-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_cgan.py
186 lines (145 loc) · 6.82 KB
/
train_cgan.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
import cv2
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
import torch
import torchvision
import torch.nn as nn
import torch.nn.functional as F
from torch.utils.data import Dataset, DataLoader
from torchvision import transforms, utils
from torchvision import models
import torch.optim as optim
from torch.autograd import Variable
import torchvision.utils as vutils
from utils.process_img import Rescale, DynamicCrop, ToTensor, CenterCrop
from utils.func import weights_init, random_annotate, gaussian_noise
from pose_dataset import PoseDataset, print_sample
from model.generator import Generator
from model.discriminator import Discriminator
from utils.process_text import tokenizer, get_embeddings, get_word2idx
import pytorch_ssim
import pickle
opt = {'b1': 0.5,
'b2': 0.999,
'batch_size': 64,
'channels': 1,
'img_size': 32,
'latent_dim': 100,
'lr': 0.0002,
'n_classes': 200,
'n_cpu': 8,
'n_epochs': 1,
'sample_interval': 400}
img_size = 64
batch_size = 2
img_shape = (3, 64, 64)
composed = transforms.Compose([Rescale(512),
DynamicCrop(30),
Rescale((img_size, img_size))])
pose_dataset = PoseDataset('./data/data_list_label.csv', './data', transform = composed, gray_scale = False)
pose_dataloader = DataLoader(pose_dataset, batch_size=batch_size, shuffle=True, num_workers=4)
adversarial_loss = torch.nn.MSELoss()
consistency_loss = pytorch_ssim.SSIM()#losses.SSIM(window_size = 5, reduction = 'sum')
generator_parsing = Generator().cuda()
discriminator_parsing = Discriminator().cuda()
generator_pose= Generator().cuda()
discriminator_pose = Discriminator().cuda()
optimizer_G_parsing = torch.optim.Adam(generator_parsing.parameters(), lr=opt['lr'], betas=(opt['b1'], opt['b2']))
optimizer_D_parsing = torch.optim.Adam(discriminator_parsing.parameters(), lr=opt['lr'], betas=(opt['b1'], opt['b2']))
optimizer_G_pose = torch.optim.Adam(generator_pose.parameters(), lr=opt['lr'], betas=(opt['b1'], opt['b2']))
optimizer_D_pose = torch.optim.Adam(discriminator_pose.parameters(), lr=opt['lr'], betas=(opt['b1'], opt['b2']))
test_batch_size = 100
torch.manual_seed(4995)
fixed_z = torch.randn(test_batch_size, opt['latent_dim']).cuda()
fixed_gen_labels = torch.from_numpy(np.arange(test_batch_size)).cuda()
gen_img_list_pose, gen_img_list_parsing = [], []
G_loss_pose, D_loss_pose, G_loss_parsing, D_loss_parsing = [], [], [], []
# Train
n_epochs = 5
for epoch in range(n_epochs):
for i, sample in enumerate(pose_dataloader):
parsing = sample['parsing'].cuda()
pose = sample['pose'].cuda()
labels = sample['label'].cuda()
b_size = parsing.shape[0]
# Adversarial ground truths
valid = torch.full((b_size, ), 1.0).cuda()
fake = torch.full((b_size, ), 0.0).cuda()
# -----------------
# Train Generator
# -----------------
# Sample noise and labels as generator input
z = torch.randn(b_size, opt['latent_dim']).cuda()
gen_labels = torch.from_numpy(np.random.randint(0, 200, b_size)).cuda()
# Pose
optimizer_G_pose.zero_grad()
gen_parsing = generator_parsing(z, gen_labels)
gen_pose = generator_pose(z, gen_labels)
loss_consistency = -consistency_loss(gen_parsing, gen_pose)
validity_pose = discriminator_pose(gen_pose, gen_labels)
g_loss_pose = adversarial_loss(validity_pose.view(-1), valid)
g_loss_pose = (g_loss_pose + loss_consistency) / 2
g_loss_pose.backward()
optimizer_G_pose.step()
# Parsing
optimizer_G_parsing.zero_grad()
gen_parsing = generator_parsing(z, gen_labels)
gen_pose = generator_pose(z, gen_labels)
loss_consistency = -consistency_loss(gen_parsing, gen_pose)
validity_parsing = discriminator_parsing(gen_parsing, gen_labels)
g_loss_parsing = adversarial_loss(validity_parsing.view(-1), valid)
g_loss_parsing = (g_loss_parsing + loss_consistency) / 2
g_loss_parsing.backward()
optimizer_G_parsing.step()
G_loss_parsing.append(g_loss_parsing.item())
G_loss_pose.append(g_loss_pose.item())
# ---------------------
# Train Discriminator
# ---------------------
optimizer_D_parsing.zero_grad()
optimizer_D_pose.zero_grad()
# Loss for real images
validity_real = discriminator_pose(pose, labels)
d_real_loss_pose = adversarial_loss(validity_real.view(-1), valid)
validity_real = discriminator_parsing(parsing, labels)
d_real_loss_parsing = adversarial_loss(validity_real.view(-1), valid)
# Loss for fake images
validity_fake = discriminator_pose(gen_pose.detach(), gen_labels)
d_fake_loss_pose = adversarial_loss(validity_fake.view(-1), fake)
validity_fake = discriminator_parsing(gen_parsing.detach(), gen_labels)
d_fake_loss_parsing = adversarial_loss(validity_fake.view(-1), fake)
# Total discriminator loss
d_loss_pose = (d_real_loss_pose + d_fake_loss_pose) / 2
d_loss_parsing = (d_real_loss_parsing + d_fake_loss_parsing) / 2
D_loss_pose.append(d_loss_pose.item())
D_loss_parsing.append(d_loss_parsing.item())
d_loss_pose.backward()
d_loss_parsing.backward()
optimizer_D_parsing.step()
optimizer_D_pose.step()
if i and i % 5 == 0:
print(
"[Epoch %d/%d] [Batch %d/%d] [D_pose loss: %f] [G_pose loss: %f] [D_parsing loss: %f] [G_parsing loss: %f]"
% (epoch, n_epochs, i, len(pose_dataloader), d_loss_pose.item(), g_loss_pose.item(), d_loss_parsing.item(), g_loss_parsing.item())
)
if i and i % 30 == 0:
gen_parsing = generator_parsing(fixed_z, fixed_gen_labels)
gen_pose = generator_pose(fixed_z, fixed_gen_labels)
gen_img_list_parsing.append(gen_parsing)
gen_img_list_pose.append(gen_pose)
img_parsing = np.reshape(gen_parsing[0].cpu().detach().numpy(), (64, 64, 3))
img_pose = np.reshape(gen_pose[0].cpu().detach().numpy(), (64, 64, 3))
fig=plt.figure(figsize=(8, 8))
fig.add_subplot(1, 2, 1)
plt.imshow(img_parsing)
fig.add_subplot(1, 2, 2)
plt.imshow(img_pose)
plt.show()
with open('gen_parsing_uni.pk', 'wb') as f:
pickle.dump(gen_img_list_parsing, f)
with open('gen_pose_uni.pk', 'wb') as f:
pickle.dump(gen_img_list_pose, f)
loss = {'G_loss_pose': G_loss_pose, 'D_loss_pose': D_loss_pose, 'G_loss_parsing': G_loss_parsing, 'D_loss_parsing': D_loss_parsing}
with open('loss_uni.pk', 'wb') as f:
pickle.dump(loss, f)