-
Notifications
You must be signed in to change notification settings - Fork 0
/
P1_A_training.py
175 lines (147 loc) · 5.07 KB
/
P1_A_training.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
import random
from pathlib import Path
import torch
import torch.nn as nn
from pytorch_fid import fid_score
from torch.utils.data import DataLoader
from torch.utils.tensorboard.writer import SummaryWriter
from torchvision import transforms
from torchvision.utils import make_grid, save_image
from tqdm import tqdm
from P1_A_model import DCGAN_D, DCGAN_G
from P1_dataloader import p1_dataset
def get_FID(device, generator, out_dir, eval_noise):
batch_size = 100
generator.eval()
idx = 0
with torch.no_grad():
gen_imgs = generator(eval_noise)
gen_imgs = UnNormalize(gen_imgs)
for img in gen_imgs:
save_image(img, out_dir / f'{idx}.png')
idx += 1
generator.train()
FID = fid_score.calculate_fid_given_paths(
[str(out_dir), 'hw2_data/face/val'],
batch_size=batch_size,
device=device,
dims=2048,
num_workers=8,
)
return FID
def rm_tree(pth: Path):
if pth.is_dir():
for child in pth.iterdir():
if child.is_file():
child.unlink()
else:
rm_tree(child)
pth.rmdir()
# [0.5696, 0.4315, 0.3593] # calculated on training set
mean = [0.5, 0.5, 0.5]
std = [0.5, 0.5, 0.5]
train_set = p1_dataset(
root='./hw2_data/face/train',
transform=transforms.Compose([
transforms.RandomHorizontalFlip(),
transforms.ToTensor(),
transforms.Normalize(mean, std),
])
)
UnNormalize = transforms.Normalize(
mean=[-u / s for u, s in zip(mean, std)],
std=[1 / s for s in std],
)
batch_size = 128
train_loader = DataLoader(
train_set, batch_size=batch_size, shuffle=True, num_workers=6)
num_epochs = 150
lr = 2e-4
ckpt_path = Path('./P1_A_ckpt')
tb_path = Path('./P1_A_tb')
out_path = Path('./P1_A_out')
rm_tree(ckpt_path)
rm_tree(tb_path)
rm_tree(out_path)
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
loss_fn = nn.BCELoss() # take y as 1 or 0 <-> choosing logx or log(1-x) in BCE
model_G = DCGAN_G().to(device) # z.shape = (b, 100, 1, 1)
model_D = DCGAN_D().to(device)
optim_G = torch.optim.Adam(model_G.parameters(), lr=lr, betas=(0.5, 0.999))
optim_D = torch.optim.Adam(model_D.parameters(), lr=lr, betas=(0.5, 0.999))
writer = SummaryWriter(tb_path)
# plot every K iterations
plot_noise = torch.randn(64, 100, 1, 1, device=device)
eval_noise = torch.randn(batch_size, 100, 1, 1, device=device)
iters = 0
ckpt_path.mkdir(exist_ok=True)
out_path.mkdir(exist_ok=True)
best_epoch = -1
best_FID = 10e10
for epoch in range(1, num_epochs + 1):
for real_imgs in tqdm(train_loader):
'''
Update Discriminator:
maximizes log(D(x)) + log(1 - D(G(z)))
'''
optim_D.zero_grad()
# train with all real batch (GAN hack)
real_imgs = real_imgs.to(device)
batch_size = real_imgs.shape[0]
out_D_real = model_D(real_imgs).flatten()
loss_D_real = loss_fn(
out_D_real,
torch.ones(batch_size, dtype=torch.float32, device=device)
)
L_D = out_D_real.mean().item()
loss_D_real.backward()
# train with all fake batch
z = torch.randn(batch_size, 100, 1, 1, device=device) # normal dist.
syn_imgs = model_G(z)
# detach since we are updating D
out_D_syn = model_D(syn_imgs.detach()).flatten()
loss_D_syn = loss_fn(
out_D_syn,
torch.zeros(batch_size, dtype=torch.float32, device=device)
)
L_G_1 = out_D_syn.mean().item()
loss_D_syn.backward()
optim_D.step()
loss_D = (loss_D_real + loss_D_syn).item()
'''
Update Generator:
~minimizes log(1 - D(G(z)))~
-> maximizes log(D(G(z)))
'''
optim_G.zero_grad()
out_G = model_D(syn_imgs).flatten()
loss_G = loss_fn(
out_G,
torch.ones(batch_size, dtype=torch.float32, device=device),
)
L_G_2 = out_G.mean().item()
loss_G.backward()
optim_G.step()
loss_G = loss_G.item()
writer.add_scalar('loss/D', loss_D, global_step=iters)
writer.add_scalar('loss/G', loss_G, global_step=iters)
writer.add_scalar('D/D(x)', L_D, global_step=iters)
writer.add_scalar('D/D(G(z)) 1', L_G_1, global_step=iters)
writer.add_scalar('D/D(G(z)) 2', L_G_2, global_step=iters)
iters += 1
with torch.no_grad():
plot_img = model_G(plot_noise).detach().cpu()
plot_img = UnNormalize(plot_img)
grid = make_grid(plot_img, padding=2)
writer.add_image('GAN results', grid, epoch)
FID = get_FID(device=device, generator=model_G,
out_dir=out_path, eval_noise=eval_noise)
if FID <= best_FID:
best_FID = FID
best_epoch = epoch
print(f"[NEW] EPOCH {epoch} BEST FID: {FID}")
torch.save(model_G.state_dict(), ckpt_path / "best_G.pth")
torch.save(model_D.state_dict(), ckpt_path / "best_D.pth")
else:
print(f"[BAD] EPOCH {epoch} FID: {FID}, BEST FID: {best_FID}")
print(f"[RST] EPOCH {best_epoch}, best FID: {best_FID}")