-
Notifications
You must be signed in to change notification settings - Fork 20
/
models_kp.py
289 lines (232 loc) · 10.2 KB
/
models_kp.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
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
import os
import time
import numpy as np
import torch
import torch.nn as nn
from torch.autograd import Variable
import torch.nn.functional as F
from data import denormalize, normalize
from utils import load_data
class SpatialSoftmax(torch.nn.Module):
def __init__(self, height, width, channel, lim=[-1., 1., -1., 1.], temperature=None, data_format='NCHW'):
super(SpatialSoftmax, self).__init__()
self.data_format = data_format
self.height = height
self.width = width
self.channel = channel
if temperature:
self.temperature = Parameter(torch.ones(1) * temperature)
else:
self.temperature = 1.
pos_x, pos_y = np.meshgrid(
np.linspace(lim[0], lim[1], self.width),
np.linspace(lim[2], lim[3], self.height))
pos_x = torch.from_numpy(pos_x.reshape(self.height * self.width)).float()
pos_y = torch.from_numpy(pos_y.reshape(self.height * self.width)).float()
self.register_buffer('pos_x', pos_x)
self.register_buffer('pos_y', pos_y)
def forward(self, feature):
# Output:
# (N, C*2) x_0 y_0 ...
if self.data_format == 'NHWC':
feature = feature.transpose(1, 3).tranpose(2, 3).view(-1, self.height * self.width)
else:
feature = feature.view(-1, self.height * self.width)
softmax_attention = F.softmax(feature / self.temperature, dim=-1)
expected_x = torch.sum(Variable(self.pos_x) * softmax_attention, dim=1, keepdim=True)
expected_y = torch.sum(Variable(self.pos_y) * softmax_attention, dim=1, keepdim=True)
expected_xy = torch.cat([expected_x, expected_y], 1)
feature_keypoints = expected_xy.view(-1, self.channel, 2)
return feature_keypoints
class KeyPointPredictor(nn.Module):
def __init__(self, args, lim=[-1., 1., -1., 1.]):
super(KeyPointPredictor, self).__init__()
nf = args.nf_hidden_kp
k = args.n_kp
norm_layer = args.norm_layer
sequence = [
# input is (ni) x 64 x 64
nn.Conv2d(3, nf, 7, 1, 3),
nn.BatchNorm2d(nf) if norm_layer == 'Batch' else nn.InstanceNorm2d(nf),
nn.LeakyReLU(0.2, inplace=True),
# feat size (nf) x 64 x 64
nn.Conv2d(nf, nf, 5, 1, 2),
nn.BatchNorm2d(nf) if norm_layer == 'Batch' else nn.InstanceNorm2d(nf),
nn.LeakyReLU(0.2, inplace=True),
# fesrcat size (nf) x 64 x 64
nn.Conv2d(nf, nf * 2, 4, 2, 1),
nn.BatchNorm2d(nf * 2) if norm_layer == 'Batch' else nn.InstanceNorm2d(nf * 2),
nn.LeakyReLU(0.2, inplace=True),
# feat size (nf * 2) x 32 x 32
nn.Conv2d(nf * 2, nf * 2, 3, 1, 1),
nn.BatchNorm2d(nf * 2) if norm_layer == 'Batch' else nn.InstanceNorm2d(nf * 2),
nn.LeakyReLU(0.2, inplace=True),
# feat size (nf * 2) x 32 x 32
nn.Conv2d(nf * 2, nf * 4, 4, 2, 1),
nn.BatchNorm2d(nf * 4) if norm_layer == 'Batch' else nn.InstanceNorm2d(nf * 4),
nn.LeakyReLU(0.2, inplace=True),
# feat size (nf * 4) x 16 x 16
nn.Conv2d(nf * 4, k, 1, 1)
# feat size (n_kp) x 16 x 16
]
self.model = nn.Sequential(*sequence)
self.integrater = SpatialSoftmax(
height=args.height//4, width=args.width//4, channel=k, lim=lim)
def integrate(self, heatmap):
return self.integrater(heatmap)
def forward(self, img):
heatmap = self.model(img)
return self.integrate(heatmap)
class FeatureExtractor(nn.Module):
def __init__(self, args):
super(FeatureExtractor, self).__init__()
nf = args.nf_hidden_kp
norm_layer = args.norm_layer
sequence = [
# input is (ni) x 64 x 64
nn.Conv2d(3, nf, 7, 1, 3),
nn.BatchNorm2d(nf) if norm_layer == 'Batch' else nn.InstanceNorm2d(nf),
nn.LeakyReLU(0.2, inplace=True),
# feat size (nf) x 64 x 64
nn.Conv2d(nf, nf, 5, 1, 2),
nn.BatchNorm2d(nf) if norm_layer == 'Batch' else nn.InstanceNorm2d(nf),
nn.LeakyReLU(0.2, inplace=True),
# feat size (nf) x 64 x 64
nn.Conv2d(nf, nf * 2, 4, 2, 1),
nn.BatchNorm2d(nf * 2) if norm_layer == 'Batch' else nn.InstanceNorm2d(nf * 2),
nn.LeakyReLU(0.2, inplace=True),
# feat size (nf * 2) x 32 x 32
nn.Conv2d(nf * 2, nf * 2, 3, 1, 1),
nn.BatchNorm2d(nf * 2) if norm_layer == 'Batch' else nn.InstanceNorm2d(nf * 2),
nn.LeakyReLU(0.2, inplace=True),
# feat size (nf * 2) x 32 x 32
nn.Conv2d(nf * 2, nf * 4, 4, 2, 1),
nn.BatchNorm2d(nf * 4) if norm_layer == 'Batch' else nn.InstanceNorm2d(nf * 4),
nn.LeakyReLU(0.2, inplace=True),
# feat size (nf * 4) x 16 x 16
]
self.model = nn.Sequential(*sequence)
def forward(self, img):
return self.model(img)
class Refiner(nn.Module):
def __init__(self, args):
super(Refiner, self).__init__()
nf = args.nf_hidden_kp
k = args.n_kp
norm_layer = args.norm_layer
sequence = [
# input is (nf * 4) x 16 x 16
nn.ConvTranspose2d(nf * 4, nf * 4, 4, 2, 1),
nn.BatchNorm2d(nf * 4) if norm_layer == 'Batch' else nn.InstanceNorm2d(nf * 4),
nn.LeakyReLU(0.2, inplace=True),
# input is (nf * 4) x 32 x 32
nn.Conv2d(nf * 4, nf * 2, 3, 1, 1),
nn.BatchNorm2d(nf * 2) if norm_layer == 'Batch' else nn.InstanceNorm2d(nf * 2),
nn.LeakyReLU(0.2, inplace=True),
# input is (nf * 4) x 32 x 32
nn.ConvTranspose2d(nf * 2, nf * 2, 4, 2, 1),
nn.BatchNorm2d(nf * 2) if norm_layer == 'Batch' else nn.InstanceNorm2d(nf * 2),
nn.LeakyReLU(0.2, inplace=True),
# input is (nf * 2) x 64 x 64
nn.Conv2d(nf * 2, nf, 5, 1, 2),
nn.BatchNorm2d(nf) if norm_layer == 'Batch' else nn.InstanceNorm2d(nf),
nn.LeakyReLU(0.2, inplace=True),
# input is (nf * 2) x 64 x 64
nn.Conv2d(nf, 3, 7, 1, 3)
]
self.model = nn.Sequential(*sequence)
def forward(self, feat):
return self.model(feat)
class KeyPointNet(nn.Module):
def __init__(self, args, use_gpu=True):
super(KeyPointNet, self).__init__()
self.args = args
self.use_gpu = use_gpu
# visual feature extractor
self.feature_extractor = FeatureExtractor(args)
# key point predictor
self.keypoint_predictor = KeyPointPredictor(args, lim=args.lim)
# map the feature back to the image
self.refiner = Refiner(args)
lim = args.lim
x = np.linspace(lim[0], lim[1], args.width // 4)
y = np.linspace(lim[2], lim[3], args.height // 4)
z = np.linspace(-1., 1., args.n_kp)
if use_gpu:
self.x = Variable(torch.FloatTensor(x)).cuda()
self.y = Variable(torch.FloatTensor(y)).cuda()
self.z = Variable(torch.FloatTensor(z)).cuda()
else:
self.x = Variable(torch.FloatTensor(x))
self.y = Variable(torch.FloatTensor(y))
self.z = Variable(torch.FloatTensor(z))
def extract_feature(self, img):
# img: B x 3 x H x W
# ret: B x (nf * 4) x (H / 4) x (W / 4)
return self.feature_extractor(img)
def predict_keypoint(self, img):
# img: B x 3 x H x W
# ret: B x n_kp x 2
return self.keypoint_predictor(img)
def keypoint_to_heatmap(self, keypoint, inv_std=10.):
# keypoint: B x n_kp x 2
# heatpmap: B x n_kp x (H / 4) x (W / 4)
# ret: B x n_kp x (H / 4) x (W / 4)
height = self.args.height // 4
width = self.args.width // 4
mu_x, mu_y = keypoint[:, :, :1].unsqueeze(-1), keypoint[:, :, 1:].unsqueeze(-1)
y = self.y.view(1, 1, height, 1)
x = self.x.view(1, 1, 1, width)
g_y = (y - mu_y)**2
g_x = (x - mu_x)**2
dist = (g_y + g_x) * inv_std**2
hmap = torch.exp(-dist)
return hmap
def transport(self, src_feat, des_feat, src_hmap, des_hmap, des_feat_hmap=None):
# src_feat: B x (nf * 4) x (H / 4) x (W / 4)
# des_feat: B x (nf * 4) x (H / 4) x (W / 4)
# src_hmap: B x n_kp x (H / 4) x (W / 4)
# des_hmap: B x n_kp x (H / 4) x (W / 4)
# des_feat_hmap = des_hmap * des_feat: B x (nf * 4) x (H / 4) * (W / 4)
# mixed_feat: B x (nf * 4) x (H / 4) x (W / 4)
src_hmap = torch.sum(src_hmap, 1, keepdim=True)
des_hmap = torch.sum(des_hmap, 1, keepdim=True)
src_digged = src_feat * (1. - src_hmap) * (1. - des_hmap)
# print(src_digged.size())
# print(des_hmap.size())
# print(des_feat.size())
if des_feat_hmap is None:
mixed_feat = src_digged + des_hmap * des_feat
else:
mixed_feat = src_digged + des_feat_hmap
return mixed_feat
def refine(self, mixed_feat):
# mixed_feat: B x (nf * 4) x (H / 4) x (W / 4)
# ret: B x 3 x H x W
return self.refiner(mixed_feat)
def kp_feat(self, feat, hmap):
# feat: B x (nf * 4) x (H / 4) x (W / 4)
# hmap: B x n_kp x (H / 4) x (W / 4)
# ret: B x n_kp x (nf * 4)
B, nf, H, W = feat.size()
n_kp = hmap.size(1)
p = feat.view(B, 1, nf, H, W) * hmap.view(B, n_kp, 1, H, W)
kp_feat = torch.sum(p, (3, 4))
return kp_feat
def forward(self, src, des):
# src: B x 3 x H x W
# des: B x 3 x H x W
# des_pred: B x 3 x H x W
cat = torch.cat([src, des], 0)
feat = self.extract_feature(cat)
kp = self.predict_keypoint(cat)
B = kp.size(0)
src_feat, des_feat = feat[:B//2], feat[B//2:]
src_kp, des_kp = kp[:B//2], kp[B//2:]
src_hmap = self.keypoint_to_heatmap(src_kp, self.args.inv_std)
des_hmap = self.keypoint_to_heatmap(des_kp, self.args.inv_std)
src_kp_feat = self.kp_feat(src_feat, src_hmap)
des_kp_feat = self.kp_feat(des_feat, des_hmap)
mixed_feat = self.transport(src_feat, des_feat, src_hmap, des_hmap)
des_pred = self.refine(mixed_feat)
return des_pred, src_kp_feat, des_kp_feat