-
Notifications
You must be signed in to change notification settings - Fork 1
/
train_pose.py
164 lines (145 loc) · 5.84 KB
/
train_pose.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
import numpy as np
import torch,pdb
import torch.nn as nn
from ConvLSTM import *
from utils import *
import time
from torch.autograd import Variable
from gan_model import *
#param
batch_size = 5
sequence_length = 8
epoch = 20
image_shape = 64
ConvLSTM_channel = 128
pose_generator_basefeature = 128
model_file = 'model_it_1943'
train_player_list = [[5,6],[7,8],[9,10],[11,12]]
test_player_list = [[13,14]]
sigma = 0.01
test_interval = 1
eval_interval = 1
#model init
pose_generator = Generator(pose_generator_basefeature)
pose_discriminator = Discriminator(ConvLSTM_channel,image_shape)
convlstm_model = ConvLSTM(ConvLSTM_channel, sequence_length)
#data loader init
train_pose_dataset = Pose_Dataset(Boxing_dir, sequence_length, player_list, len(player_list),sigma,\
epoch, grid_point = image_shape)
test_pose_dataset = Pose_Dataset(Boxing_dir, sequence_length, test_player_list, len(test_player_list),sigma,\
epoch, grid_point = image_shape, mode = 'test')
train_loader = torch.utils.data.DataLoader(dataset = train_pose_dataset,
batch_size=batch_size,
shuffle=True,
num_workers=4)
test_pose_dataset = torch.utils.data.DataLoader(dataset = train_pose_dataset,
batch_size=batch_size,
shuffle=False,
num_workers=4)
G_optmizer = torch.optim.Adam(pose_generator.parameters(),lr = lr_rate, weight_decay = weight_decay)
D_optmizer = torch.optim.Adam(pose_discriminator.parameters(),lr = lr_rate, weight_decay = weight_decay)
Convlstm_optmizer = torch.optim.Adam(convlstm_model.parameters(),lr = lr_rate, weight_decay = weight_decay)
total_data_num = train_pose_dataset.total_data_num
data_iter = iter(train_loader)
loss_fn = torch.nn.BCELoss()
loss_fn.cuda()
pose_generator.cuda()
pose_discriminator.cuda()
convlstm_model.cuda()
fake_label = Variable(torch.zeros(batch_size).cuda())
real_label = Variable(torch.ones(batch_size).cuda())
G_loss = 0
D_loss = 0
counter = 0
time_stmp = time.time()
print('Start Training')
for p1_data,p2_data in test_loader:
p1_data = p1_data.permute(1,0,4,2,3)
p2_data = p2_data.permute(1,0,4,2,3)
pred_p1 = p1_data[p1_data.size(0)-1,:,:,:,:]
input_p1 = p1_data[:p1_data.size(0)-1,:,:,:,:]
pred_p2 = p2_data[p2_data.size(0)-1,:,:,:,:]
input_p2 = p2_data[:p2_data.size(0)-1,:,:,:,:]
pred_p1 = Variable(pred_p1.cuda())
pred_p2 = Variable(pred_p2.cuda())
input_p1 = Variable(input_p1.cuda())
input_p2 = Variable(input_p2.cuda())
p1_data = Variable(p1_data.cuda())
p2_data = Variable(p2_data.cuda())
# Update D on real data
pose_discriminator.zero_grad()
p1_gt_fm = convlstm_model(p1_data)
p1_data_D_true = Discriminator(p1_gt_fm)
loss_p1_true = loss_fn(p1_data_D_true, real_label)
loss_p1_true.backward()
D_optmizer.step()
pose_discriminator.zero_grad()
p2_gt_fm = convlstm_model(p2_data)
p2_data_D_true = Discriminator(p2_gt_fm)
loss_p2_true = loss_fn(p2_data_D_true, real_label)
loss_p2_true.backward()
D_optmizer.step()
# Update D on fake data
pose_discriminator.zero_grad()
input_p1_fm = convlstm_model(input_p1)
input_p2_fm = convlstm_model(input_p2)
p1_fake_pose = pose_generator(input_p1_fm,input_p2_fm)
p1_data_fake = torch.stack(input_p1,p1_fake_pose)
p1_fake_fm = convlstm_model(p1_data_fake)
p1_data_D_fake = Discriminator(p1_fake_fm)
loss_p1_fake = loss_fn(p1_data_D_fake, fake_label)
loss_p1_fake.backward()
D_optmizer.step()
pose_discriminator.zero_grad()
p2_fake_pose = pose_generator(input_p2_fm, input_p1_fm)
p2_data_fake = torch.stack(input_p2,p2_fake_pose)
p2_fake_fm = convlstm_model(p2_data_fake)
p2_data_D_fake = Discriminator(p2_fake_fm)
loss_p2_fake = loss_fn(p2_data_D_fake, fake_label)
loss_p2_fake.backward()
D_optmizer.step()
D_loss += (loss_p1_true.data[0] + loss_p2_true.data[0] + loss_p1_fake.data[0] + loss_p2_fake.data[0])/4
# Update G
pose_generator.zero_grad()
p1_fake_pose = pose_generator(input_p1_fm,input_p2_fm)
p1_data_fake = torch.stack(input_p1,p1_fake_pose)
p1_fake_fm = convlstm_model(p1_data_fake)
p1_data_D_fake = Discriminator(p1_fake_fm)
loss_p1_fake = loss_fn(p1_data_D_fake, real_label)
loss_p1_fake.backward()
G_optmizer.step()
pose_generator.zero_grad()
p2_fake_pose = pose_generator(input_p2_fm, input_p1_fm)
p2_data_fake = torch.stack(input_p2,p2_fake_pose)
p2_fake_fm = convlstm_model(p2_data_fake)
p2_data_D_fake = Discriminator(p2_fake_fm)
loss_p2_fake = loss_fn(p2_data_D_fake, fake_label)
loss_p2_fake.backward()
G_optmizer.step()
G_loss += (loss_p1_fake.data[0] + loss_p2_fake.data[0])/2
counter+= 1
if counter % eval_interval == 0:
print('iteration: {}, Time:{}, G_loss: {}, D_loss:{}'.format(counter, time.time()- time_stmp, G_loss, D_loss))
if counter % test_interval == 0:
print('Do testing....')
test_error = 0
test_counter = 1
test_time_stmp = time.time()
for p1_data,p2_data, p1_coor in test_loader:
p1_data = p1_data.permute(1,0,4,2,3)
p2_data = p2_data.permute(1,0,4,2,3)
input_p1 = p1_data[:p1_data.size(0)-1,:,:,:,:]
input_p2 = p2_data[:p2_data.size(0)-1,:,:,:,:]
p1_fake_pose = pose_generator(input_p1,input_p2)
p1_pred = np.reshape(p1_fake_pose.data[0],[batch_size,p1_fake_pose.size(1),-1])
p1_pred_coor = np.argmax(p1_pred, axis = 2)
p1_pred_coor_x = p1_pred_coor % image_shape
p1_pred_coor_y = p1_pred_coor / image_shape
test_error += np.mean(np.sqrt((((p1_pred_coor_x - p1_coor[:,:,0])**2 \
+ (p1_pred_coor_y - p1_coor[:,:,1])**2))))
test_counter += 1
test_error = test_error/test_counter
print('Testing is Done, test_error:{}'.format(test_error))
torch.save(pose_generator.save_state_dict(), 'model/D_model_{}_{}'.format(counter,test_error))
torch.save(pose_discriminator.save_state_dict(), 'model/G_model_{}_{}'.format(counter,test_error))
print('Model Saved, test time :{}'.format(time.time()- test_time_stmp))