-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathgan_train_v2.py
225 lines (187 loc) · 10.8 KB
/
gan_train_v2.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
import time
import torch
import torch.nn.functional as F
from tensorboardX import SummaryWriter
from torch.utils.data import DataLoader
from util.config import DATASET_PARAMETERS, NETWORKS_PARAMETERS
from util.parse_dataset import csv_to_list
from network import restore_train, get_network
from utils import Meter, cycle, save_model, get_collate_fn, Logger
from dataset import VoiceDataset, FaceDataset, Voice_Face_Dataset
# dataset and dataloader
print('Parsing your dataset...')
voice_list, face_list, id_class_num, emotion_class_num = csv_to_list(DATASET_PARAMETERS)
print('voice samples num = %d, face samples num = %d' %(len(voice_list),len(face_list)))
print('Preparing the datasets...')
# voice_face_dataset = Voice_Face_Dataset(voice_list,face_list,DATASET_PARAMETERS['nframe_range'])
voice_dataset = VoiceDataset(voice_list,DATASET_PARAMETERS['nframe_range'])
face_dataset = FaceDataset(face_list)
print('Preparing the dataloaders...')
collate_fn = get_collate_fn(DATASET_PARAMETERS['nframe_range'])
voice_loader = DataLoader(voice_dataset, shuffle=True, drop_last=True,
batch_size=DATASET_PARAMETERS['batch_size'],
num_workers=DATASET_PARAMETERS['workers_num'], # 使用多进程加载的进程数
collate_fn=collate_fn
) # 如何将多个样本数据拼接成一个batch
face_loader = DataLoader(face_dataset, shuffle=True, drop_last=True,
batch_size=DATASET_PARAMETERS['batch_size'],
num_workers=DATASET_PARAMETERS['workers_num'])
voice_iterator = iter(cycle(voice_loader))
face_iterator = iter(cycle(face_loader))
print('Initializing networks...')
NETWORKS_PARAMETERS['e']['output_channel'] = id_class_num
e_net, e_optimizer = get_network('e', NETWORKS_PARAMETERS, test=True) # 部分训练
NETWORKS_PARAMETERS['g']['input_channel'][1] = emotion_class_num
g_net, g_optimizer = get_network('g', NETWORKS_PARAMETERS, train=True)
# NETWORKS_PARAMETERS['d1-condition']['input_channel'][1] = emotion_class_num
d1_net, d1_optimizer = get_network('d0', NETWORKS_PARAMETERS, train=True)
d2_net, d2_optimizer = get_network('d0', NETWORKS_PARAMETERS, train=True)
f1_net, f1_optimizer = get_network('f', NETWORKS_PARAMETERS, train=True)
f2_net, f2_optimizer = get_network('f', NETWORKS_PARAMETERS, train=True)
NETWORKS_PARAMETERS['c']['output_channel'] = id_class_num
c1_net, c1_optimizer = get_network('c', NETWORKS_PARAMETERS, train=True)
NETWORKS_PARAMETERS['c']['output_channel'] = emotion_class_num
c2_net, c2_optimizer = get_network('c', NETWORKS_PARAMETERS, train=True)
# 接力训练,载入已有的模型
if NETWORKS_PARAMETERS['finetune']:
restore_train(g_net, d1_net, f1_net, f2_net)
# label for real/fake faces
real_label = torch.full((DATASET_PARAMETERS['batch_size'], 1), 1)
fake_label = torch.full((DATASET_PARAMETERS['batch_size'], 1), 0)
D_loss_positive = torch.tensor(1, dtype=torch.float)
D_loss_negative = D_loss_positive * -1
# Meters for recording the training status 日志模块 #
writer = SummaryWriter("./models/log")
logger = Logger(DATASET_PARAMETERS['log_dir'], time.strftime("%Y-%m-%d,%H,%M"))
iteration = Meter('Iter', 'sum', ':5d')
data_time = Meter('Data', 'sum', ':4.2f')
batch_time = Meter('Time', 'sum', ':4.2f')
D_real = Meter('D_real', 'avg', ':4.3f')
D_fake = Meter('D_fake', 'avg', ':4.3f')
C1_real = Meter('C1_real', 'avg', ':4.3f')
C2_real= Meter('C2_real', 'avg', ':4.3f')
C1_fake = Meter('C1_fake', 'avg', ':4.3f')
C2_fake= Meter('C2_fake', 'avg', ':4.3f')
GD_fake = Meter('G_D_fake', 'avg', ':4.3f')
criterionL1 = torch.nn.L1Loss()
print('Training models...')
for it in range(600000+1):
# data
start_time = time.time()
# voice, face, voice_identity_label, voice_emotion_label = next(voice_face_iterator)
# face_identity_label, face_emotion_label = voice_identity_label, voice_emotion_label
voice, voice_identity_label, voice_emotion_label = next(voice_iterator)
face, face_identity_label, face_emotion_label = next(face_iterator)
noise = 0.05*torch.randn(DATASET_PARAMETERS['batch_size'], 64, 1, 1) # 标准正态分布
# use GPU or not
if NETWORKS_PARAMETERS['GPU']:
voice, voice_identity_label, voice_emotion_label = voice.cuda(), voice_identity_label.cuda(), voice_emotion_label.cuda()
face, face_identity_label, face_emotion_label = face.cuda(), face_identity_label.cuda(), face_emotion_label.cuda()
real_label, fake_label = real_label.cuda(), fake_label.cuda()
noise = noise.cuda()
D_loss_positive, D_loss_negative = D_loss_positive.cuda(), D_loss_negative.cuda()
# get embeddings and generated faces
embeddings = e_net(voice)
embeddings = F.normalize(embeddings)
# introduce some permutations
embeddings = embeddings + noise
embeddings = F.normalize(embeddings)
embeddings = embeddings.squeeze() # 压缩维度从64,128,1,1 --> 64,128
# 扩展维度从64,1 --> 64, 8, 128, 128 , nn.Embedding(emotion_class_num,emotion_class_num)
face_EM_label_ = torch.zeros((DATASET_PARAMETERS['batch_size'], emotion_class_num)).scatter_(1, face_emotion_label.type(torch.LongTensor).unsqueeze(1), 1)
face_EM_label_ = face_EM_label_.unsqueeze(2).unsqueeze(3).expand(DATASET_PARAMETERS['batch_size'], emotion_class_num, face.size(2), face.size(3))
face_EM_label_ = face_EM_label_.cuda()
voice_EM_label_ = torch.zeros((DATASET_PARAMETERS['batch_size'], emotion_class_num)).scatter_(1, voice_emotion_label.type(torch.LongTensor).unsqueeze(1), 1)
voice_EM_label_ = voice_EM_label_.unsqueeze(2).unsqueeze(3).expand(DATASET_PARAMETERS['batch_size'], emotion_class_num, face.size(2), face.size(3))
voice_EM_label_ = voice_EM_label_.cuda()
fake_face = g_net(embeddings.unsqueeze(2).unsqueeze(3)) # G条件输入
""" --------------------update Discriminators and classifers-------------------------- """
f1_optimizer.zero_grad()
f2_optimizer.zero_grad()
d1_optimizer.zero_grad()
d2_optimizer.zero_grad()
c1_optimizer.zero_grad()
c2_optimizer.zero_grad()
# ------- 真实样本score------- #
real_score_out_1 = d1_net(f1_net(face)) # D1无条件输入
real_score_out_2 = d2_net(f2_net(face)) # D2无条件输入
D1_real_loss= F.binary_cross_entropy(torch.sigmoid(real_score_out_1), real_label) # BCE loss
D2_real_loss = F.binary_cross_entropy(torch.sigmoid(real_score_out_2), real_label) # BCE loss
D_real_loss = 1*D1_real_loss+0*D2_real_loss
# ------- 生成样本score------- #
fake_score_out_1 = d1_net(f1_net(fake_face.detach())) # D1无条件输入
fake_score_out_2 = d2_net(f2_net(fake_face.detach())) # D2无条件输入
D1_fake_loss = F.binary_cross_entropy(torch.sigmoid(fake_score_out_1), fake_label)
D2_fake_loss = F.binary_cross_entropy(torch.sigmoid(fake_score_out_2), fake_label)
D_fake_loss = 1*D1_fake_loss+0*D2_fake_loss
real_id_label_out = c1_net(f1_net(face)) # 计算 c1,c2 loss
real_emotion_label_out = c2_net(f2_net(face)) # 计算 c1,c2 loss
C1_real_loss = F.nll_loss(F.log_softmax(real_id_label_out, dim=1), face_identity_label)
C2_real_loss = F.nll_loss(F.log_softmax(real_emotion_label_out, dim=1), face_emotion_label)
(D_fake_loss + D_real_loss + 1*C1_real_loss + 1*C2_real_loss).backward()
f1_optimizer.step()
f2_optimizer.step()
d1_optimizer.step()
d2_optimizer.step()
c1_optimizer.step()
c2_optimizer.step()
# ---------------------------------------------
D_real.update(D_real_loss.item())
D_fake.update(D_fake_loss.item())
C1_real.update(C1_real_loss.item())
C2_real.update(C2_real_loss.item())
# ---------------------------------------------
""" --------------------------------update Generator --------------------------------------"""
g_optimizer.zero_grad()
fake_id_label_out = c1_net(f1_net(fake_face))
fake_emotion_label_out = c2_net(f2_net(fake_face))
fake_score_out_1 = d1_net(f1_net(fake_face)) # D无条件输入
fake_score_out_2 = d2_net(f2_net(fake_face))
GD_fake_loss1 = F.binary_cross_entropy(torch.sigmoid(fake_score_out_1), real_label)
GD_fake_loss2 = F.binary_cross_entropy(torch.sigmoid(fake_score_out_2), real_label)
GD_fake_loss = 1*GD_fake_loss1 +0*GD_fake_loss2
GC1_fake_loss = F.nll_loss(F.log_softmax(fake_id_label_out, dim=1), voice_identity_label) # 用真实标签替代随机标签？
GC2_fake_loss = F.nll_loss(F.log_softmax(fake_emotion_label_out, dim=1), voice_emotion_label)
# loss_G_L1_1 = criterionL1(fake_face, face) * 100
# GD_fake_loss = fake_score_out.mul(-1).mean() # hing loss
(GD_fake_loss + 1*GC1_fake_loss + 1*GC2_fake_loss).backward()
g_optimizer.step()
# ---------------------------------------------
GD_fake.update(GD_fake_loss.item())
C1_fake.update(GC1_fake_loss.item())
C2_fake.update(GC2_fake_loss.item())
batch_time.update(time.time() - start_time)
# ---------------------------------------------
# print status
if it % 10 == 0:
logger.info([iteration.__str__() + batch_time.__str__() +
D_real.__str__() + D_fake.__str__() + C1_real.__str__() +C2_real.__str__()+C1_fake.__str__()+C2_fake.__str__()+
GD_fake.__str__() ])
writer.add_scalars('data/scalar_group', {"D_real": D_real_loss,
"D_fake": D_fake_loss,
"C1_real_loss":C1_real_loss,
"C2_real_loss":C2_real_loss,
"C1_fake_loss": GC1_fake_loss,
"C2_fake_loss": GC2_fake_loss,
"GD_fake_loss":GD_fake_loss}, it)
# info = {'image/real_images': real_images(face, 8), 'image/generated_images': generate_img(fake_face, 8)}
# writer.add_images('image/generated_images', generate_img(fake_face, 8), it)
batch_time.reset()
D_real.reset()
D_fake.reset()
C1_real.reset()
C2_real.reset()
C1_fake.reset()
C2_fake.reset()
GD_fake.reset()
# snapshot
if it % 2000 == 0:
s_time = time.strftime("%m-%d,%H,%M") + '-' + str(it) + '-'
# save_model(e_net, 'models/voice_embedding/{}voice_embedding.pth'.format(s_time))
save_model(g_net, 'models/generator/{}generator.pth'.format(s_time))
# save_model(d1_net, 'models/discriminator/{}discriminator.pth'.format(s_time))
# save_model(f1_net, 'models/face_embedding/{}face_embedding1.pth'.format(s_time))
# save_model(f2_net, 'models/face_embedding/{}face_embedding2.pth'.format(s_time))
iteration.update(1)
# writer.export_scalars_to_json("./models/log/all_scalars.json")
# writer.close()