-
Notifications
You must be signed in to change notification settings - Fork 2
/
prediction_model.py
226 lines (158 loc) · 7.22 KB
/
prediction_model.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
import yaml
import h5py
import os
import time
import torch
import torch.nn as nn
from torch import optim
from torch.utils.data import DataLoader
from torch.autograd import Variable
import numpy as np
from torch.utils.data.dataloader import DataLoader
import data_utils
import space_angle_velocity
import bone_length_loss
import model_4GRU
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
config = yaml.load(open('./config.yml'),Loader=yaml.FullLoader)
node_num = config['node_num']
input_n=config['input_n']
output_n=config['output_n']
base_path = './data/WalkDog'
input_size = config['in_features']
hidden_size = config['hidden_size']
output_size = config['out_features']
batch_size = config['batch_size']
use_node = np.array([ 0, 1, 2, 3, 6, 7, 8, 11, 12, 13, 14, 15, 16, 17, 20, 21, 22])
#load data
s80 = []
s160 = []
s320 = []
s400 = []
s560 = []
s720 = []
s1000 = []
# Enter the start frame for testing.
index_list=[]
for i in range (len(index_list)):
#test_save_path = os.path.join(base_path, 'WalkDog.npy')
test_save_path = os.path.join(base_path, 'WalkDog_1.npy')
test_save_path = test_save_path.replace("\\","/")
dataset = np.load(test_save_path,allow_pickle = True)
dataset = torch.tensor(dataset, dtype = torch.float32,requires_grad=False)
# Compared to the original sequence, our input sequence is one frame shorter.
dataset = dataset[index_list[i]-1:,:,:].cuda()
input_dataset = dataset[0:input_n]
output_dataset = dataset[input_n:input_n+output_n]
input_dataset = input_dataset.expand(batch_size,input_dataset.shape[0],input_dataset.shape[1],input_dataset.shape[2])
output_dataset = output_dataset.expand(batch_size,output_dataset.shape[0],output_dataset.shape[1],output_dataset.shape[2])
input_dataset = input_dataset
output_dataset = output_dataset
total_samples = 0
total_mse = 0
total_mpjpe = 0
model_x = torch.load(os.path.join(base_path, 'generator_x_4GRU.pkl')).to(device)
model_y = torch.load(os.path.join(base_path, 'generator_y_4GRU.pkl')).to(device)
model_z = torch.load(os.path.join(base_path, 'generator_z_4GRU.pkl')).to(device)
model_v = torch.load(os.path.join(base_path, 'generator_v_4GRU.pkl')).to(device)
input_angle = input_dataset[:, 1:, :, :3]
input_velocity = input_dataset[:, 1:, :, 3].permute(0, 2, 1)
target_angle = output_dataset[:, :, :, :3]
target_velocity = output_dataset[:, :, :, 3]
#read velocity
input_velocity = input_velocity.float()
target_velocity = target_velocity.float()
#read angle_x
input_angle_x = input_angle[:,:,:,0].permute(0, 2, 1).float()
target_angle_x = target_angle[:,:,:,0].float()
#read angle_y
input_angle_y = input_angle[:,:,:,1].permute(0, 2, 1).float()
target_angle_y = target_angle[:,:,:,1].float()
#read angle_z
input_angle_z = input_angle[:,:,:,2].permute(0, 2, 1).float()
target_angle_z = target_angle[:,:,:,2].float()
#read 3D data
input_3d_data = input_dataset[:, :, :, 4:]
target_3d_data =output_dataset[:, :, :, 4:]
output_v, _ = model_v(input_velocity, hidden_size)
output_v = output_v.view(target_velocity.shape[0],target_velocity.shape[2],output_size)
output_x, _ = model_x(input_angle_x, hidden_size)
output_x = output_x.view(target_angle_x.shape[0],target_angle_x.shape[2],output_size)
output_y, _ = model_y(input_angle_y, hidden_size)
output_y = output_y.view(target_angle_y.shape[0],target_angle_y.shape[2],output_size)
output_z, _ = model_z(input_angle_z, hidden_size)
output_z = output_z.view(target_angle_z.shape[0],target_angle_z.shape[2],output_size)
angle_x = output_x.permute(0, 2, 1)
angle_y = output_y.permute(0, 2, 1)
angle_z = output_z.permute(0, 2, 1)
pred_v = output_v.permute(0, 2, 1)
pred_angle_set = torch.stack((angle_x,angle_y,angle_z),3)
pred_angle_set = pred_angle_set.reshape(pred_angle_set.shape[0],pred_angle_set.shape[1],-1,3)
#reconstruction_loss
input_pose = torch.zeros((target_velocity.shape[0], output_n, input_3d_data.shape[-2], input_3d_data.shape[-1]))
for a in range(input_pose.shape[0]):
input_pose[a,0,:,:] = input_3d_data[a,input_n-1,:,:]
re_data = torch.FloatTensor([])
for b in range (target_3d_data.shape[0]):
for c in range (target_3d_data.shape[1]):
reconstruction_coordinate = space_angle_velocity.reconstruction_motion(pred_v[b,c,:,], pred_angle_set[b, c,:,:], input_pose[b, c, :, :],node_num)
re_data = torch.cat([re_data,reconstruction_coordinate],dim=0)
reconstruction_coordinate = reconstruction_coordinate
if c+1<target_3d_data.shape[1]:
input_pose[b,c+1,:,:] = reconstruction_coordinate
else:
continue
re_data = re_data.view(target_3d_data.shape[0],-1,node_num,3)
frame_re_data = re_data[0]
frame_target_3d_data = target_3d_data[0]
# For fair, following "Learning dynamic relationships for 3d human motion prediction, CVPR, 2020", each pose is represented as a skeleton of 17 joints.
mpjpe_set = []
for e in range (frame_re_data.shape[0]):
frame_re_data = frame_re_data.to(device)
frame_target_3d_data = frame_target_3d_data.to(device)
frame_rec_loss = torch.mean(torch.norm(frame_re_data[e,use_node,:] - frame_target_3d_data[e,use_node,:], 2, 1))
mpjpe_set.append(frame_rec_loss)
#save vis data
frame_target_3d_data = frame_target_3d_data.cpu()
frame_re_data = frame_re_data.cpu()
frame_target_3d_data = np.array(frame_target_3d_data[0])
mpjpe_set = torch.Tensor(mpjpe_set)
mpjpe_set = np.array(mpjpe_set)
# vis_mpjpe_save_path = os.path.join(base_path, 'vis_mpjpe.npy')
# np.save(vis_save_path, frame_re_data)
# np.save(vis_mpjpe_save_path, mpjpe_set)
vis_name = 'vis_' + str(i) + '.npy'
GT_name = 'GT_' + str(i) + '.npy'
vis_save_path = os.path.join(base_path, vis_name)
GT_save_path = os.path.join(base_path, GT_name)
vis_mpjpe_save_path = os.path.join(base_path, 'vis_mpjpe.npy')
np.save(vis_save_path, frame_re_data)
np.save(GT_save_path, target_3d_data.cpu())
np.save(vis_mpjpe_save_path, mpjpe_set)
#print ('-------------------')
'''
print ('mpjpe_set',mpjpe_set)
print ('frame_re_data.shape:\n',frame_re_data.shape)
print('80ms:\n',mpjpe_set[1])
print('160ms:\n',mpjpe_set[3])
print('320ms:\n',mpjpe_set[7])
print('400ms:\n',mpjpe_set[9])
print('560ms:\n',mpjpe_set[13])
print('720ms:\n',mpjpe_set[17])
print('100ms:\n',mpjpe_set[24])
'''
s80.append(mpjpe_set[1])
s160.append(mpjpe_set[3])
s320.append(mpjpe_set[7])
s400.append(mpjpe_set[9])
s560.append(mpjpe_set[13])
s720.append(mpjpe_set[17])
s1000.append(mpjpe_set[24])
print('-----------------------')
print('80ms:\n', torch.mean(torch.Tensor(s80)))
print('160ms:\n',torch.mean(torch.Tensor(s160)))
print('320ms:\n',torch.mean(torch.Tensor(s320)))
print('400ms:\n',torch.mean(torch.Tensor(s400)))
print('560ms:\n',torch.mean(torch.Tensor(s560)))
print('720ms:\n',torch.mean(torch.Tensor(s720)))
print('1000ms:\n',torch.mean(torch.Tensor(s1000)))