-
Notifications
You must be signed in to change notification settings - Fork 0
/
train.py
265 lines (197 loc) · 10.8 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
'''
Created on Jul 27, 2021
@author: Quang Tran
'''
import argparse
import os
import time
import torch
from torch import nn, optim, cuda
from torchvision.utils import save_image
from torch.optim import lr_scheduler
from nets import net
from loss import GDL, MS_SSIM
from utils import CONSTANT, utils, model_utils
from observers import checkpoint_observer, cpt_test_observer
Tensor = torch.cuda.FloatTensor if torch.cuda.is_available() else torch.FloatTensor
parser = argparse.ArgumentParser()
parser.add_argument("--n_epochs", type=int, default=20, help="number of epochs of training")
parser.add_argument("--batch_size", type=int, default=32, help="size of the batches")
parser.add_argument("--patch_size", type=int, default=3, help="size of the patch of sample")
parser.add_argument("--input_size", type=int, default=128, help="size of the input of network")
parser.add_argument("--nfg", type=int, default=32, help="feature map size of networks")
parser.add_argument("--lr", type=float, default=0.001, help="adam: learning rate")
parser.add_argument("--b1", type=float, default=0.9, help="adam: decay of first order momentum of gradient")
parser.add_argument("--b2", type=float, default=0.999, help="adam: decay of first order momentum of gradient")
parser.add_argument("--lr_decay", type=int, default=20, help="adam: learning rate decay step")
parser.add_argument("--channels", type=int, default=3, help="number of image channels")
parser.add_argument("--sample_interval", type=int, default=500, help="number of batches between image sampling")
parser.add_argument("--lambda_adv", type=float, default=0.0001, help="the default weight of adv Loss")
parser.add_argument("--lambda_px_loss", type=float, default=1.0, help="the default weight of L2 Loss")
parser.add_argument("--lambda_gdl", type=float, default=1.0, help="the default weight of GDL Loss")
parser.add_argument("--lambda_ms_ssim", type=float, default=1.0, help="the default weight of MS-SSIM Loss")
parser.add_argument("--path_gen", type=str, default=None, help="loaded generator for training")
parser.add_argument("--path_dis", type=str, default=None, help="loaded discriminator for training")
parser.add_argument("--path", type=str, default="output", help="training folder")
parser.add_argument("--is_testing", action='store_true')
parser.add_argument("--debug_mode", action='store_false')
opt = parser.parse_args()
assert not opt.is_testing
# networks
net_gen = net.FrameInterpolationGenerator(nfg=opt.nfg)
net_dis = net.FrameInterpolationDiscriminator(nfg=opt.nfg)
optimizer_G = optim.Adam(net_gen.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
optimizer_D = optim.Adam(net_dis.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
# loss functions
adversarial_loss = nn.BCEWithLogitsLoss().cuda()
l2_loss = nn.MSELoss().cuda()
gd_loss = GDL.GDL(cudaUsed=True).cuda()
ms_ssim = MS_SSIM.MS_SSIM_Loss().cuda()
# optimizers
optimizer_G = optim.Adam(net_gen.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
optimizer_D = optim.Adam(net_dis.parameters(), lr=opt.lr, betas=(opt.b1, opt.b2))
def load_generator(path=None):
if path is None:
net_gen.apply(model_utils._weights_init_normal)
return None, 0, -1
model_state_dict, optim_state_dict, epoch, version, loss = model_utils.load_model(path)
net_gen.load_state_dict(model_state_dict)
return optim_state_dict, epoch, loss
# end load_generator
def load_discriminator(path=None):
if path is None:
net_dis.apply(model_utils._weights_init_normal)
return None, 0, -1
model_state_dict, optim_state_dict, epoch, version, loss = model_utils.load_model(path)
net_dis.load_state_dict(model_state_dict)
return optim_state_dict, epoch, loss
# end load_discriminator
def _train_interval(in_pres, in_lats, gt):
# Adversarial ground truths
valid = Tensor(gt.shape[0], 1, 1, 1).fill_(0.95)
fake = Tensor(gt.shape[0], 1, 1, 1).fill_(0.1)
valid.requires_grad_(False)
fake.requires_grad_(False)
_, gen_imgs = net_gen(in_pres, in_lats) # generate output images
# ---------------------
# Train Discriminator
# ---------------------
for param in net_dis.parameters():
param.grad = None
# Calculate gradient for D
gt_distingue = net_dis(in_pres, gt, in_lats)
fake_distingue = net_dis(in_pres, gen_imgs.detach(), in_lats)
real_loss = adversarial_loss(gt_distingue, valid)
fake_loss = adversarial_loss(fake_distingue, fake)
d_loss = (real_loss + fake_loss) / 2
d_loss.backward()
optimizer_D.step() # update D's weights
# -----------------
# Train Generator
# -----------------
for param in net_gen.parameters():
param.grad = None
# Calculate gradient for G
# Loss measures generator's ability to fool the discriminator and generate similar image to ground truth
adv_loss = opt.lambda_adv * adversarial_loss(net_dis(in_pres, gen_imgs, in_lats), valid)
rec_loss = opt.lambda_px_loss * l2_loss(gen_imgs, gt) + opt.lambda_gdl * gd_loss(gen_imgs, gt) + opt.lambda_ms_ssim * ms_ssim(gen_imgs, gt)
g_loss = adv_loss + rec_loss
g_loss.backward()
optimizer_G.step() # update G's weights
return gen_imgs, real_loss.item(), fake_loss.item(), g_loss.item(), adv_loss.item()
# end _train_interval
def _show_progress(gen_imgs, groundtruth, epoch, batch_index, total_batch, d_real_loss, d_fake_loss, g_loss, g_adv_loss):
psnr = utils.cal_psnr_tensor(gen_imgs.data[0].cpu(), groundtruth.data[0].cpu())
log = ("%s: [Epoch %d] [Batch %d/%d] [D loss (real/fake): (%1.5f, %1.5f)] [G loss (adv_loss): %1.5f (%1.5f)] [PSNR: %1.2f]"
% ("FI-DUSGAN", epoch, batch_index, total_batch, d_real_loss, d_fake_loss, g_loss, g_adv_loss, psnr))
# Display result (input and output) after every sample_intervals
batches_done = epoch * total_batch + batch_index
if batches_done % (opt.sample_interval * (epoch // 10 + 1)) == 0:
save_image(gen_imgs.data[:16], opt.path + "/train_%d.png" % (batches_done), nrow=4, normalize=True)
print("Saved train_%d.png" % batches_done)
print(log)
elif batches_done % 1000 == 0:
print(log)
return "%d\t%d/%d\t%f\t%f\t%f" % (epoch, batch_index, total_batch, (d_real_loss + d_fake_loss) / 2, g_loss, psnr)
# end _show_progress
def train(epoch, dataloader, test_dataloader):
cpt_observer = checkpoint_observer.CheckPointObserver(opt.path)
cpt_testing_observer = cpt_test_observer.CheckPointTest_Observer(opt.path + "/cpt/", test_dataloader)
scheduler_G = lr_scheduler.StepLR(optimizer_G, step_size=max(opt.lr_decay, opt.n_epochs // 10), gamma=0.1)
scheduler_D = lr_scheduler.StepLR(optimizer_D, step_size=max(opt.lr_decay, opt.n_epochs // 10), gamma=0.1)
# loss functions
g_loss = 0
d_loss = 0
utils.logging(opt.path, "[Epoch]\t[Batch]\t[Total_batch]\t[D loss]\t[G loss]\t[PSNR]")
while (epoch < opt.n_epochs):
# enable training mode
net_gen.train()
net_dis.train()
temp_log = ""
for i, imgs in enumerate(dataloader):
in_pres = imgs[0].to('cuda')
in_lats = imgs[2].to('cuda')
gt = imgs[1].to('cuda')
gen_imgs, real_loss, fake_loss, g_loss, adv_loss = _train_interval(in_pres, in_lats, gt)
d_loss = (real_loss + fake_loss) / 2
# Show progress
if i > 0: temp_log += "\n"
temp_log = _show_progress(gen_imgs, gt, epoch, i, dataloader.__len__(), real_loss, fake_loss, g_loss, adv_loss)
utils.logging(opt.path, temp_log)
cpt_observer.notify(net_gen.state_dict(), net_dis.state_dict(), optimizer_G.state_dict(), optimizer_D.state_dict(), epoch + 1, g_loss, d_loss)
epoch += 1
if scheduler_G.get_lr()[0] > 10 ** -7:
scheduler_G.step()
scheduler_D.step()
# end if
cpt_testing_observer.notify(net_gen, epoch, opt)
# end while
return epoch, g_loss, d_loss
# end train
def main():
if not cuda.is_available():
print(CONSTANT.MESS_NO_CUDA)
return
data_dir = "data/tri_trainlist.txt"
val_dir = "data/tri_vallist.txt"
path = opt.path
print(opt)
os.makedirs(opt.path, exist_ok=True);
print("==============<Prepare models/>=============================")
# init training parameters and information
gen_optim_state, cur_gen_epoch, gen_loss = load_generator(opt.path_gen)
dis_optim_state, cur_dis_epoch, dis_loss = load_discriminator(opt.path_dis)
net_gen.to('cuda')
net_dis.to('cuda')
if gen_optim_state is not None:
optimizer_G.load_state_dict(gen_optim_state)
# copy tensor into GPU manually
for state in optimizer_G.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.cuda()
if dis_optim_state is not None:
optimizer_D.load_state_dict(dis_optim_state)
# copy tensor into GPU manually
for state in optimizer_D.state.values():
for k, v in state.items():
if torch.is_tensor(v):
state[k] = v.cuda()
print("==> Models are ready at %d epoch." % cur_gen_epoch)
print("==============<Prepare dataset/>=============================")
t1 = time.time()
dataloader = utils.load_dataset_from_path_file(batch_size=opt.batch_size, txt_path=data_dir, is_testing=False, patch_size=opt.patch_size)
test_dataloader = utils.load_dataset_from_path_file(batch_size=1, txt_path=val_dir, is_testing=True, patch_size=opt.patch_size)
print("==> Takes total %1.4fs" % ((time.time() - t1)))
print("==============<Training.../>====================================")
os.makedirs(path, exist_ok=True);
log = "Opts[n_epochs:%d, batch_size:%d, input_size:%d, nfg: %d lr:%f, path_gen:%s, path_dis:%s]" % (opt.n_epochs, opt.batch_size, opt.input_size, opt.nfg, opt.lr, opt.path_gen, opt.path_dis)
utils.logging(path, log, is_exist=False)
t1 = time.time()
epoch, g_loss, d_loss = train(cur_gen_epoch, dataloader, test_dataloader)
print("==> Takes total %1.2fmins" % ((time.time() - t1) / (60)))
utils.logging(path, "Training takes %1.2fmins" % ((time.time() - t1) / (60)), is_exist=True)
model_utils.save_models(net_gen.state_dict(), net_dis.state_dict(), optimizer_G.state_dict(), optimizer_D.state_dict(), epoch, path, True, g_loss, d_loss)
main()
if __name__ == '__main__':
pass