/
mmd.py
137 lines (117 loc) · 4.6 KB
/
mmd.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
import os
os.environ['MXNET_CUDNN_AUTOTUNE_DEFAULT'] = '0'
import sys
import numpy as np
import cv2
import random
import json
import mxnet as mx
import math
def calculate_distance2(data1, data2):
#total_num = batch_size
_data1 = mx.symbol.expand_dims(data1, axis=1) # B,1,C
spread_distance2 = mx.symbol.broadcast_sub(_data1, data2) #B,B,C
spread_distance2 = mx.symbol.reshape(spread_distance2, shape=(-3,-2)) #B*B,C
spread_distance2 = mx.symbol.square(spread_distance2)
distance2 = mx.symbol.sum(spread_distance2, axis=1) # B*B,
return distance2
def calculate_distance1(data1, data2):
distance = data1 - data2
#distance = mx.symbol.reshape(distance, shape=(1,-2))
distance2 = mx.symbol.square(distance) #1,C
distance2 = mx.symbol.sum(distance2, axis=1) #1,
return distance2
def multi_kernel_distance(data1, data2, batch_size, kernel_num, base_gamma):
assert batch_size==1
#distance2 = calculate_distance2(data1, data2)
distance2 = calculate_distance1(data1, data2)
kernel_mul = 2.0
coef = kernel_num*-0.5
#times = math.pow(kernel_mul, kernel_num/2.0)
ks = []
for i in xrange(kernel_num):
kernel_gamma = base_gamma * math.pow(kernel_mul, coef)
ks.append(mx.symbol.exp(distance2*-1.0/kernel_gamma))
coef += 1.0
ret = mx.symbol.add_n(*ks) #(b*b,)
return ret
def multi_kernel_distance2(distance2, kernel_num, base_gamma):
kernel_mul = 2.0
coef = kernel_num*-0.5
ks = []
for i in xrange(kernel_num):
kernel_gamma = base_gamma * math.pow(kernel_mul, coef)
ks.append(mx.symbol.exp(distance2*-1.0/kernel_gamma))
coef += 1.0
ret = mx.symbol.add_n(*ks) #(b*b,)
return ret
def mmd(data, fc, args):
batch_size = args.batch_per_gpu
print('mmd batch_size', batch_size)
assert batch_size%4==0
#static params
source_num = batch_size/2
target_num = source_num
total_num = batch_size
data_kernel_num = 4
label_kernel_num = 1
group_num = source_num/2
#data_gamma = 0.0
label_gamma = 1.3
softmax = mx.symbol.softmax(data=fc)
gt_label = mx.symbol.Variable('softmax_label')
#max_label = mx.symbol.max(gt_label)
#source_data = mx.symbol.slice_axis(data, axis=0, begin=0, end=source_num)
#target_data = mx.symbol.slice_axis(data, axis=0, begin=source_num, end=total_num)
#source_softmax = mx.symbol.slice_axis(softmax, axis=0, begin=0, end=source_num)
#target_softmax = mx.symbol.slice_axis(softmax, axis=0, begin=source_num, end=total_num)
distance2 = calculate_distance2(data, data)
bandwidth = mx.symbol.sum(distance2)
#data_gamma = (total_num * total_num - total_num)/bandwidth
data_gamma = bandwidth/(total_num * total_num - total_num)
k_list = []
#unbiased mmd
for i in xrange(group_num):
xs1 = mx.symbol.slice_axis(data, axis=0, begin=i*2, end=i*2+1)
xs2 = mx.symbol.slice_axis(data, axis=0, begin=i*2+1, end=i*2+2)
xt1 = mx.symbol.slice_axis(data, axis=0, begin=source_num+i*2, end=source_num+i*2+1)
xt2 = mx.symbol.slice_axis(data, axis=0, begin=source_num+i*2+1, end=source_num+i*2+2)
ys1 = mx.symbol.slice_axis(softmax, axis=0, begin=i*2, end=i*2+1)
ys2 = mx.symbol.slice_axis(softmax, axis=0, begin=i*2+1, end=i*2+2)
yt1 = mx.symbol.slice_axis(softmax, axis=0, begin=source_num+i*2, end=source_num+i*2+1)
yt2 = mx.symbol.slice_axis(softmax, axis=0, begin=source_num+i*2+1, end=source_num+i*2+2)
k_x = multi_kernel_distance(xs1, xs2, 1, data_kernel_num, data_gamma)
k_y = multi_kernel_distance(ys1, ys2, 1, label_kernel_num, label_gamma)
k = k_x*k_y
if args.use_dan:
k_list.append(k_x)
else:
k_list.append(k)
k_x = multi_kernel_distance(xt1, xt2, 1, data_kernel_num, data_gamma)
k_y = multi_kernel_distance(yt1, yt2, 1, label_kernel_num, label_gamma)
k = k_x*k_y
if args.use_dan:
k_list.append(k_x)
else:
k_list.append(k)
k_x = multi_kernel_distance(xs1, xt2, 1, data_kernel_num, data_gamma)*-1.0
k_y = multi_kernel_distance(ys1, yt2, 1, label_kernel_num, label_gamma)
k = k_x*k_y
if args.use_dan:
k_list.append(k_x)
else:
k_list.append(k)
k_x = multi_kernel_distance(xt1, xs2, 1, data_kernel_num, data_gamma)*-1.0
k_y = multi_kernel_distance(yt1, ys2, 1, label_kernel_num, label_gamma)
k = k_x*k_y
if args.use_dan:
k_list.append(k_x)
else:
k_list.append(k)
mmd_loss = mx.symbol.add_n(*k_list)/group_num
net = mx.symbol.SoftmaxOutput(data=fc, label = gt_label, use_ignore=True, ignore_label=args.null_label, name='softmax')
if args.train_stage>0:
grad_scale = 1.0 if args.use_dan else 2.0
mmd = mx.symbol.MakeLoss(mmd_loss, grad_scale=grad_scale)
net = mx.symbol.Group([net,mmd])
return net