forked from n1ckfg/deep3d
-
Notifications
You must be signed in to change notification settings - Fork 0
/
sym.py
118 lines (109 loc) · 6.5 KB
/
sym.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
import mxnet as mx
def make_upsample_sym(data, scale, fuse='sum', method='multi2'):
# group 1
conv1_1 = mx.symbol.Convolution(data=data, kernel=(3, 3), pad=(1, 1), num_filter=64, name="conv1_1")
relu1_1 = mx.symbol.Activation(data=conv1_1, act_type="relu", name="relu1_1")
pool1 = mx.symbol.Pooling(
data=relu1_1, pool_type="max", kernel=(2, 2), stride=(2,2), name="pool1")
# group 2
conv2_1 = mx.symbol.Convolution(
data=pool1, kernel=(3, 3), pad=(1, 1), num_filter=128, name="conv2_1")
relu2_1 = mx.symbol.Activation(data=conv2_1, act_type="relu", name="relu2_1")
pool2 = mx.symbol.Pooling(
data=relu2_1, pool_type="max", kernel=(2, 2), stride=(2,2), name="pool2")
# group 3
conv3_1 = mx.symbol.Convolution(
data=pool2, kernel=(3, 3), pad=(1, 1), num_filter=256, name="conv3_1")
relu3_1 = mx.symbol.Activation(data=conv3_1, act_type="relu", name="relu3_1")
conv3_2 = mx.symbol.Convolution(
data=relu3_1, kernel=(3, 3), pad=(1, 1), num_filter=256, name="conv3_2")
relu3_2 = mx.symbol.Activation(data=conv3_2, act_type="relu", name="relu3_2")
pool3 = mx.symbol.Pooling(
data=relu3_2, pool_type="max", kernel=(2, 2), stride=(2,2), name="pool3")
# group 4
conv4_1 = mx.symbol.Convolution(
data=pool3, kernel=(3, 3), pad=(1, 1), num_filter=512, name="conv4_1")
relu4_1 = mx.symbol.Activation(data=conv4_1, act_type="relu", name="relu4_1")
conv4_2 = mx.symbol.Convolution(
data=relu4_1, kernel=(3, 3), pad=(1, 1), num_filter=512, name="conv4_2")
relu4_2 = mx.symbol.Activation(data=conv4_2, act_type="relu", name="relu4_2")
pool4 = mx.symbol.Pooling(
data=relu4_2, pool_type="max", kernel=(2, 2), stride=(2,2), name="pool4")
# group 5
conv5_1 = mx.symbol.Convolution(
data=pool4, kernel=(3, 3), pad=(1, 1), num_filter=512, name="conv5_1")
relu5_1 = mx.symbol.Activation(data=conv5_1, act_type="relu", name="relu5_1")
conv5_2 = mx.symbol.Convolution(
data=relu5_1, kernel=(3, 3), pad=(1, 1), num_filter=512, name="conv5_2")
relu5_2 = mx.symbol.Activation(data=conv5_2, act_type="relu", name="conv1_2")
pool5 = mx.symbol.Pooling(
data=relu5_2, pool_type="max", kernel=(2, 2), stride=(2,2), name="pool5")
# group 6
flatten = mx.symbol.Flatten(data=pool5, name="flatten")
fc6 = mx.symbol.FullyConnected(data=flatten, num_hidden=512, name="fc6")
relu6 = mx.symbol.Activation(data=fc6, act_type="relu", name="relu6")
drop6 = mx.symbol.Dropout(data=relu6, p=0.5, name="drop6")
# group 7
fc7 = mx.symbol.FullyConnected(data=drop6, num_hidden=512, name="fc7")
relu7 = mx.symbol.Activation(data=fc7, act_type="relu", name="relu7")
drop7 = mx.symbol.Dropout(data=relu7, p=0.5, name="drop7")
# output
fc8 = mx.symbol.FullyConnected(data=drop7, num_hidden=33*12*5, name="pred5")
pred5 = mx.symbol.Reshape(data=fc8, target_shape=(0, 33, 5, 12))
bn_pool4 = mx.sym.CuDNNBatchNorm(data=pool4, name='bn_pool4')
pred4 = mx.symbol.Convolution(data=bn_pool4, kernel=(3,3), pad=(1,1), num_filter=33, name="pred4")
bn_pool3 = mx.sym.CuDNNBatchNorm(data=pool3, name='bn_pool3')
pred3 = mx.symbol.Convolution(data=bn_pool3, kernel=(3,3), pad=(1,1), num_filter=33, name="pred3")
bn_pool2 = mx.sym.CuDNNBatchNorm(data=pool2, name='bn_pool2')
pred2 = mx.symbol.Convolution(data=bn_pool2, kernel=(3,3), pad=(1,1), num_filter=33, name="pred2")
bn_pool1 = mx.sym.CuDNNBatchNorm(data=pool1, name='bn_pool1')
pred1 = mx.symbol.Convolution(data=bn_pool1, kernel=(3,3), pad=(1,1), num_filter=33, name="pred1")
no_bias = False
assert fuse == 'sum'
assert method == 'multi2'
workspace = 0
scale = 1
pred1 = mx.symbol.Activation(data=pred1, act_type='relu')
pred1 = mx.symbol.Deconvolution(data=pred1, kernel=(1, 1), stride=(1, 1), pad=(0, 0), num_filter=33, no_bias=no_bias, workspace=workspace, name='deconv_pred1')
scale *= 2
pred2 = mx.symbol.Activation(data=pred2, act_type='relu')
pred2 = mx.symbol.Deconvolution(data=pred2, kernel=(2*scale, 2*scale), stride=(scale, scale), pad=(scale/2, scale/2), num_filter=33, no_bias=no_bias, workspace=workspace, name='deconv_pred2')
scale *= 2
pred3 = mx.symbol.Activation(data=pred3, act_type='relu')
pred3 = mx.symbol.Deconvolution(data=pred3, kernel=(2*scale, 2*scale), stride=(scale, scale), pad=(scale/2, scale/2), num_filter=33, no_bias=no_bias, workspace=workspace, name='deconv_pred3')
scale *= 2
pred4 = mx.symbol.Activation(data=pred4, act_type='relu')
pred4 = mx.symbol.Deconvolution(data=pred4, kernel=(2*scale, 2*scale), stride=(scale, scale), pad=(scale/2, scale/2), num_filter=33, no_bias=no_bias, workspace=workspace, name='deconv_pred4')
scale *= 2
pred5 = mx.symbol.Activation(data=pred5, act_type='relu')
pred5 = mx.symbol.Deconvolution(data=pred5, kernel=(2*scale, 2*scale), stride=(scale, scale), pad=(scale/2, scale/2), num_filter=33, no_bias=no_bias, workspace=workspace, name='deconv_pred5')
feat = mx.symbol.ElementWiseSum(pred1, pred2, pred3, pred4, pred5)
feat_act = mx.symbol.Activation(data=feat, act_type='relu', name='feat_relu')
scale = 2
up = mx.symbol.Deconvolution(data=feat_act, kernel=(2*scale, 2*scale), stride=(scale, scale), pad=(scale/2, scale/2), num_filter=33, no_bias=no_bias, workspace=workspace, name='deconv_predup')
up = mx.symbol.Activation(data=up, act_type='relu')
up = mx.symbol.Convolution(data=up, kernel=(3,3), pad=(1,1), num_filter=33)
return up, [pred1, pred2, pred3, pred4, pred5]
def make_l1_sym(scale, fuse='sum', data_frames=1, flow_frames=1, get_softmax=False, method='multi2', upsample=1, tan=False):
left = mx.symbol.Variable(name="left")
flow = mx.symbol.Variable(name="flow")
left0 = mx.symbol.Variable(name='left0')
if data_frames and flow_frames:
data = mx.symbol.Concat(left, flow, dim=1, name='data')
elif data_frames:
data = left
elif flow_frames:
data = flow
data, _ = make_upsample_sym(data, scale, fuse, method=method)
if tan:
data = mx.symbol.Activation(data=data, act_type='tanh')
if method == 'direct':
depthdot = data
else:
softmax = mx.symbol.SoftmaxActivation(data=data, type='channel', name='softmax')
depthdot = mx.symbol.DepthDot(data=softmax, label=left0, scale=scale, upsample=upsample, name='depth')
loss = mx.symbol.MAERegressionOutput(data=depthdot, name='l1')
if get_softmax:
return mx.symbol.Group([depthdot, softmax])
else:
return loss