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
|
#!/usr/bin/env python3
import pickle
import math
import numpy as np
import re
import sys
model_file = sys.argv[1]
trim_trailing_zeros = re.compile('0+p')
def small_hex(f):
hf = float(f).hex()
return trim_trailing_zeros.sub('p', hf)
def process_column(v, pad):
""" process and pad """
return [small_hex(f) for f in v] + [small_hex(0.0)] * pad
def cformatM(fh, name, X, nr=None, nc=None):
nrq = int(math.ceil(X.shape[1] / 4.0))
pad = nrq * 4 - X.shape[1]
lines = map(lambda v: ', '.join(process_column(v, pad)), X)
if nr is None:
nr = X.shape[1]
else:
nrq = int(math.ceil(nr / 4.0))
if nc is None:
nc = X.shape[0]
fh.write('float {}[] = {}\n'.format('__' + name, '{'))
fh.write('\t' + ',\n\t'.join(lines))
fh.write('};\n')
fh.write('_Mat {} = {}\n\t.nr = {},\n\t.nrq = {},\n\t.nc = {},\n\t.stride = {},\n\t.data.f = {}\n{};\n'.format('_' + name, '{', nr, nrq, nc, nrq * 4, '__' + name, '}'))
fh.write('const scrappie_matrix {} = &{};\n\n'.format(name, '_' + name))
def cformatV(fh, name, X):
nrq = int(math.ceil(X.shape[0] / 4.0))
pad = nrq * 4 - X.shape[0]
lines = ', '.join(list(map(lambda f: small_hex(f), X)) + [small_hex(0.)] * pad)
fh.write('float {}[] = {}\n'.format( '__' + name, '{'))
fh.write('\t' + lines)
fh.write('};\n')
fh.write('_Mat {} = {}\n\t.nr = {},\n\t.nrq = {},\n\t.nc = {},\n\t.stride = {},\n\t.data.f = {}\n{};\n'.format('_' + name, '{', X.shape[0], nrq, 1, nrq * 4, '__' + name, '}'))
fh.write('const scrappie_matrix {} = &{};\n\n'.format(name, '_' + name))
with open(model_file, 'rb') as fh:
network = pickle.load(fh, encoding='latin1')
assert network.version == 1, "Sloika model must be version 1. Perhaps you need to run Sloika's model_upgrade.py"
sys.stdout.write("""#pragma once
#ifndef NANONET_RAW_MODEL_H
#define NANONET_RAW_MODEL_H
#include <assert.h>
#include "../util.h"
""")
""" First LSTM layer
"""
filterW = network.sublayers[0].W.get_value()
nfilter, _ , winlen = filterW.shape
cformatM(sys.stdout, 'conv_raw_W', filterW.reshape(-1, 1), nr = winlen * 4 - 3, nc=nfilter)
cformatV(sys.stdout, 'conv_raw_b', network.sublayers[0].b.get_value().reshape(-1))
sys.stdout.write("const int conv_raw_stride = {};\n".format(network.sublayers[0].stride))
sys.stdout.write("""const size_t _conv_nfilter = {};
const size_t _conv_winlen = {};
""".format(nfilter, winlen))
bigru1 = network.sublayers[1]
gru = bigru1.sublayers[0]
cformatM(sys.stdout, 'gruF1_raw_iW', gru.iW.get_value())
cformatM(sys.stdout, 'gruF1_raw_sW', gru.sW.get_value())
cformatM(sys.stdout, 'gruF1_raw_sW2', gru.sW2.get_value())
cformatV(sys.stdout, 'gruF1_raw_b', gru.b.get_value().reshape(-1))
gru = bigru1.sublayers[1].sublayers[0]
cformatM(sys.stdout, 'gruB1_raw_iW', gru.iW.get_value())
cformatM(sys.stdout, 'gruB1_raw_sW', gru.sW.get_value())
cformatM(sys.stdout, 'gruB1_raw_sW2', gru.sW2.get_value())
cformatV(sys.stdout, 'gruB1_raw_b', gru.b.get_value().reshape(-1))
""" First feed forward layer
"""
assert(network.sublayers[2].insize % 2 == 0)
size = network.sublayers[2].insize // 2
cformatM(sys.stdout, 'FF1_raw_Wf', network.sublayers[2].W.get_value()[:, : size])
cformatM(sys.stdout, 'FF1_raw_Wb', network.sublayers[2].W.get_value()[:, size : 2 * size])
cformatV(sys.stdout, 'FF1_raw_b', network.sublayers[2].b.get_value())
""" Second GRU layer
"""
bigru1 = network.sublayers[3]
gru = bigru1.sublayers[0]
cformatM(sys.stdout, 'gruF2_raw_iW', gru.iW.get_value())
cformatM(sys.stdout, 'gruF2_raw_sW', gru.sW.get_value())
cformatM(sys.stdout, 'gruF2_raw_sW2', gru.sW2.get_value())
cformatV(sys.stdout, 'gruF2_raw_b', gru.b.get_value().reshape(-1))
gru = bigru1.sublayers[1].sublayers[0]
cformatM(sys.stdout, 'gruB2_raw_iW', gru.iW.get_value())
cformatM(sys.stdout, 'gruB2_raw_sW', gru.sW.get_value())
cformatM(sys.stdout, 'gruB2_raw_sW2', gru.sW2.get_value())
cformatV(sys.stdout, 'gruB2_raw_b', gru.b.get_value().reshape(-1))
""" Second feed forward layer
"""
size = network.sublayers[4].insize // 2
assert(network.sublayers[4].insize % 2 == 0)
cformatM(sys.stdout, 'FF2_raw_Wf', network.sublayers[4].W.get_value()[:, : size])
cformatM(sys.stdout, 'FF2_raw_Wb', network.sublayers[4].W.get_value()[:, size : 2 * size])
cformatV(sys.stdout, 'FF2_raw_b', network.sublayers[4].b.get_value())
""" Softmax layer
"""
nstate = network.sublayers[5].W.get_value().shape[0]
shuffle = np.append(np.arange(nstate - 1) + 1, 0)
cformatM(sys.stdout, 'FF3_raw_W', network.sublayers[5].W.get_value()[shuffle])
cformatV(sys.stdout, 'FF3_raw_b', network.sublayers[5].b.get_value()[shuffle])
sys.stdout.write('#endif /* NANONET_RAW_MODEL_H */')
|
