-
Notifications
You must be signed in to change notification settings - Fork 3
/
model_swinmr_stgan.py
390 lines (336 loc) · 15.5 KB
/
model_swinmr_stgan.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
'''
# -----------------------------------------
Model
ST-GAN m.1.3
by Jiahao Huang (j.huang21@imperial.ac.uk)
# -----------------------------------------
'''
from collections import OrderedDict
import torch
import torch.nn as nn
from torch.optim import lr_scheduler
from torch.optim import Adam
from models.select_network import define_G, define_D
from models.model_base import ModelBase
from models.loss import GANLoss, CharbonnierLoss, PerceptualLoss
from models.loss_ssim import SSIMLoss
from utils.utils_model import test_mode
from utils.utils_regularizers import regularizer_orth, regularizer_clip
from utils.utils_swinmr import *
import matplotlib.pyplot as plt
import einops
import wandb
from math import ceil
import copy
class MRI_STGAN(ModelBase):
def __init__(self, opt):
super(MRI_STGAN, self).__init__(opt)
# ------------------------------------
# define network
# ------------------------------------
self.opt_train = self.opt['train'] # training option
self.opt_dataset = self.opt['datasets']
self.netG = define_G(opt)
self.netG = self.model_to_device(self.netG)
if self.is_train:
self.netD = define_D(opt)
self.netD = self.model_to_device(self.netD)
if self.opt_train['E_decay'] > 0:
self.netE = define_G(opt).to(self.device).eval()
if opt['rank'] == 0:
wandb.watch(self.netG)
if self.is_train:
wandb.watch(self.netD)
"""
# ----------------------------------------
# Preparation before training with data
# Save model during training
# ----------------------------------------
"""
# ----------------------------------------
# initialize training
# ----------------------------------------
def init_train(self):
self.load() # load model
self.netG.train() # set training mode,for BN
self.netD.train() # set training mode,for BN
self.define_loss() # define loss
self.define_optimizer() # define optimizer
self.load_optimizers() # load optimizer
self.define_scheduler() # define scheduler
self.log_dict = OrderedDict() # log
# ----------------------------------------
# load pre-trained G D and E model
# ----------------------------------------
def load(self):
load_path_G = self.opt['path']['pretrained_netG']
if load_path_G is not None:
print('Loading model for G [{:s}] ...'.format(load_path_G))
self.load_network(load_path_G, self.netG, strict=self.opt_train['G_param_strict'])
load_path_E = self.opt['path']['pretrained_netE']
if self.opt_train['E_decay'] > 0:
if load_path_E is not None:
print('Loading model for E [{:s}] ...'.format(load_path_E))
self.load_network(load_path_E, self.netE, strict=self.opt_train['E_param_strict'])
else:
print('Copying model for E')
self.update_E(0)
self.netE.eval()
load_path_D = self.opt['path']['pretrained_netD']
if self.opt['is_train'] and load_path_D is not None:
print('Loading model for D [{:s}] ...'.format(load_path_D))
self.load_network(load_path_D, self.netD, strict=self.opt_train['D_param_strict'])
# ----------------------------------------
# load optimizerG and optimizerD
# ----------------------------------------
def load_optimizers(self):
load_path_optimizerG = self.opt['path']['pretrained_optimizerG']
if load_path_optimizerG is not None and self.opt_train['G_optimizer_reuse']:
print('Loading optimizerG [{:s}] ...'.format(load_path_optimizerG))
self.load_optimizer(load_path_optimizerG, self.G_optimizer)
load_path_optimizerD = self.opt['path']['pretrained_optimizerD']
if load_path_optimizerD is not None and self.opt_train['D_optimizer_reuse']:
print('Loading optimizerD [{:s}] ...'.format(load_path_optimizerD))
self.load_optimizer(load_path_optimizerD, self.D_optimizer)
# ----------------------------------------
# save model / optimizer(optional)
# ----------------------------------------
def save(self, iter_label):
self.save_network(self.save_dir, self.netG, 'G', iter_label)
self.save_network(self.save_dir, self.netD, 'D', iter_label)
if self.opt_train['E_decay'] > 0:
self.save_network(self.save_dir, self.netE, 'E', iter_label)
if self.opt_train['G_optimizer_reuse']:
self.save_optimizer(self.save_dir, self.G_optimizer, 'optimizerG', iter_label)
if self.opt_train['D_optimizer_reuse']:
self.save_optimizer(self.save_dir, self.D_optimizer, 'optimizerD', iter_label)
# ----------------------------------------
# define loss
# ----------------------------------------
def define_loss(self):
# ------------------------------------
# G_loss
# ------------------------------------
G_lossfn_type = self.opt_train['G_lossfn_type']
if G_lossfn_type == 'l1':
self.G_lossfn = nn.L1Loss().to(self.device)
elif G_lossfn_type == 'l2':
self.G_lossfn = nn.MSELoss().to(self.device)
elif G_lossfn_type == 'l2sum':
self.G_lossfn = nn.MSELoss(reduction='sum').to(self.device)
elif G_lossfn_type == 'ssim':
self.G_lossfn = SSIMLoss().to(self.device)
elif G_lossfn_type == 'charbonnier':
self.G_lossfn = CharbonnierLoss(self.opt_train['G_charbonnier_eps']).to(self.device)
else:
raise NotImplementedError('Loss type [{:s}] is not found.'.format(G_lossfn_type))
self.G_lossfn_weight = self.opt_train['G_lossfn_weight']
self.perceptual_lossfn = PerceptualLoss().to(self.device)
# ------------------------------------
# D_loss
# ------------------------------------
self.D_lossfn = GANLoss(self.opt_train['gan_type'], 1.0, 0.0).to(self.device)
self.D_lossfn_weight = self.opt_train['D_lossfn_weight']
self.D_update_ratio = self.opt_train['D_update_ratio'] if self.opt_train['D_update_ratio'] else 1
self.D_init_iters = self.opt_train['D_init_iters'] if self.opt_train['D_init_iters'] else 0
def total_loss(self):
self.alpha = self.opt_train['alpha']
self.beta = self.opt_train['beta']
self.gamma = self.opt_train['gamma']
# H HR, E Recon, L LR
self.H_k_real, self.H_k_imag = fft_map(self.H)
self.E_k_real, self.E_k_imag = fft_map(self.E)
self.loss_image = self.G_lossfn(self.E, self.H)
self.loss_freq = (self.G_lossfn(self.E_k_real, self.H_k_real) + self.G_lossfn(self.E_k_imag, self.H_k_imag)) / 2
self.loss_perc = self.perceptual_lossfn(self.E, self.H)
return self.alpha * self.loss_image + self.beta * self.loss_freq + self.gamma * self.loss_perc
# ----------------------------------------
# define optimizer for G and D
# ----------------------------------------
def define_optimizer(self):
G_optim_params = []
for k, v in self.netG.named_parameters():
if v.requires_grad:
G_optim_params.append(v)
else:
print('Params [{:s}] will not optimize.'.format(k))
self.G_optimizer = Adam(G_optim_params, lr=self.opt_train['G_optimizer_lr'], weight_decay=0)
self.D_optimizer = Adam(self.netD.parameters(), lr=self.opt_train['D_optimizer_lr'], weight_decay=0)
# ----------------------------------------
# define scheduler, only "MultiStepLR"
# ----------------------------------------
def define_scheduler(self):
self.schedulers.append(lr_scheduler.MultiStepLR(self.G_optimizer,
self.opt_train['G_scheduler_milestones'],
self.opt_train['G_scheduler_gamma']
))
self.schedulers.append(lr_scheduler.MultiStepLR(self.D_optimizer,
self.opt_train['D_scheduler_milestones'],
self.opt_train['D_scheduler_gamma']
))
"""
# ----------------------------------------
# Optimization during training with data
# Testing/evaluation
# ----------------------------------------
"""
# ----------------------------------------
# feed L/H data
# ----------------------------------------
def feed_data(self, data, need_H=True):
self.H = data['H'].to(self.device)
self.L = data['L'].to(self.device)
# self.mask = data['mask'].to(self.device)
# ----------------------------------------
# feed L to netG
# ----------------------------------------
def netG_forward(self):
self.E = self.netG(self.L)
# ----------------------------------------
# update parameters and get loss
# ----------------------------------------
def optimize_parameters(self, current_step):
# ------------------------------------
# optimize G
# ------------------------------------
for p in self.netD.parameters():
p.requires_grad = False
self.G_optimizer.zero_grad()
self.netG_forward()
if current_step % self.D_update_ratio == 0 and current_step > self.D_init_iters: # updata D first
pred_g_fake = self.netD(self.E)
self.loss_adversarial = self.D_lossfn(pred_g_fake, True)
loss_G_total = self.G_lossfn_weight * self.total_loss() + self.D_lossfn_weight * self.loss_adversarial
loss_G_total.backward()
self.G_optimizer.step()
# ------------------------------------
# optimize D
# ------------------------------------
for p in self.netD.parameters():
p.requires_grad = True
self.D_optimizer.zero_grad()
# real
pred_d_real = self.netD(self.H) # 1) real data
l_d_real = self.D_lossfn(pred_d_real, True)
l_d_real.backward()
# fake
pred_d_fake = self.netD(self.E.detach().clone()) # 2) fake data, detach to avoid BP to G
l_d_fake = self.D_lossfn(pred_d_fake, False)
l_d_fake.backward()
self.D_optimizer.step()
# ------------------------------------
# record log
# ------------------------------------
if current_step % self.D_update_ratio == 0 and current_step > self.D_init_iters:
self.log_dict['G_loss'] = loss_G_total.item()
self.log_dict['G_loss_image'] = self.loss_image.item()
self.log_dict['G_loss_frequency'] = self.loss_freq.item()
self.log_dict['G_loss_preceptual'] = self.loss_perc.item()
self.log_dict['G_loss_adversarial'] = self.loss_adversarial.item()
self.log_dict['D_loss_real'] = torch.mean(l_d_real.detach())
self.log_dict['D_loss_fake'] = torch.mean(l_d_fake.detach())
if self.opt_train['E_decay'] > 0:
self.update_E(self.opt_train['E_decay'])
def record_loss_for_val(self):
G_loss = self.G_lossfn_weight * self.total_loss()
self.log_dict['G_loss'] = G_loss.item()
self.log_dict['G_loss_image'] = self.loss_image.item()
self.log_dict['G_loss_frequency'] = self.loss_freq.item()
self.log_dict['G_loss_preceptual'] = self.loss_perc.item()
self.log_dict['G_loss_adversarial'] = self.loss_adversarial.item()
def check_windowsize(self):
self.window_size = self.opt['netG']['window_size']
_, _, h_old, w_old = self.H.size()
h_pad = ceil(h_old / self.window_size) * self.window_size - h_old # downsampling for 3 times (2^3=8)
w_pad = ceil(w_old / self.window_size) * self.window_size - w_old
self.h_old = h_old
self.w_old = w_old
self.H = torch.cat([self.H, torch.flip(self.H, [2])], 2)[:, :, :h_old + h_pad, :]
self.H = torch.cat([self.H, torch.flip(self.H, [3])], 3)[:, :, :, :w_old + w_pad]
self.L = torch.cat([self.L, torch.flip(self.L, [2])], 2)[:, :, :h_old + h_pad, :]
self.L = torch.cat([self.L, torch.flip(self.L, [3])], 3)[:, :, :, :w_old + w_pad]
def recover_windowsize(self):
self.L = self.L[..., :self.h_old, :self.w_old]
self.H = self.H[..., :self.h_old, :self.w_old]
self.E = self.E[..., :self.h_old, :self.w_old]
# ----------------------------------------
# test / inference
# ----------------------------------------
def test(self):
self.netG.eval()
with torch.no_grad():
self.netG_forward()
self.netG.train()
# ----------------------------------------
# get log_dict
# ----------------------------------------
def current_log(self):
return self.log_dict
# ----------------------------------------
# get L, E, H image
# ----------------------------------------
def current_visuals(self, need_H=True):
out_dict = OrderedDict()
out_dict['L'] = self.L.detach()[0].float().cpu()
out_dict['E'] = self.E.detach()[0].float().cpu()
if need_H:
out_dict['H'] = self.H.detach()[0].float().cpu()
return out_dict
def current_visuals_gpu(self, need_H=True):
out_dict = OrderedDict()
out_dict['L'] = self.L.detach()[0].float()
out_dict['E'] = self.E.detach()[0].float()
if need_H:
out_dict['H'] = self.H.detach()[0].float()
return out_dict
# ----------------------------------------
# get L, E, H batch images
# ----------------------------------------
def current_results(self, need_H=True):
out_dict = OrderedDict()
out_dict['L'] = self.L.detach().float().cpu()
out_dict['E'] = self.E.detach().float().cpu()
if need_H:
out_dict['H'] = self.H.detach().float().cpu()
return out_dict
def current_results_gpu(self, need_H=True):
out_dict = OrderedDict()
out_dict['L'] = self.L.detach().float()
out_dict['E'] = self.E.detach().float()
if need_H:
out_dict['H'] = self.H.detach().float()
return out_dict
"""
# ----------------------------------------
# Information of netG, netD
# ----------------------------------------
"""
# ----------------------------------------
# print network
# ----------------------------------------
def print_network(self):
msg = self.describe_network(self.netG)
print(msg)
if self.is_train:
msg = self.describe_network(self.netD)
print(msg)
# ----------------------------------------
# print params
# ----------------------------------------
def print_params(self):
msg = self.describe_params(self.netG)
print(msg)
# ----------------------------------------
# network information
# ----------------------------------------
def info_network(self):
msg = self.describe_network(self.netG)
if self.is_train:
msg += self.describe_network(self.netD)
return msg
# ----------------------------------------
# params information
# ----------------------------------------
def info_params(self):
msg = self.describe_params(self.netG)
return msg