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
|
require 'htk_io'
gconf = {lrate = 0.8, wcost = 1e-6, momentum = 0.9, frm_ext = 5,
rearrange = true, -- just to make the context order consistent with old results, deprecated
frm_trim = 5, -- trim the first and last 5 frames, TNet just does this, deprecated
tr_scp = "/speechlab/users/mfy43/swb50/train_bp.scp",
cv_scp = "/speechlab/users/mfy43/swb50/train_cv.scp",
htk_conf = "/speechlab/users/mfy43/swb50/plp_0_d_a.conf",
initialized_param = {"/speechlab/users/mfy43/swb50/swb_init.nerv",
"/speechlab/users/mfy43/swb50/swb_global_transf.nerv"}}
function make_layer_repo(param_repo)
local layer_repo = nerv.LayerRepo(
{
-- global transf
["nerv.BiasLayer"] =
{
blayer1 = {dim_in = {429}, dim_out = {429}, params = {bias = "bias1"}},
blayer2 = {dim_in = {429}, dim_out = {429}, params = {bias = "bias2"}}
},
["nerv.WindowLayer"] =
{
wlayer1 = {dim_in = {429}, dim_out = {429}, params = {window = "window1"}},
wlayer2 = {dim_in = {429}, dim_out = {429}, params = {window = "window2"}}
},
-- biased linearity
["nerv.AffineLayer"] =
{
affine0 = {dim_in = {429}, dim_out = {2048},
params = {ltp = "affine0_ltp", bp = "affine0_bp"}},
affine1 = {dim_in = {2048}, dim_out = {2048},
params = {ltp = "affine1_ltp", bp = "affine1_bp"}},
affine2 = {dim_in = {2048}, dim_out = {2048},
params = {ltp = "affine2_ltp", bp = "affine2_bp"}},
affine3 = {dim_in = {2048}, dim_out = {2048},
params = {ltp = "affine3_ltp", bp = "affine3_bp"}},
affine4 = {dim_in = {2048}, dim_out = {2048},
params = {ltp = "affine4_ltp", bp = "affine4_bp"}},
affine5 = {dim_in = {2048}, dim_out = {2048},
params = {ltp = "affine5_ltp", bp = "affine5_bp"}},
affine6 = {dim_in = {2048}, dim_out = {2048},
params = {ltp = "affine6_ltp", bp = "affine6_bp"}},
affine7 = {dim_in = {2048}, dim_out = {3001},
params = {ltp = "affine7_ltp", bp = "affine7_bp"}}
},
["nerv.SigmoidLayer"] =
{
sigmoid0 = {dim_in = {2048}, dim_out = {2048}},
sigmoid1 = {dim_in = {2048}, dim_out = {2048}},
sigmoid2 = {dim_in = {2048}, dim_out = {2048}},
sigmoid3 = {dim_in = {2048}, dim_out = {2048}},
sigmoid4 = {dim_in = {2048}, dim_out = {2048}},
sigmoid5 = {dim_in = {2048}, dim_out = {2048}},
sigmoid6 = {dim_in = {2048}, dim_out = {2048}}
},
["nerv.SoftmaxCELayer"] = -- softmax + ce criterion layer for finetune output
{
ce_crit = {dim_in = {3001, 1}, dim_out = {1}, compressed = true}
},
["nerv.SoftmaxLayer"] = -- softmax for decode output
{
softmax = {dim_in = {3001}, dim_out = {3001}}
}
}, param_repo, gconf)
layer_repo:add_layers(
{
["nerv.GraphLayer"] =
{
global_transf = {
dim_in = {429}, dim_out = {429},
layer_repo = layer_repo,
connections = {
{"<input>[1]", "blayer1[1]", 0},
{"blayer1[1]", "wlayer1[1]", 0},
{"wlayer1[1]", "blayer2[1]", 0},
{"blayer2[1]", "wlayer2[1]", 0},
{"wlayer2[1]", "<output>[1]", 0}
}
},
main = {
dim_in = {429}, dim_out = {3001},
layer_repo = layer_repo,
connections = {
{"<input>[1]", "affine0[1]", 0},
{"affine0[1]", "sigmoid0[1]", 0},
{"sigmoid0[1]", "affine1[1]", 0},
{"affine1[1]", "sigmoid1[1]", 0},
{"sigmoid1[1]", "affine2[1]", 0},
{"affine2[1]", "sigmoid2[1]", 0},
{"sigmoid2[1]", "affine3[1]", 0},
{"affine3[1]", "sigmoid3[1]", 0},
{"sigmoid3[1]", "affine4[1]", 0},
{"affine4[1]", "sigmoid4[1]", 0},
{"sigmoid4[1]", "affine5[1]", 0},
{"affine5[1]", "sigmoid5[1]", 0},
{"sigmoid5[1]", "affine6[1]", 0},
{"affine6[1]", "sigmoid6[1]", 0},
{"sigmoid6[1]", "affine7[1]", 0},
{"affine7[1]", "<output>[1]", 0}
}
}
}
}, param_repo, gconf)
layer_repo:add_layers(
{
["nerv.GraphLayer"] =
{
ce_output = {
dim_in = {429, 1}, dim_out = {1},
layer_repo = layer_repo,
connections = {
{"<input>[1]", "main[1]", 0},
{"main[1]", "ce_crit[1]", 0},
{"<input>[2]", "ce_crit[2]", 0},
{"ce_crit[1]", "<output>[1]", 0}
}
},
softmax_output = {
dim_in = {429}, dim_out = {3001},
layer_repo = layer_repo,
connections = {
{"<input>[1]", "main[1]", 0},
{"main[1]", "softmax[1]", 0},
{"softmax[1]", "<output>[1]", 0}
}
}
}
}, param_repo, gconf)
return layer_repo
end
function get_network(layer_repo)
return layer_repo:get_layer("ce_output")
end
function get_decode_network(layer_repo)
return layer_repo:get_layer("softmax_output")
end
function get_global_transf(layer_repo)
return layer_repo:get_layer("global_transf")
end
function make_readers(scp_file, layer_repo)
return {
{reader = nerv.TNetReader(gconf,
{
id = "main_scp",
scp_file = scp_file,
conf_file = gconf.htk_conf,
frm_ext = gconf.frm_ext,
mlfs = {
phone_state = {
file = "/speechlab/users/mfy43/swb50/ref.mlf",
format = "map",
format_arg = "/speechlab/users/mfy43/swb50/dict",
dir = "*/",
ext = "lab"
}
}
}),
data = {main_scp = 429, phone_state = 1}}
}
end
function make_buffer(readers)
return nerv.SGDBuffer(gconf,
{
buffer_size = gconf.buffer_size,
batch_size = gconf.batch_size,
randomize = gconf.randomize,
readers = readers,
use_gpu = true
})
end
function get_input_order()
return {{id = "main_scp", global_transf = true},
{id = "phone_state"}}
end
function get_decode_input_order()
return {{id = "main_scp", global_transf = true}}
end
function get_accuracy(layer_repo)
local ce_crit = layer_repo:get_layer("ce_crit")
return ce_crit.total_correct / ce_crit.total_frames * 100
end
function print_stat(layer_repo)
local ce_crit = layer_repo:get_layer("ce_crit")
nerv.info("*** training stat begin ***")
nerv.printf("cross entropy:\t\t%.8f\n", ce_crit.total_ce)
nerv.printf("correct:\t\t%d\n", ce_crit.total_correct)
nerv.printf("frames:\t\t\t%d\n", ce_crit.total_frames)
nerv.printf("err/frm:\t\t%.8f\n", ce_crit.total_ce / ce_crit.total_frames)
nerv.printf("accuracy:\t\t%.3f%%\n", get_accuracy(layer_repo))
nerv.info("*** training stat end ***")
end
|