forked from cxisdawn/bigxigua_space
-
Notifications
You must be signed in to change notification settings - Fork 0
/
train_test.py
165 lines (152 loc) · 5.73 KB
/
train_test.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
import torch.nn as nn
import torch
import dataloader
from torch.types import Device
from torch.utils.data import Dataset, DataLoader
import torchvision
from models import model, funtion
import cv2
import numpy as np
from synergy3DMM import SynergyNet
import torch.optim as optim
import torch.optim.lr_scheduler as lr_scheduler
from configparser import ConfigParser
from itertools import zip_longest
# 如果有多片gpu时使用
# device_ids = [0, 1, 2] # 多个GPU的ID
# model = MyModel() # 自定义模型
# model = nn.DataParallel(model, device_ids=device_ids)
# 对于一片gpu
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
# 读取config文件
config = ConfigParser()
config.read('/content/SPACE/config.ini')
audio_length,_ = config.get('S2L_LSTM', 'audio_length').split(';', 1)
lmk_dim,_ = config.get('S2L_LSTM', 'lmk_dim').split(';', 1)
hidden_size = int(config.get('S2L_LSTM', 'hidden_size'))
batchsize = int(config.get('train', 'batchsize'))
epoch = int(config.get('train', 'epoch'))
pretrain = config.get('train', 'pretrain')
#####定义模型########
audio_net = model.AudioEncoder().to(device)
lmk_net = model.Encoder68().to(device)
s2l_lstm = model.S2L_LSTMModel(int(audio_length)+int(lmk_dim)*68, hidden_size, 2, 68*3).to(device)
if pretrain == True:
nets_loaded = torch.load('models1.pth')
audio_net.load_state_dict(nets_loaded['audio_net'])
lmk_net.load_state_dict(nets_loaded['lmk_net'])
s2l_lstm.load_state_dict(nets_loaded['s2l_lstm'])
params = []
params += list(audio_net.parameters())
params += list(lmk_net.parameters())
params += list(s2l_lstm.parameters())
# 创建Adam优化器实例
optimizer = optim.Adam(params, lr=1e-5)
# 创建学习率调度器实例
# scheduler = lr_scheduler.LinearLR(optimizer, 10000, 5e-4)
#########################
#提取脸部标志所需模型
synergynet = SynergyNet()
#加载数据 dataloader
audio_folder = "data/audio"
image_folder = "data/images"
dataset = dataloader.AudioImageDataset(audio_folder, image_folder, transform=None)
dataloader = DataLoader(dataset, batch_size=batchsize, shuffle=True)
#绝对误差损失函数
loss1 = nn.L1Loss()
step = 0
for i in range(epoch):
print('----------------第{}轮训练开始-------------------'.format(i+1))
print("dataloader读取")
for mfccs, images in dataloader:
# 初始化h0,c0
h0 = torch.randn(2, batchsize, 1024).to(device)
c0 = torch.randn(2, batchsize, 1024).to(device)
# print("11111")
"""
mcffs[batchsize,帧数,channels,系数]
images[batchsize,帧数,H,W,C]
都是tensor类型
"""
images = np.array(images)
lmklist,lmk_step = funtion.extract_lankmarks(images,synergynet)
if lmklist is None:
print("数据错误,跳过该轮循环")
continue
# print(lmklist.shape,lmk_step.shape,111)
"""
lmklist:[batchsize,fps,points,dim] 正常帧
lmk_step:[batchsize,fps,points-1,dim] 少了第一帧
"""
fps_length = lmklist.shape[1] #帧数
#将每一帧的音频信息送进网络处理
mfccs = mfccs.to(device)
processed_outputs = []
for mfcc in mfccs:
#[fps,c,length]
audio_feature = audio_net(mfcc)
# print(audio_feature.shape)
processed_outputs.append(audio_feature)
# 将处理后的结果拼接成一个batch [batch,fps,c,len]
batch_mfcc = torch.stack(processed_outputs, dim=0)
# print(batch_mfcc.shape,222)
mfcc_cut = batch_mfcc[:,1:,:,:] # 帧数对齐
mfcc_cut = mfcc_cut.to(device)
audio_feature = mfcc_cut.reshape(batchsize,fps_length,1,-1) #[batch,帧数,1,feature_length]]
#对数据进行维度转换[fps,batch......]
audio_feature = audio_feature.transpose(1,0)
lmklist = lmklist.transpose(1,0)
lmk_step = lmk_step.transpose(1,0)
# print(audio_feature.shape)
# print(lmklist.shape)
# print(lmk_step.shape)
#将数据送入GPU
lmklist = lmklist.to(device)
lmk_step = lmk_step.to(device)
"""
lmklist:[fps,batchsize,points,dim] 正常帧
lmk_step:[fps-1,batchsize,points,dim] 少了第一帧
audio_feature:[fps,batchsize,1,feature_length]
"""
############################################
count = 0
loss_all = 0
for audio_fps, image_fps, next_step in zip_longest(audio_feature, lmklist, lmk_step, fillvalue=None):
"""
audio_fps:[batch,1,length]
image_fps:[batch,68,3]
next_step:[batch,68,3]
"""
if next_step is None:
print(11111)
break
image_code = lmk_net(image_fps) #经过脸部标志编码器 [batch,204]
# print(image_code.shape,111) #[2,204]
feature_add = funtion.lm_splice(audio_fps, image_code) #[T,batchsize,length]
# print(feature_add.shape,3333)
predict,h0,c0 = s2l_lstm(feature_add,h0,c0)
# print(predict.shape)
predict = predict.reshape(batchsize,68,3)
##只对循环里面的网络梯度清零
s2l_lstm.zero_grad()
lmk_net.zero_grad()
loss = loss1(predict,next_step)
loss.backward(retain_graph=True)
optimizer.step()
# audio_net.zero_grad()
# lmk_net.zero_grad()
# s2l_lstm.zero_grad()
loss_all += loss
count += 1
step += 1
if step%740==0:
print('第{}次训练'.format(step))
print('参数保存')
# 保存模型参数
state_dicts = {'audio_net': audio_net.state_dict(), 'lmk_net': lmk_net.state_dict(),
's2l_lstm': s2l_lstm.state_dict()}
torch.save(state_dicts, 'models{}.pth'.format(step))
print(loss_all/count)
##对所有网络梯度清零
optimizer.zero_grad()
# # print(mfcc.shape,images.shape)