forked from razor1179/pytorch-kaldi-CGS
-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathhcgs.py
executable file
·143 lines (140 loc) · 10.2 KB
/
hcgs.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
import numpy as np
import scipy.io as sio
import cgs_base
import torch
from torch.nn.parameter import Parameter
import torch.nn as nn
def conn_mat(n_in, n_out, block_sizes, drop_ratios, mat_num='1', dir='/home/dkadetot/saved_mat', equal_blks_for_input = True, for_test = False):
if not len(block_sizes) == len(drop_ratios):
print('block size and drop ratio should have the same length!')
exit()
block_sizes.reverse()
drop_ratios.reverse()
recursive_call = len(block_sizes)
block_size = block_sizes.pop()
drop_ratio = drop_ratios.pop()
sparsity = 1 - float(drop_ratio) / 100
n_blk_rows = n_in / block_size
conn_mat = np.full((n_in, n_out), 0, dtype='float32')
if n_in % block_size != 0:
n_blk_rows += 1
n_blk_cols = n_out / block_size
if n_out % block_size != 0:
n_blk_cols += 1
if equal_blks_for_input:
n_blk_sels = int(round(n_blk_cols * sparsity))
n_blk_rows = int(n_blk_rows)
n_blk_cols = int(n_blk_cols)
for i in range(n_blk_rows-1):
rand_choices = np.random.choice(n_blk_cols, n_blk_sels, False)
# print rand_choices
for j in range(n_blk_sels):
if rand_choices[j] == n_blk_cols-1 and n_out % block_size != 0:
conn_mat[i*block_size:(i+1)*block_size, rand_choices[j]*block_size:n_out] = 1
r_h, c_h = conn_mat[i*block_size:(i+1)*block_size, rand_choices[j]*block_size:n_out].shape
conn_mat[i * block_size:(i + 1) * block_size, rand_choices[j] * block_size:n_out] = cgs_base.conn_mat(r_h, c_h, block_sizes[:], drop_ratios[:], equal_blks_for_input=equal_blks_for_input, recursive_call=recursive_call)
else:
conn_mat[i*block_size:(i+1)*block_size, rand_choices[j]*block_size:(rand_choices[j]+1)*block_size] = 1
r_h, c_h = conn_mat[i * block_size:(i + 1) * block_size, rand_choices[j] * block_size:(rand_choices[j] + 1) * block_size].shape
conn_mat[i * block_size:(i + 1) * block_size, rand_choices[j] * block_size:(rand_choices[j] + 1) * block_size] = cgs_base.conn_mat(r_h, c_h, block_sizes[:], drop_ratios[:], equal_blks_for_input=equal_blks_for_input, recursive_call=recursive_call)
rand_choices = np.random.choice(n_blk_cols, n_blk_sels, False)
for j in range(n_blk_sels):
conn_mat[(n_blk_rows-1)*block_size:n_in, rand_choices[j]*block_size:(rand_choices[j]+1)*block_size] = 1
r_h, c_h = conn_mat[(n_blk_rows - 1) * block_size:n_in, rand_choices[j] * block_size:(rand_choices[j] + 1) * block_size].shape
conn_mat[(n_blk_rows - 1) * block_size:n_in, rand_choices[j] * block_size:(rand_choices[j] + 1) * block_size] = cgs_base.conn_mat(r_h, c_h, block_sizes[:], drop_ratios[:], equal_blks_for_input=equal_blks_for_input, recursive_call=recursive_call)
else:
n_blk_sels = int(round(n_blk_rows * sparsity))
for i in range(n_blk_cols-1):
rand_choices = np.random.choice(n_blk_rows, n_blk_sels, False)
for j in range(n_blk_sels):
if rand_choices[j] == n_blk_rows-1 and n_in % block_size != 0:
conn_mat[rand_choices[j]*block_size:n_in, i*block_size:(i+1)*block_size] = 1
r_h, c_h = conn_mat[rand_choices[j] * block_size:n_in, i * block_size:(i + 1) * block_size].shape
conn_mat[rand_choices[j] * block_size:n_in, i * block_size:(i + 1) * block_size] = cgs_base.conn_mat(r_h, c_h, block_sizes[:], drop_ratios[:], equal_blks_for_input=equal_blks_for_input, recursive_call=recursive_call)
else:
conn_mat[rand_choices[j]*block_size:(rand_choices[j]+1)*block_size, i*block_size:(i+1)*block_size] = 1
r_h, c_h = conn_mat[rand_choices[j] * block_size:(rand_choices[j] + 1) * block_size, i * block_size:(i + 1) * block_size].shape
conn_mat[rand_choices[j] * block_size:(rand_choices[j] + 1) * block_size, i * block_size:(i + 1) * block_size] = cgs_base.conn_mat(r_h, c_h, block_sizes[:], drop_ratios[:], equal_blks_for_input=equal_blks_for_input, recursive_call=recursive_call)
rand_choices = np.random.choice(n_blk_rows, n_blk_sels, False)
for j in range(n_blk_sels):
conn_mat[rand_choices[j]*block_size:(rand_choices[j]+1)*block_size, (n_blk_cols-1)*block_size:n_out] = 1
r_h, c_h = conn_mat[rand_choices[j] * block_size:(rand_choices[j] + 1) * block_size, (n_blk_cols - 1) * block_size:n_out].shape
conn_mat[rand_choices[j] * block_size:(rand_choices[j] + 1) * block_size, (n_blk_cols - 1) * block_size:n_out] = cgs_base.conn_mat(r_h, c_h, block_sizes[:], drop_ratios[:], equal_blks_for_input=equal_blks_for_input, recursive_call=recursive_call)
if for_test:
# Save conn_mat in mat file
sio.savemat(dir + '/conn_mat%s.mat' % mat_num, {'CM%s' % mat_num: conn_mat})
return conn_mat
else:
conn_mat_torch = torch.Tensor(n_in, n_out)
conn_mat_torch = torch.from_numpy(conn_mat)
device = torch.device("cuda")
conn_mat_torch = conn_mat_torch.to(device)
conn_mat_torch.requires_grad_(False)
return conn_mat_torch
def conn_mat(n_in, n_out, block_sizes, drop_ratios, mat_num='1', dir='/home/dkadetot/saved_mat', equal_blks_for_input = True, for_test = False):
if not len(block_sizes) == len(drop_ratios):
print('block size and drop ratio should have the same length!')
exit()
block_sizes.reverse()
drop_ratios.reverse()
recursive_call = len(block_sizes)
block_size = block_sizes.pop()
drop_ratio = drop_ratios.pop()
sparsity = 1 - float(drop_ratio) / 100
n_blk_rows = n_in / block_size
conn_mat = np.full((n_in, n_out), 0, dtype='float32')
if n_in % block_size != 0:
n_blk_rows += 1
n_blk_cols = n_out / block_size
if n_out % block_size != 0:
n_blk_cols += 1
if equal_blks_for_input:
n_blk_sels = int(round(n_blk_cols * sparsity))
n_blk_rows = int(n_blk_rows)
n_blk_cols = int(n_blk_cols)
for i in range(n_blk_rows-1):
rand_choices = np.random.choice(n_blk_cols, n_blk_sels, False)
# print rand_choices
for j in range(n_blk_sels):
if rand_choices[j] == n_blk_cols-1 and n_out % block_size != 0:
conn_mat[i*block_size:(i+1)*block_size, rand_choices[j]*block_size:n_out] = 1
r_h, c_h = conn_mat[i*block_size:(i+1)*block_size, rand_choices[j]*block_size:n_out].shape
conn_mat[i * block_size:(i + 1) * block_size, rand_choices[j] * block_size:n_out] = cgs_base.conn_mat(r_h, c_h, block_sizes[:], drop_ratios[:], equal_blks_for_input=equal_blks_for_input, recursive_call=recursive_call)
else:
conn_mat[i*block_size:(i+1)*block_size, rand_choices[j]*block_size:(rand_choices[j]+1)*block_size] = 1
r_h, c_h = conn_mat[i * block_size:(i + 1) * block_size, rand_choices[j] * block_size:(rand_choices[j] + 1) * block_size].shape
conn_mat[i * block_size:(i + 1) * block_size, rand_choices[j] * block_size:(rand_choices[j] + 1) * block_size] = cgs_base.conn_mat(r_h, c_h, block_sizes[:], drop_ratios[:], equal_blks_for_input=equal_blks_for_input, recursive_call=recursive_call)
rand_choices = np.random.choice(n_blk_cols, n_blk_sels, False)
for j in range(n_blk_sels):
conn_mat[(n_blk_rows-1)*block_size:n_in, rand_choices[j]*block_size:(rand_choices[j]+1)*block_size] = 1
r_h, c_h = conn_mat[(n_blk_rows - 1) * block_size:n_in, rand_choices[j] * block_size:(rand_choices[j] + 1) * block_size].shape
conn_mat[(n_blk_rows - 1) * block_size:n_in, rand_choices[j] * block_size:(rand_choices[j] + 1) * block_size] = cgs_base.conn_mat(r_h, c_h, block_sizes[:], drop_ratios[:], equal_blks_for_input=equal_blks_for_input, recursive_call=recursive_call)
else:
n_blk_sels = int(round(n_blk_rows * sparsity))
for i in range(n_blk_cols-1):
rand_choices = np.random.choice(n_blk_rows, n_blk_sels, False)
for j in range(n_blk_sels):
if rand_choices[j] == n_blk_rows-1 and n_in % block_size != 0:
conn_mat[rand_choices[j]*block_size:n_in, i*block_size:(i+1)*block_size] = 1
r_h, c_h = conn_mat[rand_choices[j] * block_size:n_in, i * block_size:(i + 1) * block_size].shape
conn_mat[rand_choices[j] * block_size:n_in, i * block_size:(i + 1) * block_size] = cgs_base.conn_mat(r_h, c_h, block_sizes[:], drop_ratios[:], equal_blks_for_input=equal_blks_for_input, recursive_call=recursive_call)
else:
conn_mat[rand_choices[j]*block_size:(rand_choices[j]+1)*block_size, i*block_size:(i+1)*block_size] = 1
r_h, c_h = conn_mat[rand_choices[j] * block_size:(rand_choices[j] + 1) * block_size, i * block_size:(i + 1) * block_size].shape
conn_mat[rand_choices[j] * block_size:(rand_choices[j] + 1) * block_size, i * block_size:(i + 1) * block_size] = cgs_base.conn_mat(r_h, c_h, block_sizes[:], drop_ratios[:], equal_blks_for_input=equal_blks_for_input, recursive_call=recursive_call)
rand_choices = np.random.choice(n_blk_rows, n_blk_sels, False)
for j in range(n_blk_sels):
conn_mat[rand_choices[j]*block_size:(rand_choices[j]+1)*block_size, (n_blk_cols-1)*block_size:n_out] = 1
r_h, c_h = conn_mat[rand_choices[j] * block_size:(rand_choices[j] + 1) * block_size, (n_blk_cols - 1) * block_size:n_out].shape
conn_mat[rand_choices[j] * block_size:(rand_choices[j] + 1) * block_size, (n_blk_cols - 1) * block_size:n_out] = cgs_base.conn_mat(r_h, c_h, block_sizes[:], drop_ratios[:], equal_blks_for_input=equal_blks_for_input, recursive_call=recursive_call)
if for_test:
# Save conn_mat in mat file
sio.savemat(dir + '/conn_mat%s.mat' % mat_num, {'CM%s' % mat_num: conn_mat})
return conn_mat
else:
conn_mat_torch = torch.Tensor(n_in, n_out)
conn_mat_torch = torch.from_numpy(conn_mat)
device = torch.device("cuda")
conn_mat_torch = conn_mat_torch.to(device)
conn_mat_torch.requires_grad_(False)
return conn_mat_torch