/
train.py
228 lines (186 loc) · 8.69 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
import os
import time
import datetime
import itertools
import argparse
import torch
import torch.nn as nn
import torch.optim as optim
from torch.utils.data import DataLoader
from torchvision.utils import save_image, make_grid
import model
from model import Generator, Discriminator
from datasets import ImageDataset, FakeImageBuffer
if torch.cuda.is_available():
device = 'cuda'
else:
device = 'cpu'
parser = argparse.ArgumentParser()
parser.add_argument("--start_from", type=int, default=0, help="epoch number to start from; 0 for training from scratch")
parser.add_argument("--num_epochs", type=int, default=200, help="number of epochs")
parser.add_argument("--dataset_name", type=str, default="horse2zebra", help="name of the dataset")
parser.add_argument("--batch_size", type=int, default=10, help="number of samples in batch")
parser.add_argument("--img_height", type=int, default=256, help="image height in pixels")
parser.add_argument("--img_width", type=int, default=256, help="image width in pixels")
parser.add_argument("--channels", type=int, default=3, help="number of image channels")
parser.add_argument("--num_residual", type=int, default=9, help="number of residual blocks")
parser.add_argument("--lambda_cycle", type=float, default=10.0, help="cycle loss weight")
args = parser.parse_args()
print(args)
#############################
# Dataloaders; fake buffers #
#############################
# TODO: make path_to_data exactly path_to_data.
# Now script requires to be launched only from location of *dataset_name* dir.
train_loader = DataLoader(
ImageDataset(path_to_data='datasets/%s' % args.dataset_name, size=(args.img_height,args.img_width), mode='train'),
batch_size=args.batch_size,
num_workers=2,
shuffle=True
)
test_loader = DataLoader(
ImageDataset(path_to_data='datasets/%s' % args.dataset_name, size=(args.img_height,args.img_width), mode='test'),
batch_size=5,
num_workers=2,
shuffle=True
)
fake_A_buffer = FakeImageBuffer()
fake_B_buffer = FakeImageBuffer()
##########
# Models #
##########
G_AB = Generator(in_channels=args.channels, num_residual=args.num_residual).to(device=device)
G_BA = Generator(in_channels=args.channels, num_residual=args.num_residual).to(device=device)
D_A = Discriminator().to(device=device)
D_B = Discriminator().to(device=device)
if args.start_from == 0:
G_AB.apply(model.init_weights_func)
G_BA.apply(model.init_weights_func)
D_A.apply(model.init_weights_func)
D_B.apply(model.init_weights_func)
else:
G_AB.load_state_dict(torch.load("saved_models/%s/G_AB_%d.pth" % (args.dataset_name, args.start_from)))
G_BA.load_state_dict(torch.load("saved_models/%s/G_BA_%d.pth" % (args.dataset_name, args.start_from)))
D_A.load_state_dict(torch.load("saved_models/%s/D_A_%d.pth" % (args.dataset_name, args.start_from)))
D_B.load_state_dict(torch.load("saved_models/%s/D_B_%d.pth" % (args.dataset_name, args.start_from)))
##############
# Criterions #
##############
criterion_adv = nn.MSELoss()
criterion_cycle = nn.L1Loss()
lambda_cycle = args.lambda_cycle
##############
# Optimizers #
##############
G_optimizer = torch.optim.Adam(itertools.chain(G_AB.parameters(), G_BA.parameters()), 2e-4)
D_A_optimizer = torch.optim.Adam(D_A.parameters(), 2e-4)
D_B_optimizer = torch.optim.Adam(D_B.parameters(), 2e-4)
lr_scheduler_G = torch.optim.lr_scheduler.LambdaLR(G_optimizer, lr_lambda=lambda epoch: 2 - epoch/100)
lr_scheduler_D_A = torch.optim.lr_scheduler.LambdaLR(D_A_optimizer, lr_lambda=lambda epoch: 2 - epoch/100)
lr_scheduler_D_B = torch.optim.lr_scheduler.LambdaLR(D_B_optimizer, lr_lambda=lambda epoch: 2 - epoch/100)
############################################
# Generating and saving images #
# From https://github.com/eriklindernoren/ #
############################################
def sample_images(batches_done):
"""Saves a generated sample from the test set"""
imgs = next(iter(test_loader))
G_AB.eval()
G_BA.eval()
real_A = imgs["A"].permute(0, 3, 1, 2).to(device=device).float()
fake_B = G_AB(real_A)
real_B = imgs["B"].permute(0, 3, 1, 2).to(device=device).float()
fake_A = G_BA(real_B)
# Arange images along x-axis
real_A = make_grid(real_A, nrow=5, normalize=True)
real_B = make_grid(real_B, nrow=5, normalize=True)
fake_A = make_grid(fake_A, nrow=5, normalize=True)
fake_B = make_grid(fake_B, nrow=5, normalize=True)
# Arange images along y-axis
image_grid = torch.cat((real_A, fake_B, real_B, fake_A), 1)
save_image(image_grid, "generated_images/%s/%s.png" % (args.dataset_name, batches_done), normalize=False)
############
# Training #
############
batch_time = time.time()
for epoch in range(args.start_from, args.num_epochs):
for i, batch in enumerate(train_loader):
real_A = batch['A'].permute(0, 3, 1, 2).to(device=device).float()
real_B = batch['B'].permute(0, 3, 1, 2).to(device=device).float()
#######################
# Training generators #
#######################
G_AB.train()
G_BA.train()
G_optimizer.zero_grad()
fake_B = G_AB(real_A)
fake_A = G_BA(real_B)
rec_A = G_BA(fake_B)
rec_B = G_AB(fake_A)
# Adversarial Loss
out_DA = D_A(fake_A) # TODO: need to avoid this; but how to get output shape of discriminator?
out_DB = D_B(fake_B) # TODO: need to avoid this; but how to get output shape of discriminator?
loss_adv_AB = criterion_adv(out_DB, torch.ones(out_DB.shape, device=device))
loss_adv_BA = criterion_adv(out_DA, torch.ones(out_DA.shape, device=device))
loss_adv = (loss_adv_AB + loss_adv_BA) / 2
# Cycle Loss
loss_cycle_A = criterion_cycle(real_A, rec_A)
loss_cycle_B = criterion_cycle(real_B, rec_B)
loss_cycle = (loss_cycle_A + loss_cycle_B) / 2
# Total generator loss
loss_gen = loss_adv + lambda_cycle * loss_cycle
loss_gen.backward()
G_optimizer.step()
###########################
# Training discriminators #
###########################
D_A_optimizer.zero_grad()
preds_real = D_A(real_A)
loss_real = criterion_adv(preds_real, torch.ones(preds_real.shape, device=device))
fake = fake_A_buffer.push_and_pop(fake_A)
preds_fake = D_A(fake.detach())
loss_fake = criterion_adv(preds_fake, torch.zeros(preds_fake.shape, device=device))
loss_D_A = (loss_real + loss_fake) / 2
loss_D_A.backward()
D_A_optimizer.step()
D_B_optimizer.zero_grad()
preds_real = D_B(real_B)
loss_real = criterion_adv(preds_real, torch.ones(preds_real.shape, device=device))
fake = fake_B_buffer.push_and_pop(fake_B)
preds_fake = D_B(fake.detach())
loss_fake = criterion_adv(preds_fake, torch.zeros(preds_fake.shape, device=device))
loss_D_B = (loss_real + loss_fake) / 2
loss_D_B.backward()
D_B_optimizer.step()
loss_D = (loss_D_A + loss_D_B) / 2
###########
# Logging #
###########
batches_done = epoch * len(train_loader) + i
batches_left = args.num_epochs * len(train_loader) - batches_done
time_left = datetime.timedelta(seconds=batches_left * (time.time() - batch_time))
batch_time = time.time()
print("\r[Epoch %d/%d] [Iteration %d/%d] [D loss: %f] [G loss: %f, adv: %f, cycle: %f] ETA: %s"
% (
epoch, args.num_epochs, i, len(train_loader),
loss_D.item(), loss_gen.item(), loss_adv.item(), loss_cycle.item(),
time_left
)
)
# If at sample interval save image
if batches_done > 0 and batches_done % 100 == 0:
sample_images(batches_done)
if epoch >= 100:
lr_scheduler_G.step(epoch)
lr_scheduler_D_A.step(epoch)
lr_scheduler_D_B.step(epoch)
if epoch % 30 == 0:
torch.save(G_AB.state_dict(), "saved_models/%s/G_AB_%d.pth" % (args.dataset_name, epoch))
torch.save(G_BA.state_dict(), "saved_models/%s/G_BA_%d.pth" % (args.dataset_name, epoch))
torch.save(D_A.state_dict(), "saved_models/%s/D_A_%d.pth" % (args.dataset_name, epoch))
torch.save(D_B.state_dict(), "saved_models/%s/D_B_%d.pth" % (args.dataset_name, epoch))
if(args.num_epoch % 30 != 0):
torch.save(G_AB.state_dict(), "saved_models/%s/G_AB_%d.pth" % (args.dataset_name, args.num_epoch))
torch.save(G_BA.state_dict(), "saved_models/%s/G_BA_%d.pth" % (args.dataset_name, args.num_epoch))
torch.save(D_A.state_dict(), "saved_models/%s/D_A_%d.pth" % (args.dataset_name, args.num_epoch))
torch.save(D_B.state_dict(), "saved_models/%s/D_B_%d.pth" % (args.dataset_name, args.num_epoch))