-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathtrain.py
184 lines (146 loc) · 6.93 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
import pandas as pd
import torch
import torch.nn.functional as F
import torch.optim as optim
from torch.utils.data import DataLoader
from torchvision.utils import save_image
from model import FC, CNN
from parser_util import get_train_parser
from data.utils import get_dataset, inv_normalize
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
print('Device on: ', device)
parser = get_train_parser()
args = parser.parse_args()
def criterion(xs, x_preds, mu, logvar):
recon_loss = F.smooth_l1_loss(x_preds, xs, reduction='sum')
# See Appendix B from VAE paper:
# Solution of KL divergence, Gaussian case.
# https://arxiv.org/abs/1312.6114
kl_loss = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())
return recon_loss, kl_loss
def train(epoch, model, img_shape, optimizer, train_loader):
model.train()
for batch_idx, (xs, _) in enumerate(train_loader):
xs = xs.to(device)
x_preds, mu, logvar = model(xs)
x_preds = x_preds.view(-1, *img_shape)
recon_loss, kl_loss = criterion(xs, x_preds, mu, logvar)
loss = recon_loss + kl_loss
if epoch == 1 and batch_idx == 0:
print(
'Initial loss: {:.6f} Initial reconstruction loss: {:.6f} Initial kl loss: {:.6f}'.
format(loss.item(), recon_loss.item(), kl_loss.item()))
optimizer.zero_grad()
loss.backward()
optimizer.step()
# To make loss invariant to batch size, compute mean.
mean_loss = loss.item() / len(xs)
mean_recon_loss = recon_loss.item() / len(xs)
mean_kl_loss = kl_loss.item() / len(xs)
if batch_idx == 0:
running_loss = mean_loss
running_recon_loss = mean_recon_loss
running_kl_loss = mean_kl_loss
else:
running_loss = 0.05 * mean_loss + (1 - 0.05) * running_loss
running_recon_loss = 0.05 * mean_recon_loss + (1 - 0.05) * running_recon_loss
running_kl_loss = 0.05 * mean_kl_loss + (1 - 0.05) * running_kl_loss
if batch_idx % args.log_interval == 0:
print(
'Train Epoch: {}[{}/{} ({:.0f})%]\tLoss: {:.8f}\tReconstruction Loss: {:.8f}\tKL Divergence Loss: {:.8f}'
.format(epoch, batch_idx * len(xs), len(train_loader.dataset),
100. * batch_idx / len(train_loader), running_loss, running_recon_loss,
running_kl_loss))
return running_loss, running_recon_loss, running_kl_loss
def test(epoch, model, img_shape, test_loader):
model.eval()
sum_recon_loss = 0
sum_kl_loss = 0
sum_loss = 0
with torch.no_grad():
for batch_idx, (xs, _) in enumerate(test_loader):
xs = xs.to(device)
x_preds, mu, logvar = model(xs)
x_preds = x_preds.view(-1, *img_shape)
recon_loss, kl_loss = criterion(xs, x_preds, mu, logvar)
loss = recon_loss + kl_loss
sum_recon_loss += recon_loss.item()
sum_kl_loss += kl_loss.item()
sum_loss += loss.item()
if batch_idx == 0:
nrow = 8
xs = xs.detach().cpu()
x_preds = x_preds.detach().cpu()
comparision = torch.cat([xs[:nrow], x_preds[:nrow]])
comparision = inv_normalize(comparision, args.dataset)
save_image(comparision,
'results/' + args.dataset + f'/{args.model}' + '/reconstruction_' +
'epoch_' + str(epoch) + '.png',
nrow=nrow)
avg_loss = sum_loss / len(test_loader.dataset)
avg_recon_loss = sum_recon_loss / len(test_loader.dataset)
avg_kl_loss = sum_kl_loss / len(test_loader.dataset)
print("======> Test ")
print("Epoch: {} Test loss: {:.8f}\tTest reconstruction loss: {:.8f}\tTest kl loss: {:.8f}".
format(epoch, avg_loss, avg_recon_loss, avg_kl_loss))
return avg_loss, avg_recon_loss, avg_kl_loss
def main():
train_dataset = get_dataset(args.dataset, train=True)
test_dataset = get_dataset(args.dataset, train=False)
cuda_kwargs = {}
if torch.cuda.is_available():
cuda_kwargs = {'num_workers': 4, 'pin_memory': True}
train_loader = DataLoader(train_dataset,
batch_size=args.batch_size,
shuffle=True,
**cuda_kwargs)
test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=True, **cuda_kwargs)
img, _ = train_dataset[0]
img_shape = img.shape
if args.model == 'fc':
input_size = torch.flatten(img).shape[0]
model = FC(input_size=input_size, z_size=args.latent_dim).to(device)
elif args.model == 'cnn':
model = CNN(img_shape=img_shape, z_size=args.latent_dim).to(device)
optimizer = optim.Adam(model.parameters(), lr=args.lr)
scheduler = optim.lr_scheduler.StepLR(optimizer,
step_size=args.scheduler_step,
gamma=args.scheduler_gamma)
train_losses = []
train_recon_losses = []
train_kl_losses = []
test_losses = []
test_recon_losses = []
test_kl_losses = []
print("======> Start training")
for epoch in range(1, args.num_epochs + 1):
train_loss, train_recon_loss, train_kl_loss = train(epoch, model, img_shape, optimizer,
train_loader)
test_loss, test_recon_loss, test_kl_loss = test(epoch, model, img_shape, test_loader)
train_losses.append(train_loss)
train_recon_losses.append(train_recon_loss)
train_kl_losses.append(train_kl_loss)
test_losses.append(test_loss)
test_recon_losses.append(test_recon_loss)
test_kl_losses.append(test_kl_loss)
scheduler.step()
if args.save_model:
model_save_path = args.model + '_' + args.dataset + ".pt" if args.save_path is None else args.save_path
torch.save(model.state_dict(), model_save_path)
print(f"Model saved at {model_save_path}")
train_split = ['train'] * args.num_epochs
test_split = ['test'] * args.num_epochs
epochs = list(range(1, args.num_epochs + 1))
columns = ['Split', 'Epochs', 'Loss', 'Reconstruction_Loss', 'KL_Loss']
train_df = pd.DataFrame(list(
zip(train_split, epochs, train_losses, train_recon_losses, train_kl_losses)),
columns=columns)
test_df = pd.DataFrame(list(
zip(test_split, epochs, test_losses, test_recon_losses, test_kl_losses)),
columns=columns)
learning_curve_csv = args.learning_curve_csv if args.learning_curve_csv is not None else f'{args.model}_' + args.dataset + '.csv'
learning_curve_df = pd.concat([train_df, test_df], ignore_index=True)
learning_curve_df.to_csv(learning_curve_csv, mode='w')
print(f"Learning curve saved at: {learning_curve_csv}")
if __name__ == '__main__':
main()