/
symbol_octConv.py
162 lines (116 loc) · 9.88 KB
/
symbol_octConv.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
import mxnet as mx
from symbol_basic import *
def firstOctConv(data, settings, ch_in, ch_out, name, kernel=(1,1), pad=(0,0), stride=(1,1)):
alpha_in, alpha_out = settings
hf_ch_in = int(ch_in * (1 - alpha_in))
hf_ch_out = int(ch_out * (1 - alpha_out))
lf_ch_in = ch_in - hf_ch_in
lf_ch_out = ch_out - hf_ch_out
hf_data = data
if stride == (2, 2):
hf_data = mx.symbol.Pooling(data=hf_data, pool_type='avg', kernel=(2,2), stride=(2,2), name=('%s_hf_down' % name))
hf_conv = Conv(data=hf_data, num_filter=hf_ch_out, kernel=kernel, pad=pad, stride=(1,1), name=('%s_hf_conv' % name))
hf_pool = mx.symbol.Pooling(data=hf_data, pool_type='avg', kernel=(2,2), stride=(2,2), name=('%s_hf_pool' % name))
hf_pool_conv = Conv(data=hf_pool, num_filter=lf_ch_out, kernel=kernel, pad=pad, stride=(1,1), name=('%s_hf_pool_conv' % name))
out_h = hf_conv
out_l = hf_pool_conv
return out_h, out_l
def lastOctConv(hf_data, lf_data, settings, ch_in, ch_out, name, kernel=(1,1), pad=(0,0), stride=(1,1)):
alpha_in, alpha_out = settings
hf_ch_in = int(ch_in * (1 - alpha_in))
hf_ch_out = int(ch_out * (1 - alpha_out))
if stride == (2, 2):
hf_data = mx.symbol.Pooling(data=hf_data, pool_type='avg', kernel=(2,2), stride=(2,2), name=('%s_hf_down' % name))
hf_conv = Conv(data=hf_data, num_filter=hf_ch_out, kernel=kernel, pad=pad, stride=(1,1), name=('%s_hf_conv' % name))
lf_conv = Conv(data=lf_data, num_filter=hf_ch_out, kernel=kernel, pad=pad, stride=(1,1), name=('%s_lf_conv' % name))
out_h = hf_conv + lf_conv
return out_h
def OctConv(hf_data, lf_data, settings, ch_in, ch_out, name, kernel=(1,1), pad=(0,0), stride=(1,1)):
alpha_in, alpha_out = settings
hf_ch_in = int(ch_in * (1 - alpha_in))
hf_ch_out = int(ch_out * (1 - alpha_out))
lf_ch_in = ch_in - hf_ch_in
lf_ch_out = ch_out - hf_ch_out
if stride == (2, 2):
hf_data = mx.symbol.Pooling(data=hf_data, pool_type='avg', kernel=(2,2), stride=(2,2), name=('%s_hf_down' % name))
hf_conv = Conv(data=hf_data, num_filter=hf_ch_out, kernel=kernel, pad=pad, stride=(1,1), name=('%s_hf_conv' % name))
hf_pool = mx.symbol.Pooling(data=hf_data, pool_type='avg', kernel=(2,2), stride=(2,2), name=('%s_hf_pool' % name))
hf_pool_conv = Conv(data=hf_pool, num_filter=lf_ch_out, kernel=kernel, pad=pad, stride=(1,1), name=('%s_hf_pool_conv' % name))
lf_conv = Conv(data=lf_data, num_filter=hf_ch_out, kernel=kernel, pad=pad, stride=(1,1), name=('%s_lf_conv' % name))
if stride == (2, 2):
lf_upsample = lf_conv
lf_down = mx.symbol.Pooling(data=lf_data, pool_type='avg', kernel=(2,2), stride=(2,2), name=('%s_lf_down' % name))
else:
lf_upsample = mx.symbol.UpSampling(lf_conv, scale=2, sample_type='nearest',num_args=1, name='%s_lf_upsample' % name)
lf_down = lf_data
lf_down_conv = Conv(data=lf_down, num_filter=lf_ch_out, kernel=kernel, pad=pad, stride=(1,1), name=('%s_lf_down_conv' % name))
out_h = hf_conv + lf_upsample
out_l = hf_pool_conv + lf_down_conv
return out_h, out_l
def firstOctConv_BN_AC(data, alpha, num_filter_in, num_filter_out, kernel, pad, stride=(1,1), name=None, w=None, b=None, no_bias=True, attr=None, num_group=1):
hf_data, lf_data = firstOctConv(data=data, settings=(0, alpha), ch_in=num_filter_in, ch_out=num_filter_out, name=name, kernel=kernel, pad=pad, stride=stride)
out_hf = BN_AC(data=hf_data, name=('%s_hf') % name)
out_lf = BN_AC(data=lf_data, name=('%s_lf') % name)
return out_hf, out_lf
def lastOctConv_BN_AC(hf_data, lf_data, alpha, num_filter_in, num_filter_out, kernel, pad, stride=(1,1), name=None, w=None, b=None, no_bias=True, attr=None, num_group=1):
conv = lastOctConv(hf_data=hf_data, lf_data=lf_data, settings=(alpha, 0), ch_in=num_filter_in, ch_out=num_filter_out, name=name, kernel=kernel, pad=pad, stride=stride)
out = BN_AC(data=conv, name=name)
return out
def octConv_BN_AC(hf_data, lf_data, alpha, num_filter_in, num_filter_out, kernel, pad, stride=(1,1), name=None, w=None, b=None, no_bias=True, attr=None, num_group=1):
hf_data, lf_data = OctConv(hf_data=hf_data, lf_data=lf_data, settings=(alpha, alpha), ch_in=num_filter_in, ch_out=num_filter_out, name=name, kernel=kernel, pad=pad, stride=stride)
out_hf = BN_AC(data=hf_data, name=('%s_hf') % name)
out_lf = BN_AC(data=lf_data, name=('%s_lf') % name)
return out_hf, out_lf
def firstOctConv_BN(data, alpha, num_filter_in, num_filter_out, kernel, pad, stride=(1,1), name=None, w=None, b=None, no_bias=True, attr=None, num_group=1):
hf_data, lf_data = firstOctConv(data=data, settings=(0, alpha), ch_in=num_filter_in, ch_out=num_filter_out, name=name, kernel=kernel, pad=pad, stride=stride)
out_hf = BN(data=hf_data, name=('%s_hf') % name)
out_lf = BN(data=lf_data, name=('%s_lf') % name)
return out_hf, out_lf
def lastOctConv_BN(hf_data, lf_data, alpha, num_filter_in, num_filter_out, kernel, pad, stride=(1,1), name=None, w=None, b=None, no_bias=True, attr=None, num_group=1):
conv = lastOctConv(hf_data=hf_data, lf_data=lf_data, settings=(alpha, 0), ch_in=num_filter_in, ch_out=num_filter_out, name=name, kernel=kernel, pad=pad, stride=stride)
out = BN(data=conv, name=name)
return out
def octConv_BN(hf_data, lf_data, alpha, num_filter_in, num_filter_out, kernel, pad, stride=(1,1), name=None, w=None, b=None, no_bias=True, attr=None, num_group=1):
hf_data, lf_data = OctConv(hf_data=hf_data, lf_data=lf_data, settings=(alpha, alpha), ch_in=num_filter_in, ch_out=num_filter_out, name=name, kernel=kernel, pad=pad, stride=stride)
out_hf = BN(data=hf_data, name=('%s_hf') % name)
out_lf = BN(data=lf_data, name=('%s_lf') % name)
return out_hf, out_lf
def Residual_Unit_norm(data, num_in, num_mid, num_out, name, first_block=False, stride=(1, 1), g=1):
conv_m1 = Conv_BN_AC( data=data, num_filter=num_mid, kernel=( 1, 1), pad=( 0, 0), name=('%s_conv-m1' % name))
conv_m2 = Conv_BN_AC( data=conv_m1, num_filter=num_mid, kernel=( 3, 3), pad=( 1, 1), name=('%s_conv-m2' % name), stride=stride, num_group=g)
conv_m3 = Conv_BN( data=conv_m2, num_filter=num_out, kernel=( 1, 1), pad=( 0, 0), name=('%s_conv-m3' % name))
if first_block:
data = Conv_BN( data=data, num_filter=num_out, kernel=( 1, 1), pad=( 0, 0), name=('%s_conv-w1' % name), stride=stride)
outputs = mx.symbol.ElementWiseSum(*[data, conv_m3], name=('%s_sum' % name))
return AC(outputs)
def Residual_Unit_last(hf_data, lf_data, alpha, num_in, num_mid, num_out, name, first_block=False, stride=(1, 1), g=1):
hf_data_m, lf_data_m = octConv_BN_AC( hf_data=hf_data, lf_data=lf_data, alpha=alpha, num_filter_in=num_in, num_filter_out=num_mid, kernel=( 1, 1), pad=( 0, 0), name=('%s_conv-m1' % name))
conv_m2 = lastOctConv_BN_AC(hf_data=hf_data_m, lf_data=lf_data_m, alpha=alpha, num_filter_in=num_mid, num_filter_out=num_mid, name=('%s_conv-m2' % name), kernel=(3,3), pad=(1,1), stride=stride)
conv_m3 = Conv_BN( data=conv_m2, num_filter=num_out, kernel=( 1, 1), pad=( 0, 0), name=('%s_conv-m3' % name))
if first_block:
data = lastOctConv_BN(hf_data=hf_data, lf_data=lf_data, alpha=alpha, num_filter_in=num_in, num_filter_out=num_out, name=('%s_conv-w1' % name), kernel=(1,1), pad=(0,0), stride=stride)
outputs = mx.symbol.ElementWiseSum(*[data, conv_m3], name=('%s_sum' % name))
outputs = AC(outputs, name=('%s_act' % name))
return outputs
def Residual_Unit_first(data, alpha, num_in, num_mid, num_out, name, first_block=False, stride=(1, 1), g=1):
hf_data_m, lf_data_m = firstOctConv_BN_AC(data=data, alpha=alpha, num_filter_in=num_in, num_filter_out=num_mid, kernel=( 1, 1), pad=( 0, 0), name=('%s_conv-m1' % name))
hf_data_m, lf_data_m = octConv_BN_AC( hf_data=hf_data_m, lf_data=lf_data_m, alpha=alpha, num_filter_in=num_mid, num_filter_out=num_mid, kernel=( 3, 3), pad=( 1, 1), name=('%s_conv-m2' % name), stride=stride, num_group=g)
hf_data_m, lf_data_m = octConv_BN( hf_data=hf_data_m, lf_data=lf_data_m, alpha=alpha, num_filter_in=num_mid, num_filter_out=num_out, kernel=( 1, 1), pad=( 0, 0), name=('%s_conv-m3' % name))
if first_block:
hf_data, lf_data = firstOctConv_BN( data=data, alpha=alpha, num_filter_in=num_in, num_filter_out=num_out, kernel=( 1, 1), pad=( 0, 0), name=('%s_conv-w1' % name), stride=stride)
hf_outputs = mx.symbol.ElementWiseSum(*[hf_data, hf_data_m], name=('%s_hf_sum' % name))
lf_outputs = mx.symbol.ElementWiseSum(*[lf_data, lf_data_m], name=('%s_lf_sum' % name))
hf_outputs = AC(hf_outputs, name=('%s_hf_act' % name))
lf_outputs = AC(lf_outputs, name=('%s_lf_act' % name))
return hf_outputs, lf_outputs
def Residual_Unit(hf_data, lf_data, alpha, num_in, num_mid, num_out, name, first_block=False, stride=(1, 1), g=1):
hf_data_m, lf_data_m = octConv_BN_AC( hf_data=hf_data, lf_data=lf_data, alpha=alpha, num_filter_in=num_in, num_filter_out=num_mid, kernel=( 1, 1), pad=( 0, 0), name=('%s_conv-m1' % name))
hf_data_m, lf_data_m = octConv_BN_AC( hf_data=hf_data_m, lf_data=lf_data_m, alpha=alpha, num_filter_in=num_mid, num_filter_out=num_mid, kernel=( 3, 3), pad=( 1, 1), name=('%s_conv-m2' % name), stride=stride, num_group=g)
hf_data_m, lf_data_m = octConv_BN( hf_data=hf_data_m, lf_data=lf_data_m, alpha=alpha, num_filter_in=num_mid, num_filter_out=num_out, kernel=( 1, 1), pad=( 0, 0), name=('%s_conv-m3' % name))
if first_block:
hf_data, lf_data = octConv_BN( hf_data=hf_data, lf_data=lf_data, alpha=alpha, num_filter_in=num_in, num_filter_out=num_out, kernel=( 1, 1), pad=( 0, 0), name=('%s_conv-w1' % name), stride=stride)
hf_outputs = mx.symbol.ElementWiseSum(*[hf_data, hf_data_m], name=('%s_hf_sum' % name))
lf_outputs = mx.symbol.ElementWiseSum(*[lf_data, lf_data_m], name=('%s_lf_sum' % name))
hf_outputs = AC(hf_outputs, name=('%s_hf_act' % name))
lf_outputs = AC(lf_outputs, name=('%s_lf_act' % name))
return hf_outputs, lf_outputs