forked from razor1179/pytorch-kaldi-CGS
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathPattern.py
executable file
·218 lines (199 loc) · 10.2 KB
/
Pattern.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
'''
Description:
version:
Author: Wang Yanhong
email: [email protected]
Date: 2020-10-20 06:22:15
LastEditors: Wang Yanhong
LastEditTime: 2020-11-09 12:28:11
'''
import math
import os
import torch
from torch.nn.parameter import Parameter
from torch.nn import functional as F
from torch.nn.modules.module import Module
import hcgs
import guided_hcgs
import scipy.io as sio
import numpy as np
from data_io import read_mat
from sparsity import sparsity
class Pattern(Module):
"""Creates HCGS layer
Args:
in_features: size of each input sample
out_features: size of each output sample
Attributes:
mask: the non-learnable weights of the module of shape
`(out_features x in_features)`
"""
def __init__(self, dense_features, pattern_mode, pattern_shape, pattern_nnz, pattern_num, save_dir, pattern_from_file=False):
super(Pattern, self).__init__()
self.dense_features = dense_features
self.pattern_from_file = pattern_from_file
self.pattern_mode = pattern_mode
self.pattern_shape = np.array(pattern_shape)
self.pattern_nnz = int(pattern_nnz)
self.pattern_num = int(pattern_num)
self.mask = self.get_mask()
def get_mask(self, dense=True):
if self.pattern_mode == 'pattern_from_weight':
self.pattern = self.update_pattern_by_weight()
return(Parameter(torch.ones(self.dense_features.shape)))
elif self.pattern_mode == 'pattern':
return self.pattern_mask()
elif self.pattern_mode == 'coo':
return self.coo_mask()
elif self.pattern_mode == 'pattern_coo':
return self.pattern_coo_mask()
else:
return(Parameter(torch.ones(self.dense_features.shape)))
def get_pattern(self, pattern_nnz):
# block_num = self.dense_features//self.block_size
if self.pattern_from_file:
if not os.path.exists(pattern_from_file):
print("Error: Pattern file \"%s\" does not exists!" %
pattern_from_file)
exit()
pattern = np.load(pattern_from_file)
pattern = np.split(pattern, self.pattern_num, 0)
else:
pattern = []
for i in range(self.pattern_num):
mask = np.zeros(np.prod(self.pattern_shape))
mask[np.random.choice(mask.shape[0], pattern_nnz)] = 1
pattern.append(mask.reshape(self.pattern_shape))
return pattern
def pattern_mask(self):
self.pattern = self.get_pattern(self.pattern_nnz)
assert self.dense_features.shape[0] % self.pattern_shape[
0] == 0, f'Error:{self.dense_features.shape[0]} can not be divisible by {self.pattern_shape[0]}'
assert self.dense_features.shape[1] % self.pattern_shape[
1] == 0, f'Error:{self.dense_features.shape[1]} can not be divisible by {self.pattern_shape[1]}'
self.mask = np.zeros(self.dense_features.shape)
row_block_num = self.dense_features.shape[0]//self.pattern_shape[0]
col_block_num = self.dense_features.shape[1]//self.pattern_shape[1]
for i in range(row_block_num):
for j in range(col_block_num):
self.mask[i*self.pattern_shape[0]:(i+1)*self.pattern_shape[0],
j*self.pattern_shape[1]:(j+1)*self.pattern_shape[1]] = \
self.pattern[np.random.choice(self.pattern_num, 1)[0]]
return(Parameter(torch.from_numpy(self.mask)))
def coo_mask(self):
assert self.dense_features.shape[0] % self.pattern_shape[
0] == 0, f'Error:{self.dense_features.shape[0]} can not be divisible by {self.pattern_shape[0]}'
assert self.dense_features.shape[1] % self.pattern_shape[
1] == 0, f'Error:{self.dense_features.shape[1]} can not be divisible by {self.pattern_shape[1]}'
self.mask = np.zeros(self.dense_features.shape)
row_block_num = self.dense_features.shape[0]//self.pattern_shape[0]
col_block_num = self.dense_features.shape[1]//self.pattern_shape[1]
for i in range(row_block_num):
for j in range(col_block_num):
dense_features_block = self.dense_features[i*self.pattern_shape[0]:(i+1)*self.pattern_shape[0],
j*self.pattern_shape[1]:(j+1)*self.pattern_shape[1]].flatten()
mask_block = np.zeros(dense_features_block.shape)
mask_block[np.argsort(
np.abs(dense_features_block))[-self.pattern_nnz:]] = 1
self.mask[i*self.pattern_shape[0]:(i+1)*self.pattern_shape[0],
j*self.pattern_shape[1]:(j+1)*self.pattern_shape[1]] = mask_block.reshape(self.pattern_shape)
return(Parameter(torch.from_numpy(self.mask)))
def pattern_coo_mask(self):
assert self.dense_features.shape[0] % self.pattern_shape[
0] == 0, f'Error:{self.dense_features.shape[0]} can not be divisible by {self.pattern_shape[0]}'
assert self.dense_features.shape[1] % self.pattern_shape[
1] == 0, f'Error:{self.dense_features.shape[1]} can not be divisible by {self.pattern_shape[1]}'
row_block_num = self.dense_features.shape[0]//self.pattern_shape[0]
col_block_num = self.dense_features.shape[1]//self.pattern_shape[1]
self.mask = np.zeros(self.dense_features.shape)
self.pat_nnz = math.ceil(self.pattern_nnz/2)
self.coo_nnz = self.pattern_nnz - self.pat_nnz
self.pattern = self.get_pattern(self.pat_nnz)
for i in range(row_block_num):
for j in range(col_block_num):
mask_block = self.pattern[np.random.choice(
self.pattern_num, 1)[0]].flatten()
dense_features_block = self.dense_features[i*self.pattern_shape[0]:(i+1)*self.pattern_shape[0],
j*self.pattern_shape[1]:(j+1)*self.pattern_shape[1]].flatten()
mask_block[np.argsort(np.abs(np.array(
dense_features_block)*(np.ones_like(mask_block) - mask_block)))[-self.coo_nnz:]] = 1
self.mask[i*self.pattern_shape[0]:(i+1)*self.pattern_shape[0],
j*self.pattern_shape[1]:(j+1)*self.pattern_shape[1]] = mask_block.reshape(self.pattern_shape)
return(Parameter(torch.from_numpy(self.mask)))
def update(self, dense_features):
self.dense_features = dense_features
self.mask = Parameter(hcgs.conn_mat(
out_features, in_features, block_sizes[:], drop_ratios[:], des))
def read_mat(self, file):
""" [mat] = read_mat(file_or_fd)
Reads single kaldi matrix, supports ascii and binary.
file_or_fd : file, gzipped file, pipe or opened file descriptor.
"""
fd = open(file)
try:
binary = fd.read(2).decode()
if binary == '\0B':
mat = _read_mat_binary(fd)
else:
assert (binary == ' [')
mat = _read_mat_ascii(fd)
finally:
fd.close()
return mat
def update_pattern_by_weight(self):
pattern_candidates = sparsity.generate_complete_pattern_set(
self.pattern_shape, self.pattern_nnz)
self.pattern = sparsity.find_top_k_by_similarity(
self.dense_features, pattern_candidates, stride=self.pattern_shape, pattern_num=self.pattern_nnz)
for i, p in enumerate(self.pattern):
print("top", i, len(self.pattern))
print(p)
def update_mask(self):
self.mask = sparsity.apply_patterns(self.dense_features, self.pattern)
# def update_pattern_by_weight(self, input, prune_perc):
# assert self.dense_features.shape[0] % self.pattern_shape[
# 0] == 0, f'Error:{self.dense_features.shape[0]} can not be divisible by {self.pattern_shape[0]}'
# assert self.dense_features.shape[1] % self.pattern_shape[
# 1] == 0, f'Error:{self.dense_features.shape[1]} can not be divisible by {self.pattern_shape[1]}'
# row_block_num = self.dense_features.shape[0]//self.pattern_shape[0]
# col_block_num = self.dense_features.shape[1]//self.pattern_shape[1]
# pattern_candidates = list()
# pattern_scores = list()
# pattern_len = self.pattern_shape[0]*self.pattern_shape[1]
# for i in range(pattern_len):
# for j in range(pattern_len):
# if not i == j:
# pattern = torch.zeros(self.pattern_shape).flatten()
# pattern[i] = 1
# pattern[j] = 1
# pattern_score = cal_pattern_score(input, pattern)
# pattern_candidates.append(pattern)
# pattern_scores.append(pattern_score)
# sorted_patterns = sorted(
# pattern_candidates, key=lambda i: pattern_scores[i], reverse=True)
# def cal_pattern_score(self, input, pattern):
# pattern_score = 0
# assert self.dense_features.shape[0] % self.pattern_shape[
# 0] == 0, f'Error:{self.dense_features.shape[0]} can not be divisible by {self.pattern_shape[0]}'
# assert self.dense_features.shape[1] % self.pattern_shape[
# 1] == 0, f'Error:{self.dense_features.shape[1]} can not be divisible by {self.pattern_shape[1]}'
# row_block_num = self.dense_features.shape[0]//self.pattern_shape[0]
# col_block_num = self.dense_features.shape[1]//self.pattern_shape[1]
# for i in range(row_block_num):
# for j in range(col_block_num):
# pattern_score += pattern * input[i*self.pattern_shape[0]:(i+1)*self.pattern_shape[0],
# j*self.pattern_shape[1]:(j+1)*self.pattern_shape[1]].flatten()
# return pattern_score
# for i in range(16):
# mask = np.zeros(64)
# mask[i*4:(i+1)*4]=1
# pattern.append(mask.reshape(8,8))
# np.save('pattern_file/b08b08_k04_n16_pattern_v1.npy',np.concatenate(pattern,0))
# pattern=[]
# for i in range(16):
# mask = np.zeros(64)
# # mask[i*4:(i+1)*4]=1
# start = int(np.random.choice(60,1))
# mask[start:start+4]=1
# pattern.append(mask.reshape([8,8]))
# np.save('pattern_file/b08b08_k04_n16_pattern_v2.npy',np.concatenate(pattern,0))