-
Notifications
You must be signed in to change notification settings - Fork 9
/
trainer_stackgan.py
435 lines (365 loc) · 20.1 KB
/
trainer_stackgan.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
import torch
import json
from models.StackGAN import *
from models.Encoder import baseEncoder,baseEncoderv2,harchyEncoder
from models.tools import cal_gradient_penalty, init_weights
from torch import nn
from utils.MIMICDataSet import MIMICDataset2
from utils.OpeniDataSet import OpeniDataset2
from torch.utils.data import DataLoader, random_split
from torchvision import transforms
from utils import get_time, matplotlib_imshow, deNorm, Rescale, ToTensor, Equalize
from tqdm import tqdm
from torch.utils.tensorboard import SummaryWriter
import torchvision
import os
import numpy as np
import torch.nn.functional as F
from torch.optim.lr_scheduler import MultiStepLR
def weights_init(m):
classname = m.__class__.__name__
if classname.find('Conv') != -1:
nn.init.orthogonal(m.weight.data, 1.0)
elif classname.find('BatchNorm') != -1:
m.weight.data.normal_(1.0, 0.02)
m.bias.data.fill_(0)
elif classname.find('Linear') != -1:
nn.init.orthogonal(m.weight.data, 1.0)
if m.bias is not None:
m.bias.data.fill_(0.0)
class Trainer:
def __init__(self):
self.cfg_json = "config/MIMIC_StackGAN.json"
self.cfg = self.pare_cfg(self.cfg_json)
self.exp_name = self.cfg["EXPER_NAME"]
self.max_epoch = self.cfg["MAX_EPOCH"]
self.encoder_checkpoint = self.cfg["CHECKPOINT_ENCODER"]
self.decoder_checkpoint = self.cfg["CHECKPOINT_DECODER"]
self.D_checkpoint = self.cfg["CHECKPOINT_D"]
self.check_create_checkpoint()
self.encoder_resume = self.cfg["RESUME_ENCODER"]
self.decoder_resume_F = self.cfg["RESUME_DECODER_F"]
self.decoder_resume_L = self.cfg["RESUME_DECODER_L"]
self.D_resume_F,self.D_resume_L = self.cfg["RESUME_D"]
self.train_csv = self.cfg["TRAIN_CSV"]
self.val_csv = self.cfg["VAL_CSV"]
self.test_csv = self.cfg["TEST_CSV"]
self.text_csv = self.cfg["TEXT_CSV"]
self.img_csv = self.cfg["IMG_CSV"]
self.data_root = self.cfg["DATA_ROOT"]
self.batch_size = self.cfg["BATCH_SIZE"]
self.image_size = tuple(self.cfg["IMAGE_SIZE"])
self.name = self.cfg["EXPER_NAME"]
self.pix_loss_ratio = self.cfg["PIXEL_LOSS_RATIO"]
self.adv_loss_ratio = self.cfg["ADV_LOSS_RATIO"]
self.checkpoint_epoch = self.cfg["CHECKPOINT_EPOCH"]
self.beta1 = self.cfg["beta1"]
self.word_dict = self.cfg["DICTIONARY"]
self.writer = SummaryWriter(os.path.join("runs", self.exp_name))
self.dataset = {
"OPENI": OpeniDataset2,
"MIMIC-CXR": MIMICDataset2
}
##################################################
################# Dataset Setup ##################
##################################################
self.t2i_dataset = self.dataset[self.cfg["DATASET"]](csv_txt=self.text_csv,
csv_img=self.img_csv,
root=self.data_root,
word_dict=self.word_dict,
transform=transforms.Compose([
Rescale(self.image_size),
Equalize(),
ToTensor()
]))
self.trainset = self.dataset[self.cfg["DATASET"]][0](csv_txt=self.train_csv,
csv_img=self.img_csv,
root=self.data_root,
word_dict=self.word_dict,
transform=transforms.Compose([
Rescale(self.image_size),
Equalize(),
ToTensor()
]))
self.valset = self.dataset[self.cfg["DATASET"]][0](csv_txt=self.val_csv,
csv_img=self.img_csv,
root=self.data_root,
word_dict=self.word_dict,
transform=transforms.Compose([
Rescale(self.image_size),
Equalize(),
ToTensor()
]))
self.device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
s_gpus = self.cfg["GPU_ID"].split(',')
self.gpus = [int(ix) for ix in s_gpus]
self.num_gpus = len(self.gpus)
#########################################
############ Loss Function ##############
#########################################
content_losses = {"L2": nn.MSELoss(),
"L1": nn.L1Loss()}
self.G_criterion = content_losses[self.cfg["CONTENT_LOSS"]].to(self.device)
#########################################
############ Network Init ###############
#########################################
self.decoder_L,self.D_L= self.define_nets()
self.decoder_F,self.D_F= self.define_nets()
self.encoder = harchyEncoder(vocab_size=self.t2i_dataset.vocab_size,
embed_size=self.cfg["E_EMBED_SIZE"],
hidden_size=self.cfg["E_HIDEN_SIZE"],
max_len=[self.t2i_dataset.max_len_finding,
self.t2i_dataset.max_len_impression],
unit=self.cfg["RNN_CELL"],
feature_base_dim=self.cfg["D_CHANNEL_SIZE"]
).to(self.device)
self.encoder = nn.DataParallel(self.encoder, device_ids=self.gpus)
def define_nets(self):
netG = G_NET().to(self.device)
# netG.apply(weights_init)
netG = nn.DataParallel(netG, device_ids=self.gpus).to(self.device)
print(netG)
netsD = []
netsD.append(D_NET64().to(self.device))
netsD.append(D_NET128().to(self.device))
netsD.append(D_NET256().to(self.device))
for i in range(len(netsD)):
# netsD[i].apply(weights_init)
netsD[i] = torch.nn.DataParallel(netsD[i], device_ids=self.gpus).to(self.device)
# print(netsD[i])
print('# of netsD', len(netsD))
return netG,netsD
def define_dataloader(self):
self.train_dataloader = DataLoader(self.trainset,
batch_size=self.batch_size,
shuffle=True,
num_workers=8,
drop_last=True)
self.val_dataloader = DataLoader(self.valset,
batch_size=self.batch_size,
shuffle=True,
num_workers=8,
drop_last=True)
def define_opt(self, only_G=False):
'''Define optimizer'''
self.G_optimizer = torch.optim.Adam([{'params': self.encoder.parameters()}] +
[{'params': self.decoder_F.parameters()}] +
[{'params': self.decoder_L.parameters()}],
lr=self.cfg["G_LR"], betas=(self.beta1, 0.999))
self.G_lr_scheduler = MultiStepLR(self.G_optimizer, milestones=self.cfg["LR_DECAY_EPOCH"], gamma=0.2)
self.D_optimizer = torch.optim.Adam([{'params': D.parameters()} for D in self.D_F] +
[{'params': D.parameters()}for D in self.D_L]
, lr=self.cfg["D_LR"], betas=(self.beta1, 0.999))
self.D_lr_scheduler = MultiStepLR(self.D_optimizer, milestones=self.cfg["LR_DECAY_EPOCH"], gamma=0.2)
def check_create_checkpoint(self):
'''Check for the checkpoint path exists or not
If not exist, create folder'''
if os.path.exists(self.encoder_checkpoint) == False:
os.makedirs(self.encoder_checkpoint)
if os.path.exists(self.decoder_checkpoint) == False:
os.makedirs(self.decoder_checkpoint)
if os.path.exists(self.D_checkpoint) == False:
os.makedirs(self.D_checkpoint)
def load_model(self):
print("Model Loading....................")
self.encoder.load_state_dict(torch.load(self.encoder_resume))
self.decoder_F.load_state_dict(torch.load(self.decoder_resume_F))
self.decoder_L.load_state_dict(torch.load(self.decoder_resume_L))
for i in range(3):
self.D_L[i].load_state_dict(torch.load(self.D_resume_L[i]))
self.D_F[i].load_state_dict(torch.load(self.D_resume_F[i]))
def get_lr(self, optimizer):
for param_group in optimizer.param_groups:
return param_group['lr']
def Loss_on_layer(self, image, finding, impression, decoder):
'''
Pretrain genertaor with batch
:image image batch
:text text batch
'''
image1 = F.interpolate(image, size=(64,64))
image2 = F.interpolate(image, size=(128,128))
image3 = F.interpolate(image, size=(256,256))
real_img = [image1, image2, image3]
txt_emded, hidden = self.encoder(finding, impression)
self.G_optimizer.zero_grad()
pre_image,mu, logvar = decoder(txt_emded)
total_loss = 0
for i in range(3):
loss = self.G_criterion(pre_image[i].float(), real_img[i].float())
loss.backward(retain_graph=True)
total_loss+=loss
self.G_optimizer.step()
return total_loss, pre_image[-1], real_img[-1]
def Loss_on_layer_GAN(self, image, finding, impression, decoder, D):
'''
Pretrain genertaor with batch
:image image batch
:text text batch
'''
image1 = F.interpolate(image, size=(64, 64))
image2 = F.interpolate(image, size=(128, 128))
image3 = F.interpolate(image, size=(256, 256))
real_img = [image1,image2,image3]
txt_emded, hidden = self.encoder(finding, impression)
pre_image, mu, logvar = decoder(txt_emded)
# Train Discriminator
self.D_optimizer.zero_grad()
total_D_loss = 0
for i in range(3):
pre_fake = D[i](pre_image[i], mu.detach())
pre_real = D[i](real_img[i], mu.detach())
gradient_penalty, gradients = cal_gradient_penalty(D[i], real_img[i], pre_image[i], mu.detach(), "cuda")
D_loss = pre_fake.mean() - pre_real.mean() + gradient_penalty
total_D_loss += D_loss
D_loss.backward(retain_graph=True)
self.D_optimizer.step()
# Train Generator
self.G_optimizer.zero_grad()
G_loss = 0
for i in range(3):
pre_fake = D[i](pre_image[i], mu.detach())
adv_loss = - self.adv_loss_ratio * pre_fake.mean()
adv_loss.backward(retain_graph=True)
content_loss = self.pix_loss_ratio * self.G_criterion(pre_image[i].float(),
real_img[i].float())
content_loss.backward(retain_graph=True)
G_loss += content_loss
G_loss += adv_loss
self.G_optimizer.step()
return total_D_loss, G_loss, pre_image[-1], real_img[-1]
def train_layer(self):
DISP_FREQ = 10
for epoch in range(20):
print('Generator Epoch {}'.format(epoch))
self.encoder.train()
self.decoder_F.train()
self.decoder_L.train()
for idx, batch in enumerate(tqdm(self.train_dataloader)):
finding = batch['finding'].to(self.device)
impression = batch['impression'].to(self.device)
image_f = batch['image_F'].to(self.device)
image_l = batch['image_L'].to(self.device)
loss_f, pre_image_f, r_image_f = self.Loss_on_layer(image_f, finding, impression,
self.decoder_F)
loss_l, pre_image_l, r_image_l = self.Loss_on_layer(image_l, finding, impression,
self.decoder_L)
# print('Loss: {:.4f}'.format(loss.item()))
if ((idx + 1) % DISP_FREQ == 0) and idx != 0:
self.writer.add_scalar('Train_front loss',
loss_f.item(),
epoch * len(self.train_dataloader) + idx)
self.writer.add_scalar('Train_lateral loss',
loss_l.item(),
epoch * len(self.train_dataloader) + idx)
# write to tensorboard
self.writer.add_images("Train_front_Original",
deNorm(r_image_f),
epoch * len(self.train_dataloader) + idx)
self.writer.add_images("Train_front_Predicted",
deNorm(pre_image_f),
epoch * len(self.train_dataloader) + idx)
self.writer.add_images("Train_lateral_Original",
deNorm(r_image_l),
epoch * len(self.train_dataloader) + idx)
self.writer.add_images("Train_lateral_Predicted",
deNorm(pre_image_l),
epoch * len(self.train_dataloader) + idx)
self.G_lr_scheduler.step(epoch)
def set_D(self,Ds, training = True):
if training:
for D in Ds:
D.train()
else:
for D in Ds:
D.eval()
def train_GAN_layer(self):
DISP_FREQ = 10
self.encoder.train()
self.decoder_F.train()
self.decoder_L.train()
self.set_D(self.D_F)
self.set_D(self.D_L)
for epoch in range(self.max_epoch):
print('GAN Epoch {}'.format(epoch))
for idx, batch in enumerate(tqdm(self.train_dataloader)):
finding = batch['finding'].to(self.device)
impression = batch['impression'].to(self.device)
image_f = batch['image_F'].to(self.device)
image_l = batch['image_L'].to(self.device)
D_loss_f, G_loss_f, pre_image_f, image_f = self.Loss_on_layer_GAN(image_f, finding, impression,self.decoder_F, self.D_F)
D_loss_l, G_loss_l, pre_image_l, image_l = self.Loss_on_layer_GAN(image_l, finding, impression,self.decoder_L, self.D_L)
if ((idx + 1) % DISP_FREQ == 0) and idx != 0:
# ...log the running loss
# self.writer.add_scalar("Train_{}_SSIM".format(layer_id), ssim.ssim(r_image, pre_image).item(),
# epoch * len(self.train_dataloader) + idx)
self.writer.add_scalar('GAN_G_train_Layer_front_loss',
G_loss_f.item(),
epoch * len(self.train_dataloader) + idx)
self.writer.add_scalar('GAN_D_train_Layer_front_loss',
D_loss_f.item(),
epoch * len(self.train_dataloader) + idx)
self.writer.add_scalar('GAN_G_train_Layer_lateral_loss',
G_loss_l.item(),
epoch * len(self.train_dataloader) + idx)
self.writer.add_scalar('GAN_D_train_Layer_lateral_loss',
D_loss_l.item(),
epoch * len(self.train_dataloader) + idx)
# write to tensorboard
self.writer.add_images("GAN_Train_Original_front",
deNorm(image_f),
epoch * len(self.train_dataloader) + idx)
self.writer.add_images("GAN_Train_Predicted_front",
deNorm(pre_image_f),
epoch * len(self.train_dataloader) + idx)
self.writer.add_images("GAN_Train_Original_lateral",
deNorm(image_l),
epoch * len(self.train_dataloader) + idx)
self.writer.add_images("GAN_Train_Predicted_lateral",
deNorm(pre_image_l),
epoch * len(self.train_dataloader) + idx)
self.G_lr_scheduler.step(epoch)
self.D_lr_scheduler.step(epoch)
if epoch%10==0 and epoch!=0:
torch.save(self.encoder.state_dict(), os.path.join(self.encoder_checkpoint,
"Encoder_{}_epoch_{}_checkpoint.pth".format(
self.cfg["ENCODER"],
epoch)))
for i in range(3):
torch.save(self.D_F[i].state_dict(), os.path.join(self.D_checkpoint,
"D_{}_F_epoch_{}_layer_{}_checkpoint.pth".format(
self.cfg["DISCRIMINATOR"],epoch,i)))
torch.save(self.D_L[i].state_dict(), os.path.join(self.D_checkpoint,
"D_{}_L_epoch_{}_layer_{}_checkpoint.pth".format(
self.cfg["DISCRIMINATOR"], epoch,i)))
torch.save(self.decoder_F.state_dict(), os.path.join(self.decoder_checkpoint,
"Decoder_{}_F_epoch_{}_checkpoint.pth".format(
self.cfg["DECODER"], epoch)))
torch.save(self.decoder_L.state_dict(), os.path.join(self.decoder_checkpoint,
"Decoder_{}_L_epoch_{}_checkpoint.pth".format(
self.cfg["DECODER"], epoch)))
def train(self):
# self.load_model()
self.define_opt()
self.define_dataloader()
#########################################################
############### Train Generator by layer ################
#########################################################
print("Start training on Decoder")
# self.train_layer()
#########################################################
################## Train GAN by layer ###################
#########################################################
print("Start training GAN")
self.train_GAN_layer()
def pare_cfg(self, cfg_json):
with open(cfg_json) as f:
cfg = f.read()
print(cfg)
print("Config Loaded")
return json.loads(cfg)
def main():
trainer = Trainer()
trainer.train()
if __name__ == "__main__":
main()