-
Notifications
You must be signed in to change notification settings - Fork 0
/
dkgvibe.py
224 lines (184 loc) · 6.92 KB
/
dkgvibe.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
224
# -*- coding: utf-8 -*-
# Max-Planck-Gesellschaft zur Förderung der Wissenschaften e.V. (MPG) is
# holder of all proprietary rights on this computer program.
# You can only use this computer program if you have closed
# a license agreement with MPG or you get the right to use the computer
# program from someone who is authorized to grant you that right.
# Any use of the computer program without a valid license is prohibited and
# liable to prosecution.
#
# Copyright©2019 Max-Planck-Gesellschaft zur Förderung
# der Wissenschaften e.V. (MPG). acting on behalf of its Max Planck Institute
# for Intelligent Systems. All rights reserved.
#
# Contact: ps-license@tuebingen.mpg.de
import os
import torch
import os.path as osp
import torch.nn as nn
import torch.nn.functional as F
from lib.core.config import VIBE_DATA_DIR
from lib.models.spin import Regressor, hmr
class TemporalEncoder(nn.Module):
def __init__(
self,
n_layers=1,
hidden_size=2048,
add_linear=False,
bidirectional=False,
use_residual=True
):
super(TemporalEncoder, self).__init__()
self.gru = nn.GRU(
input_size=2048,
hidden_size=hidden_size,
bidirectional=bidirectional,
num_layers=n_layers
)
self.linear = None
if bidirectional:
self.linear = nn.Linear(hidden_size*2, 2048)
elif add_linear:
self.linear = nn.Linear(hidden_size, 2048)
self.use_residual = use_residual
def forward(self, x):
n,t,f = x.shape
x = x.permute(1,0,2) # NTF -> TNF
y, _ = self.gru(x)
if self.linear:
y = F.relu(y)
y = self.linear(y.view(-1, y.size(-1)))
y = y.view(t,n,f)
if self.use_residual and y.shape[-1] == 2048:
y = y + x
y = y.permute(1,0,2) # TNF -> NTF
return y
class DKGEncoder(nn.Module):
def __init__(
self,
n_layers=1,
hidden_size=2048,
add_linear=False,
bidirectional=False,
use_residual=True
):
super(DKGEncoder, self).__init__()
self.dkg = nn.graph(
input_size=2048,
hidden_size=hidden_size,
bidirectional=bidirectional,
num_layers=n_layers
)
self.linear = None
if bidirectional:
self.linear = nn.Linear(hidden_size*2, 2048)
elif add_linear:
self.linear = nn.Linear(hidden_size, 2048)
self.use_residual = use_residual
def forward(self, x):
n,t,f = x.shape
x = x.permute(1,0,2) # NTF -> TNF
y, _ = self.dkg(x)
if self.linear:
y = F.relu(y)
y = self.linear(y.view(-1, y.size(-1)))
y = y.view(t,n,f)
if self.use_residual and y.shape[-1] == 2048:
y = y + x
y = y.permute(1,0,2) # TNF -> NTF
return y
class VIBE(nn.Module):
def __init__(
self,
seqlen,
batch_size=64,
n_layers=1,
hidden_size=2048,
add_linear=False,
bidirectional=False,
use_residual=True,
pretrained=osp.join(VIBE_DATA_DIR, 'spin_model_checkpoint.pth.tar'),
):
super(VIBE, self).__init__()
self.seqlen = seqlen
self.batch_size = batch_size
self.encoder = TemporalEncoder(
n_layers=n_layers,
hidden_size=hidden_size,
bidirectional=bidirectional,
add_linear=add_linear,
use_residual=use_residual,
)
self.dkg = DKGEncoder(
n_layers=n_layers,
hidden_size=hidden_size,
bidirectional=bidirectional,
add_linear=add_linear,
use_residual=use_residual,
)
# regressor can predict cam, pose and shape params in an iterative way
self.regressor = Regressor()
if pretrained and os.path.isfile(pretrained):
pretrained_dict = torch.load(pretrained)['model']
self.regressor.load_state_dict(pretrained_dict, strict=False)
print(f'=> loaded pretrained model from \'{pretrained}\'')
def forward(self, input, J_regressor=None):
# input size NTF
batch_size, seqlen = input.shape[:2]
feature = self.encoder(input)
feature = self.dkg(feature)
feature = feature.reshape(-1, feature.size(-1))
smpl_output = self.regressor(feature, J_regressor=J_regressor)
for s in smpl_output:
s['theta'] = s['theta'].reshape(batch_size, seqlen, -1)
s['verts'] = s['verts'].reshape(batch_size, seqlen, -1, 3)
s['kp_2d'] = s['kp_2d'].reshape(batch_size, seqlen, -1, 2)
s['kp_3d'] = s['kp_3d'].reshape(batch_size, seqlen, -1, 3)
s['rotmat'] = s['rotmat'].reshape(batch_size, seqlen, -1, 3, 3)
return smpl_output
class VIBE_Demo(nn.Module):
def __init__(
self,
seqlen,
batch_size=64,
n_layers=1,
hidden_size=2048,
add_linear=False,
bidirectional=False,
use_residual=True,
pretrained=osp.join(VIBE_DATA_DIR, 'spin_model_checkpoint.pth.tar'),
):
super(VIBE_Demo, self).__init__()
self.seqlen = seqlen
self.batch_size = batch_size
self.encoder = TemporalEncoder(
n_layers=n_layers,
hidden_size=hidden_size,
bidirectional=bidirectional,
add_linear=add_linear,
use_residual=use_residual,
)
self.hmr = hmr()
checkpoint = torch.load(pretrained)
self.hmr.load_state_dict(checkpoint['model'], strict=False)
# regressor can predict cam, pose and shape params in an iterative way
self.regressor = Regressor()
if pretrained and os.path.isfile(pretrained):
pretrained_dict = torch.load(pretrained)['model']
self.regressor.load_state_dict(pretrained_dict, strict=False)
print(f'=> loaded pretrained model from \'{pretrained}\'')
def forward(self, input, J_regressor=None):
# input size NTF
batch_size, seqlen, nc, h, w = input.shape
feature = self.hmr.feature_extractor(input.reshape(-1, nc, h, w))
feature = feature.reshape(batch_size, seqlen, -1)
feature = self.encoder(feature)
feature = feature.reshape(-1, feature.size(-1))
smpl_output = self.regressor(feature, J_regressor=J_regressor)
for s in smpl_output:
s['theta'] = s['theta'].reshape(batch_size, seqlen, -1)
s['verts'] = s['verts'].reshape(batch_size, seqlen, -1, 3)
s['kp_2d'] = s['kp_2d'].reshape(batch_size, seqlen, -1, 2)
s['kp_3d'] = s['kp_3d'].reshape(batch_size, seqlen, -1, 3)
s['rotmat'] = s['rotmat'].reshape(batch_size, seqlen, -1, 3, 3)
return smpl_output