, const1], dim=1)
gen_adv_loss1 = torch.mean(loss1(logits_fake1, gen_gt))
gen_loss1 = 100 * MF_loss1 + 10 * gen_adv_loss1 + 1 * Lgc
summary_writer.add_scalar('loss/g1', gen_loss1, it)
# logger.info(" g1 loss is {}".format(gen_loss1))
gen_loss1.backward()
optimizer_g1.step()
scheduler_g1.step()
# g2
dis.eval()
g1.eval()
g2.train()
optimizer_g1.zero_grad()
optimizer_g2.zero_grad()
optimizer_dis.zero_grad()
with torch.no_grad():
g1_out = g1(input_images) # [B, 1, 128, 128]
g1_out = torch.clamp(g1_out, 0.0, 1.0)
g2_out = g2(input_images) # [B, 1, 128, 128]
g2_out = torch.clamp(g2_out, 0.0, 1.0)
axes = tuple(range(2, len(output_images.shape)))
precision = torch.mean(torch.sum(g2_out * output_images, axes) / (torch.sum(g2_out, axes) + 1e-6))
recall = torch.mean(torch.sum(g2_out * output_images, axes) / (torch.sum(output_images, axes) + 1e-6))
MF_loss2 = -50 * torch.pow((1 - precision), 3) * precision.log() - torch.pow((1 - recall), 3) * recall.log()
pos1 = torch.cat([input_images, 2 * output_images - 1], dim=1) # [B, 2, 128, 128]
neg1 = torch.cat([input_images, 2 * g1_out - 1], dim=1) # [B, 2, 128, 128]
neg2 = torch.cat([input_images, 2 * g2_out - 1], dim=1) # [B, 2, 128, 128]
disc_input = torch.cat([pos1, neg1, neg2], dim=0) # [3*B, 2, 128, 128]
with torch.no_grad():
logits_real, logits_fake1, logits_fake2, Lgc = dis(disc_input) # [B, 3] [B, 3] [B, 3] [B, 1]
const1 = torch.ones(config.mini_batch_size, 1).cuda(non_blocking=True).float()
const0 = torch.zeros(config.mini_batch_size, 1).cuda(non_blocking=True).float()
gen_gt = torch.cat([const1, const0, const0], dim=1)
gen_gt1 = torch.cat([const0, const1, const0], dim=1)
gen_gt2 = torch.cat([const0, const0, const1], dim=1)
gen_adv_loss2 = torch.mean(loss1(logits_fake2, gen_gt))
gen_loss2 = 100 * MF_loss2 + 10 * gen_adv_loss2 + 1 * Lgc
summary_writer.add_scalar('loss/g2', gen_loss2, it)
gen_loss2.backward()
optimizer_g2.step()
scheduler_g2.step()
# test
sum_val_loss_g1 = 0
sum_val_false_ratio_g1 = 0
sum_val_detect_ratio_g1 = 0
sum_val_F1_g1 = 0
sum_val_loss_g2 = 0
sum_val_false_ratio_g2 = 0
sum_val_detect_ratio_g2 = 0
sum_val_F1_g2 = 0
sum_val_loss_g3 = 0
sum_val_false_ratio_g3 = 0
sum_val_detect_ratio_g3 = 0
sum_val_F1_g3 = 0
for bt_idx_test, data in enumerate(tqdm(testset)):
g1.eval()
g2.eval()
dis.eval()
optimizer_g1.zero_grad()
optimizer_g2.zero_grad()
optimizer_dis.zero_grad()
with torch.no_grad():
input_images, output_images = data['input_images'], data['output_images'] # [B, 1, 128, 128]
input_images = input_images.cuda(non_blocking=True).float()
output_images = output_images.cuda(non_blocking=True).float()
g1_out = g1(input_images) # [B, 1, 128, 128]
g1_out = torch.clamp(g1_out, 0.0, 1.0)
g2_out = g2(input_images) # [B, 1, 128, 128]
g2_out = torch.clamp(g2_out, 0.0, 1.0)
pos1 = torch.cat([input_images, 2 * output_images - 1], dim=1) # [B, 2, 128, 128]
neg1 = torch.cat([input_images, 2 * g1_out - 1], dim=1) # [B, 2, 128, 128]
neg2 = torch.cat([input_images, 2 * g2_out - 1], dim=1) # [B, 2, 128, 128]
disc_input = torch.cat([pos1, neg1, neg2], dim=0) # [3*B, 2, 128, 128]
_, logits_fake1, logits_fake2, _ = dis(disc_input)
g3_out = (g1_out * (logits_fake1[:, 0] / (logits_fake1[:, 0] + logits_fake2[:, 0])) + g2_out * (
logits_fake2[:, 0] / (logits_fake1[:, 0] + logits_fake2[:, 0]))) # jiaquan的方式进行融合
output_images = output_images.cpu().numpy()
g1_out = g1_out.detach().cpu().numpy()
g2_out = g2_out.detach().cpu().numpy()
g3_out = g3_out.detach().cpu().numpy()
# g1
val_loss_g1 = np.mean(np.square(g1_out - output_images))
sum_val_loss_g1 += val_loss_g1
val_false_ratio_g1 = np.mean(np.maximum(0, g1_out - output_images))
sum_val_false_ratio_g1 += val_false_ratio_g1
val_detect_ratio_g1 = np.sum(g1_out * output_images) / np.maximum(np.sum(output_images), 1)
sum_val_detect_ratio_g1 += val_detect_ratio_g1
val_F1_g1 = calculateF1Measure(g1_out, output_images, 0.5)
sum_val_F1_g1 += val_F1_g1
# g2
val_loss_g2 = np.mean(np.square(g2_out - output_images))
sum_val_loss_g2 += val_loss_g2
val_false_ratio_g2 = np.mean(np.maximum(0, g2_out - output_images))
sum_val_false_ratio_g2 += val_false_ratio_g2
val_detect_ratio_g2 = np.sum(g2_out * output_images) / np.maximum(np.sum(output_images), 1)
sum_val_detect_ratio_g2 += val_detect_ratio_g2
val_F1_g2 = calculateF1Measure(g2_out, output_images, 0.5)
sum_val_F1_g2 += val_F1_g2
# g3
val_loss_g3 = np.mean(np.square(g3_out - output_images))
sum_val_loss_g3 += val_loss_g3
val_false_ratio_g3 = np.mean(np.maximum(0, g3_out - output_images))
sum_val_false_ratio_g3 += val_false_ratio_g3
val_detect_ratio_g3 = np.sum(g3_out * output_images) / np.maximum(np.sum(output_images), 1)
sum_val_detect_ratio_g3 += val_detect_ratio_g3
val_F1_g3 = calculateF1Measure(g3_out, output_images, 0.5)
sum_val_F1_g3 += val_F1_g3
# save pic
output_image1 = np.squeeze(g1_out * 255.0) # /np.maximum(output_image1.max(),0.0001))
output_image2 = np.squeeze(g2_out * 255.0) # /np.maximum(output_image2.max(),0.0001))
output_image3 = np.squeeze(g3_out * 255.0) # /np.maximum(output_image3.max(),0.0001))
cv2.imwrite(os.path.join(images_dir, '%05d_G1.png' % (bt_idx_test)), np.uint8(output_image1))
cv2.imwrite(os.path.join(images_dir, '%05d_G2.png' % (bt_idx_test)), np.uint8(output_image2))
cv2.imwrite(os.path.join(images_dir, '%05d_Res.png' % (bt_idx_test)), np.uint8(output_image3))
# logger.info("======================== g1 results ============================")
avg_val_loss_g1 = sum_val_loss_g1 / len(testsplit)
avg_val_false_ratio_g1 = sum_val_false_ratio_g1 / len(testsplit)
avg_val_detect_ratio_g1 = sum_val_detect_ratio_g1 / len(testsplit)
avg_val_F1_g1 = sum_val_F1_g1 / len(testsplit)
summary_writer.add_scalar('valloss/g1', avg_val_loss_g1, epoch + 1)
summary_writer.add_scalar('false_alarm_rate/g1', avg_val_false_ratio_g1, epoch + 1)
summary_writer.add_scalar('detection_rate/g1', avg_val_detect_ratio_g1, epoch + 1)
summary_writer.add_scalar('F1_measure/g1', avg_val_F1_g1, epoch + 1)
# logger.info("======================== g2 results ============================")
avg_val_loss_g2 = sum_val_loss_g2 / len(testsplit)
avg_val_false_ratio_g2 = sum_val_false_ratio_g2 / len(testsplit)
avg_val_detect_ratio_g2 = sum_val_detect_ratio_g2 / len(testsplit)
avg_val_F1_g2 = sum_val_F1_g2 / len(testsplit)
summary_writer.add_scalar('valloss/g2', avg_val_loss_g2, epoch + 1)
summary_writer.add_scalar('false_alarm_rate/g2', avg_val_false_ratio_g2, epoch + 1)
summary_writer.add_scalar('detection_rate/g2', avg_val_detect_ratio_g2, epoch + 1)
summary_writer.add_scalar('F1_measure/g2', avg_val_F1_g2, epoch + 1)
# logger.info("======================== g3 results ============================")
avg_val_loss_g3 = sum_val_loss_g3 / len(testsplit)
avg_val_false_ratio_g3 = sum_val_false_ratio_g3 / len(testsplit)
avg_val_detect_ratio_g3 = sum_val_detect_ratio_g3 / len(testsplit)
avg_val_F1_g3 = sum_val_F1_g3 / len(testsplit)
summary_writer.add_scalar('valloss/g3', avg_val_loss_g3, epoch + 1)
summary_writer.add_scalar('false_alarm_rate/g3', avg_val_false_ratio_g3, epoch + 1)
summary_writer.add_scalar('detection_rate/g3', avg_val_detect_ratio_g3, epoch + 1)
summary_writer.add_scalar('F1_measure/g3', avg_val_F1_g3, epoch + 1)
print('current epoch {}/{}, total iteration: {}, g1 F1: {}, g2 F1: {}, g3 F1: {}'.format(
epoch + 1, config.max_epoch_num, it, avg_val_F1_g1, avg_val_F1_g2, avg_val_F1_g3))
############# save model
ckpt_name1 = os.path.join(model_result_dir, 'g1_epoch_{}_batch_{}'.format(epoch + 1, bt_idx + 1))
ckpt_name2 = os.path.join(model_result_dir, 'g2_epoch_{}_batch_{}'.format(epoch + 1, bt_idx + 1))
ckpt_name3 = os.path.join(model_result_dir, 'dis_epoch_{}_batch_{}'.format(epoch + 1, bt_idx + 1))
save_checkpoint(checkpoint_state(g1, optimizer_g1, epoch + 1, it), filename=ckpt_name1)
save_checkpoint(checkpoint_state(g2, optimizer_g2, epoch + 1, it), filename=ckpt_name2)
save_checkpoint(checkpoint_state(dis, optimizer_dis, epoch + 1, it), filename=ckpt_name3)
# scheduler_g1.step()
# scheduler_g2.step()
scheduler_dis.step()
if __name__ == '__main__':
train()