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
|
#!/usr/bin/env python3
__author__ = "David Huggins-Daines <dhdaines@gmail.com>"
import numpy
import sys
import os
import struct
from cmusphinx import s3mixw, sendump
import sphinxbase
def mixw_kmeans_iter(lmw, cb):
cbacc = numpy.zeros(len(cb))
cbcnt = numpy.zeros(len(cb))
tdist = 0
for m in lmw:
dist = (cb - m)
dist *= dist
cw = dist.argmin()
tdist += dist.min()
cbacc[cw] += m
cbcnt[cw] += 1
cb[:] = cbacc / cbcnt
return tdist
def map_mixw_cb(mixw, cb, zero=0.0):
n_sen, n_feat, n_gau = mixw.shape
lmw = numpy.log(mixw)
mwmap = numpy.zeros(mixw.shape, 'uint8')
for s in range(0, n_sen):
for f in range(0, n_feat):
for g in range(0, n_gau):
x = mixw[s, f, g]
if x <= zero:
mwmap[s, f, g] = len(cb)
else:
dist = (cb - lmw[s, f, g])
dist *= dist
mwmap[s, f, g] = dist.argmin()
return mwmap
def mixw_freq(mixwmap):
hist = numpy.zeros(mixwmap.max() + 1, 'i')
for cw in mixwmap.ravel():
hist[cw] += 1
return hist
try:
from qmwx import mixw_kmeans_iter, map_mixw_cb, mixw_freq
except ImportError:
pass
def quantize_mixw_kmeans(mixw, k, zero=0.0):
mw = mixw.ravel()
lmw = numpy.log(mw.take(numpy.greater(mw, zero).nonzero()[0]))
mmw = lmw.min()
xmw = lmw.max()
rmw = xmw - mmw
print("min log mixw: %f range: %f" % (mmw, rmw))
cb = numpy.random.random(k) * rmw + mmw
pdist = 1e+50
for i in range(0, 10):
tdist = mixw_kmeans_iter(lmw, cb)
conv = (pdist - tdist) / pdist
print("Total distortion: %e convergence ratio: %e" % (tdist, conv))
if conv < 0.01:
print("Training has converged, stopping")
break
pdist = tdist
return cb
def hb_encode(mixw):
comp = []
for i in range(0, len(mixw) - 1, 2):
comp.append((mixw[i + 1] << 4) | mixw[i])
if len(mixw) % 2:
comp.append(mixw[-1])
return comp
fmtdesc3 = \
"""BEGIN FILE FORMAT DESCRIPTION
(int32) <length(string)> (including trailing 0)
<string> (including trailing 0)
... preceding 2 items repeated any number of times
(int32) 0 (length(string)=0 terminates the header)
cluster_count centroids
cluster index array (feature_count x mixture_count x model_count)
... preceding 2 items repeated codebook_count times
END FILE FORMAT DESCRIPTION
feature_count %d
codebook_count 1
mixture_count %d
model_count %d
cluster_count %d
cluster_bits 4
logbase 1.0001
mixw_shift 10"""
def write_sendump_hb(mixwmap, cb, outfile):
n_sen, n_feat, n_gau = mixwmap.shape
fh = open(outfile, "wb")
# Write the header
fmtdesc0 = fmtdesc3 % (n_feat, n_gau, n_sen, len(cb))
for line in fmtdesc0.split('\n'):
fh.write(struct.pack('>I', len(line) + 1))
fh.write(line)
fh.write('\0')
fh.write(struct.pack('>I', 0))
# Add one extra index to the end to hold the "zero" value
qcb = numpy.resize(-(cb / numpy.log(1.0001)).astype('i') >> 10,
len(cb) + 1)
qcb[-1] = 159
qcb.astype('uint8').tofile(fh)
for f in range(0, n_feat):
for g in range(0, n_gau):
mm = numpy.array(hb_encode(mixwmap[:, f, g]), 'uint8')
mm.tofile(fh)
fh.close()
fmtdesc4 = \
"""BEGIN FILE FORMAT DESCRIPTION
(int32) <length(string)> (including trailing 0)
<string> (including trailing 0)
... preceding 2 items repeated any number of times
(int32) 0 (length(string)=0 terminates the header)
codebook_count <codebook_count>
mixture_count <mixture_count>
model_count <model_count>
cluster_count <cluster_count>
huffman_coded 1
logbase <logarithm_base>
mixw_shift <log_bits_downshifted>
<huffman codebook>
<compressed arrays of mixture weights>
END FILE FORMAT DESCRIPTION
feature_count %d
codebook_count 1
mixture_count %d
model_count %d
cluster_count %d
huffman_coded 1
logbase 1.0001
mixw_shift 10"""
def write_sendump_huff(mixwmap, cb, outfile):
n_sen, n_feat, n_gau = mixwmap.shape
fh = open(outfile, "wb")
# Write the header
fmtdesc0 = fmtdesc3 % (n_feat, n_gau, n_sen, len(cb))
for line in fmtdesc0.split('\n'):
fh.write(struct.pack('>I', len(line) + 1))
fh.write(line)
fh.write('\0')
# Terminate it with a null entry
fh.write(struct.pack('>I', 0))
# If there's an extra "floor" value then add it to the codebook
if mixwmap.max() == len(cb):
qcb = numpy.resize(-(cb / numpy.log(1.0001)).astype('i') >> 10,
len(cb) + 1)
qcb[-1] = 159
else:
qcb = numpy.resize(-(cb / numpy.log(1.0001)).astype('i') >> 10,
len(cb))
# Histogram the mixture weight map and build a Huffman codebook
hist = mixw_freq(mixwmap)
# Write the codebook (we code directly to quantized mixw values)
huff = sphinxbase.HuffCode(list(zip(qcb, hist)))
huff.write(fh)
# Now Huffman code the mixture weights (not their codebook indices
# mind you!) to the output file.
huff.attach(fh, "wb")
for f in range(0, n_feat):
for g in range(0, n_gau):
syms = [qcb[x] for x in mixwmap[:, f, g]]
huff.encode_to_file(syms)
huff.detach()
fh.close()
def norm_floor_mixw(mixw, floor=1e-7):
return (mixw.T / mixw.T.sum(0)).T.clip(floor, 1.0)
if __name__ == '__main__':
ifn, ofn = sys.argv[1:]
if os.path.basename(ifn).startswith('sendump'):
mixw = sendump.Sendump(ifn).mixw()
else:
mixw = norm_floor_mixw(s3mixw.open(ifn).getall(), 1e-7)
cb = quantize_mixw_kmeans(mixw, 15, 1e-7)
mwmap = map_mixw_cb(mixw, cb, 1e-7)
write_sendump_hb(mwmap, cb, ofn)
|
