-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathtrain.py
executable file
·196 lines (166 loc) · 8.77 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
import argparse
import os
import random
import numpy as np
import pandas as pd
import torch
from pytorch_metric_learning.losses import NormalizedSoftmaxLoss
from torch import nn
from torch.backends import cudnn
from torch.optim import Adam
from torch.utils.data.dataloader import DataLoader
from tqdm import tqdm
from model import Extractor, Discriminator, Generator, set_bn_eval
from utils import DomainDataset, compute_metric
# for reproducibility
random.seed(1)
np.random.seed(1)
torch.manual_seed(1)
cudnn.deterministic = True
cudnn.benchmark = False
# train for one epoch
def train(backbone, data_loader):
backbone.train()
# fix bn on backbone
backbone.apply(set_bn_eval)
generator.train()
discriminator.train()
total_extractor_loss, total_generator_loss, total_identity_loss, total_discriminator_loss = 0.0, 0.0, 0.0, 0.0
total_num, train_bar = 0, tqdm(data_loader, dynamic_ncols=True)
for sketch, photo, label in train_bar:
sketch, photo, label = sketch.cuda(), photo.cuda(), label.cuda()
optimizer_generator.zero_grad()
optimizer_extractor.zero_grad()
# generator #
fake = generator(sketch)
pred_fake = discriminator(fake)
# generator loss
target_fake = torch.ones(pred_fake.size(), device=pred_fake.device)
gg_loss = adversarial_criterion(pred_fake, target_fake)
total_generator_loss += gg_loss.item() * sketch.size(0)
# identity loss
ii_loss = identity_criterion(generator(photo), photo)
total_identity_loss += ii_loss.item() * sketch.size(0)
# extractor #
sketch_proj = backbone(sketch)
photo_proj = backbone(photo)
fake_proj = backbone(fake)
# extractor loss
class_loss = (class_criterion(sketch_proj, label) + class_criterion(photo_proj, label) +
class_criterion(fake_proj, label)) / 3
total_extractor_loss += class_loss.item() * sketch.size(0)
(gg_loss + 0.1 * ii_loss + 10 * class_loss).backward()
optimizer_generator.step()
optimizer_extractor.step()
# discriminator loss #
optimizer_discriminator.zero_grad()
pred_photo = discriminator(photo)
target_photo = torch.ones(pred_photo.size(), device=pred_photo.device)
pred_fake = discriminator(fake.detach())
target_fake = torch.zeros(pred_fake.size(), device=pred_fake.device)
adversarial_loss = (adversarial_criterion(pred_photo, target_photo) +
adversarial_criterion(pred_fake, target_fake)) / 2
total_discriminator_loss += adversarial_loss.item() * sketch.size(0)
adversarial_loss.backward()
optimizer_discriminator.step()
total_num += sketch.size(0)
e_loss = total_extractor_loss / total_num
g_loss = total_generator_loss / total_num
i_loss = total_identity_loss / total_num
d_loss = total_discriminator_loss / total_num
train_bar.set_description('Train Epoch: [{}/{}] E-Loss: {:.4f} G-Loss: {:.4f} I-Loss: {:.4f} D-Loss: {:.4f}'
.format(epoch, epochs, e_loss, g_loss, i_loss, d_loss))
return e_loss, g_loss, i_loss, d_loss
# val for one epoch
def val(backbone, encoder, data_loader):
backbone.eval()
encoder.eval()
vectors, domains, labels = [], [], []
with torch.no_grad():
for img, domain, label in tqdm(data_loader, desc='Feature extracting', dynamic_ncols=True):
img = img.cuda()
photo = img[domain == 0]
sketch = img[domain == 1]
photo_emb = backbone(photo)
sketch_emb = backbone(encoder(sketch))
emb = torch.cat((photo_emb, sketch_emb), dim=0)
vectors.append(emb.cpu())
label = torch.cat((label[domain == 0], label[domain == 1]), dim=0)
labels.append(label)
domain = torch.cat((domain[domain == 0], domain[domain == 1]), dim=0)
domains.append(domain)
vectors = torch.cat(vectors, dim=0)
domains = torch.cat(domains, dim=0)
labels = torch.cat(labels, dim=0)
acc = compute_metric(vectors, domains, labels)
results['P@100'].append(acc['P@100'] * 100)
results['P@200'].append(acc['P@200'] * 100)
results['mAP@200'].append(acc['mAP@200'] * 100)
results['mAP@all'].append(acc['mAP@all'] * 100)
print('Val Epoch: [{}/{}] | P@100:{:.1f}% | P@200:{:.1f}% | mAP@200:{:.1f}% | mAP@all:{:.1f}%'
.format(epoch, epochs, acc['P@100'] * 100, acc['P@200'] * 100, acc['mAP@200'] * 100,
acc['mAP@all'] * 100))
return acc['precise'], vectors
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='Train Model')
# common args
parser.add_argument('--data_root', default='/data', type=str, help='Datasets root path')
parser.add_argument('--data_name', default='sketchy', type=str, choices=['sketchy', 'tuberlin'],
help='Dataset name')
parser.add_argument('--backbone_type', default='resnet50', type=str, choices=['resnet50', 'vgg16'],
help='Backbone type')
parser.add_argument('--emb_dim', default=512, type=int, help='Embedding dim')
parser.add_argument('--batch_size', default=64, type=int, help='Number of images in each mini-batch')
parser.add_argument('--epochs', default=10, type=int, help='Number of epochs over the model to train')
parser.add_argument('--warmup', default=1, type=int, help='Number of warmups over the extractor to train')
parser.add_argument('--save_root', default='result', type=str, help='Result saved root path')
# args parse
args = parser.parse_args()
data_root, data_name, backbone_type, emb_dim = args.data_root, args.data_name, args.backbone_type, args.emb_dim
batch_size, epochs, warmup, save_root = args.batch_size, args.epochs, args.warmup, args.save_root
# data prepare
train_data = DomainDataset(data_root, data_name, split='train')
val_data = DomainDataset(data_root, data_name, split='val')
train_loader = DataLoader(train_data, batch_size=batch_size // 2, shuffle=True, num_workers=8)
val_loader = DataLoader(val_data, batch_size=batch_size, shuffle=False, num_workers=8)
# model define
extractor = Extractor(backbone_type, emb_dim).cuda()
generator = Generator(in_channels=8, num_block=8).cuda()
discriminator = Discriminator(in_channels=8).cuda()
# loss setup
class_criterion = NormalizedSoftmaxLoss(len(train_data.classes), emb_dim).cuda()
adversarial_criterion = nn.MSELoss()
identity_criterion = nn.L1Loss()
# optimizer config
optimizer_extractor = Adam([{'params': extractor.parameters()}, {'params': class_criterion.parameters(),
'lr': 1e-3}], lr=1e-5)
optimizer_generator = Adam(generator.parameters(), lr=1e-3, betas=(0.5, 0.999))
optimizer_discriminator = Adam(discriminator.parameters(), lr=1e-4, betas=(0.5, 0.999))
# training loop
results = {'extractor_loss': [], 'generator_loss': [], 'identity_loss': [], 'discriminator_loss': [],
'precise': [], 'P@100': [], 'P@200': [], 'mAP@200': [], 'mAP@all': []}
save_name_pre = '{}_{}_{}'.format(data_name, backbone_type, emb_dim)
if not os.path.exists(save_root):
os.makedirs(save_root)
best_precise = 0.0
for epoch in range(1, epochs + 1):
# warmup, not update the parameters of extractor, except the final fc layer
for param in list(extractor.backbone.parameters())[:-2]:
param.requires_grad = False if epoch <= warmup else True
extractor_loss, generator_loss, identity_loss, discriminator_loss = train(extractor, train_loader)
results['extractor_loss'].append(extractor_loss)
results['generator_loss'].append(generator_loss)
results['identity_loss'].append(identity_loss)
results['discriminator_loss'].append(discriminator_loss)
precise, features = val(extractor, generator, val_loader)
results['precise'].append(precise * 100)
# save statistics
data_frame = pd.DataFrame(data=results, index=range(1, epoch + 1))
data_frame.to_csv('{}/{}_results.csv'.format(save_root, save_name_pre), index_label='epoch')
if precise > best_precise:
best_precise = precise
torch.save(extractor.state_dict(), '{}/{}_extractor.pth'.format(save_root, save_name_pre))
torch.save(generator.state_dict(), '{}/{}_generator.pth'.format(save_root, save_name_pre))
torch.save(discriminator.state_dict(), '{}/{}_discriminator.pth'.format(save_root, save_name_pre))
torch.save(class_criterion.state_dict(), '{}/{}_proxies.pth'.format(save_root, save_name_pre))
torch.save(features, '{}/{}_vectors.pth'.format(save_root, save_name_pre))