-
Notifications
You must be signed in to change notification settings - Fork 606
/
Copy pathlabel_encoder.py
216 lines (173 loc) · 9.13 KB
/
label_encoder.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
import torch
import torch.nn as nn
import math
from ltr.models.backbone.resnet import BasicBlock
from ltr.models.layers.blocks import conv_block
from ltr.models.lwl.utils import interpolate
class ResidualDS16SW(nn.Module):
""" Outputs the few-shot learner label and spatial importance weights given the segmentation mask """
def __init__(self, layer_dims, use_bn=True):
super().__init__()
self.conv_block = conv_block(1, layer_dims[0], kernel_size=3, stride=2, padding=1, batch_norm=use_bn)
self.pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
ds1 = nn.Conv2d(layer_dims[0], layer_dims[1], kernel_size=3, padding=1, stride=2)
self.res1 = BasicBlock(layer_dims[0], layer_dims[1], stride=2, downsample=ds1, use_bn=use_bn)
ds2 = nn.Conv2d(layer_dims[1], layer_dims[2], kernel_size=3, padding=1, stride=2)
self.res2 = BasicBlock(layer_dims[1], layer_dims[2], stride=2, downsample=ds2, use_bn=use_bn)
self.label_pred = conv_block(layer_dims[2], layer_dims[3], kernel_size=3, stride=1, padding=1,
relu=True, batch_norm=use_bn)
self.samp_w_pred = nn.Conv2d(layer_dims[2], layer_dims[3], kernel_size=3, padding=1, stride=1)
for m in self.modules():
if isinstance(m, nn.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
self.samp_w_pred.weight.data.fill_(0)
self.samp_w_pred.bias.data.fill_(1)
def forward(self, label_mask, feature=None):
# label_mask: frames, seq, h, w
assert label_mask.dim() == 4
label_shape = label_mask.shape
label_mask = label_mask.view(-1, 1, *label_mask.shape[-2:])
out = self.pool(self.conv_block(label_mask))
out = self.res2(self.res1(out))
label_enc = self.label_pred(out)
sample_w = self.samp_w_pred(out)
label_enc = label_enc.view(label_shape[0], label_shape[1], *label_enc.shape[-3:])
sample_w = sample_w.view(label_shape[0], label_shape[1], *sample_w.shape[-3:])
# Out dim is (num_seq, num_frames, layer_dims[-1], h, w)
return label_enc, sample_w
class ResidualDS16FeatSWBoxCatMultiBlock(nn.Module):
def __init__(self, layer_dims, feat_dim, use_final_relu=True, use_gauss=True, use_bn=True,
non_default_init=True, init_bn=1, gauss_scale=0.25, final_bn=True):
super().__init__()
in_layer_dim = (feat_dim+1,) + tuple(list(layer_dims)[:-2])
out_layer_dim = tuple(list(layer_dims)[:-1])
self.use_gauss = use_gauss
res = []
for in_d, out_d in zip(in_layer_dim, out_layer_dim):
ds = nn.Conv2d(in_d, out_d, kernel_size=3, padding=1, stride=1)
res.append(BasicBlock(in_d, out_d, stride=1, downsample=ds, use_bn=use_bn))
self.res = nn.Sequential(*res)
self.label_pred = conv_block(layer_dims[-2], layer_dims[-1], kernel_size=3, stride=1, padding=1,
relu=use_final_relu, batch_norm=final_bn)
self.gauss_scale = gauss_scale
if non_default_init:
for m in self.modules():
if isinstance(m, nn.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(init_bn)
m.bias.data.zero_()
def bbox_to_mask(self, bbox, sz):
mask = torch.zeros((bbox.shape[0],1,*sz), dtype=torch.float32, device=bbox.device)
for i, bb in enumerate(bbox):
x1, y1, w, h = list(map(int, bb))
x1 = int(x1+0.5)
y1 = int(y1+0.5)
h = int(h+0.5)
w = int(w+0.5)
mask[i, :, y1:(y1+h), x1:(x1+w)] = 1.0
return mask
def bbox_to_gauss(self, bbox, sz):
mask = torch.zeros((bbox.shape[0],1,*sz), dtype=torch.float32, device=bbox.device)
x_max, y_max = sz[-1], sz[-2]
for i, bb in enumerate(bbox):
x1, y1, w, h = list(map(int, bb))
cx, cy = x1+w/2, y1+h/2
xcoords = torch.arange(0, x_max).unsqueeze(dim=0).to(bbox.device).float()
ycoords = torch.arange(0, y_max).unsqueeze(dim=0).T.to(bbox.device).float()
d_xcoords = xcoords - cx
d_ycoords = ycoords - cy
dtotsqr = d_xcoords**2/(self.gauss_scale*w)**2 + d_ycoords**2/(self.gauss_scale*h)**2
mask[i,0] = torch.exp(-0.5*dtotsqr)
return mask
def forward(self, bb, feat, sz):
if self.use_gauss:
label_mask = self.bbox_to_gauss(bb, sz[-2:])
else:
label_mask = self.bbox_to_mask(bb, sz[-2:])
label_shape = label_mask.shape
label_mask = label_mask.view(-1, 1, *label_mask.shape[-2:])
feat = feat.view(-1, *feat.shape[-3:])
feat_mask_enc = torch.cat([feat, interpolate(label_mask, feat.shape[-2:])], dim=1)
out = self.res(feat_mask_enc)
label_enc = self.label_pred(out)
label_enc = label_enc.view(label_shape[0], label_shape[1], *label_enc.shape[-3:])
return label_enc
class ResidualDS16FeatSWBox(nn.Module):
def __init__(self, layer_dims, feat_dim, use_final_relu=True, use_gauss=True, use_bn=False, use_sample_w=True):
super().__init__()
self.use_sample_w = use_sample_w
self.use_gauss = use_gauss
self.conv_block = conv_block(1, layer_dims[0], kernel_size=3, stride=2, padding=1, batch_norm=use_bn)
self.pool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
ds1 = nn.Conv2d(layer_dims[0], layer_dims[1], kernel_size=3, padding=1, stride=2)
self.res1 = BasicBlock(layer_dims[0], layer_dims[1], stride=2, downsample=ds1, use_bn=use_bn)
ds2 = nn.Conv2d(layer_dims[1], layer_dims[2], kernel_size=3, padding=1, stride=2)
self.res2 = BasicBlock(layer_dims[1], layer_dims[2], stride=2, downsample=ds2, use_bn=use_bn)
ds3 = nn.Conv2d(layer_dims[2] + feat_dim, layer_dims[3], kernel_size=3, padding=1, stride=1)
self.res3 = BasicBlock(layer_dims[2] + feat_dim, layer_dims[3], stride=1, downsample=ds3, use_bn=use_bn)
self.label_pred = conv_block(layer_dims[3], layer_dims[4], kernel_size=3, stride=1, padding=1,
relu=use_final_relu)
if self.use_sample_w:
self.samp_w_pred = nn.Conv2d(layer_dims[3], layer_dims[4], kernel_size=3, padding=1, stride=1)
for m in self.modules():
if isinstance(m, nn.Conv2d):
n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels
m.weight.data.normal_(0, math.sqrt(2. / n))
elif isinstance(m, nn.BatchNorm2d):
m.weight.data.fill_(1)
m.bias.data.zero_()
if self.use_sample_w:
self.samp_w_pred.weight.data.fill_(0)
self.samp_w_pred.bias.data.fill_(1)
def bbox_to_mask(self, bbox, sz):
mask = torch.zeros((bbox.shape[0],1,*sz), dtype=torch.float32, device=bbox.device)
for i, bb in enumerate(bbox):
x1, y1, w, h = list(map(int, bb))
x1 = int(x1+0.5)
y1 = int(y1+0.5)
h = int(h+0.5)
w = int(w+0.5)
mask[i, :, max(0,y1):(y1+h), max(0,x1):(x1+w)] = 1.0
return mask
def bbox_to_gauss(self, bbox, sz):
mask = torch.zeros((bbox.shape[0],1,*sz), dtype=torch.float32, device=bbox.device)
x_max, y_max = sz[-1], sz[-2]
for i, bb in enumerate(bbox):
x1, y1, w, h = list(map(int, bb))
cx, cy = x1+w/2, y1+h/2
xcoords = torch.arange(0, x_max).unsqueeze(dim=0).to(bbox.device).float()
ycoords = torch.arange(0, y_max).unsqueeze(dim=0).T.to(bbox.device).float()
d_xcoords = xcoords - cx
d_ycoords = ycoords - cy
dtotsqr = d_xcoords**2/(0.25*w)**2 + d_ycoords**2/(0.25*h)**2
mask[i,0] = torch.exp(-0.5*dtotsqr)
return mask
def forward(self, bb, feat, sz):
assert bb.dim() == 3
num_frames = bb.shape[0]
batch_sz = bb.shape[1]
bb = bb.reshape(-1, 4)
if self.use_gauss:
label_mask = self.bbox_to_gauss(bb, sz[-2:])
else:
label_mask = self.bbox_to_mask(bb, sz[-2:])
label_mask = label_mask.view(-1, 1, *label_mask.shape[-2:])
mask_enc = self.pool(self.conv_block(label_mask))
mask_enc = self.res2(self.res1(mask_enc))
feat = feat.view(-1, *feat.shape[-3:])
feat_mask_enc = torch.cat((mask_enc, feat), dim=1)
out = self.res3(feat_mask_enc)
label_enc = self.label_pred(out)
label_enc = label_enc.view(num_frames, batch_sz, *label_enc.shape[-3:])
sample_w = None
if self.use_sample_w:
sample_w = self.samp_w_pred(out)
sample_w = sample_w.view(num_frames, batch_sz, *sample_w.shape[-3:])
# Out dim is (num_seq, num_frames, layer_dims[-1], h, w)
return label_enc, sample_w