forked from ai-adv-lab/deepspeech.mxnet
/
arch_deepspeech_vgg.py
291 lines (259 loc) · 14.4 KB
/
arch_deepspeech_vgg.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
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
# pylint: disable=C0111, too-many-statements, too-many-locals
# pylint: too-many-arguments,too-many-instance-attributes,too-many-locals,redefined-outer-name,fixme
# pylint: disable=superfluous-parens, no-member, invalid-name
"""
architecture file for deep speech 2 model
"""
import json
import math
import argparse
import mxnet as mx
from stt_layer_batchnorm import batchnorm
from stt_layer_conv import conv
from stt_layer_fc import sequence_fc
from stt_layer_gru import bi_gru_unroll, gru_unroll
from stt_layer_lstm import bi_lstm_unroll
from stt_layer_slice import slice_symbol_to_seq_symobls
from stt_layer_warpctc import warpctc_layer
def prepare_data(args, is_val=False):
"""
set atual shape of data
"""
rnn_type = args.config.get("arch", "rnn_type")
num_rnn_layer = args.config.getint("arch", "num_rnn_layer")
num_hidden_rnn_list = json.loads(args.config.get("arch", "num_hidden_rnn_list"))
num_hidden_proj = args.config.getint("arch", "num_hidden_proj")
batch_size = args.config.getint("common", "batch_size")
if is_val:
batch_size = args.config.getint("common", "val_batch_size")
if rnn_type == 'lstm':
init_c = [('l%d_init_c' % l, (batch_size, num_hidden_rnn_list[l]))
for l in range(num_rnn_layer)]
init_h = [('l%d_init_h' % l, (batch_size, num_hidden_rnn_list[l]))
for l in range(num_rnn_layer)]
elif rnn_type == 'bilstm':
forward_init_c = [('forward_l%d_init_c' % l, (batch_size, num_hidden_rnn_list[l]))
for l in range(num_rnn_layer)]
backward_init_c = [('backward_l%d_init_c' % l, (batch_size, num_hidden_rnn_list[l]))
for l in range(num_rnn_layer)]
init_c = forward_init_c + backward_init_c
if num_hidden_proj > 0:
forward_init_h = [('forward_l%d_init_h' % l, (batch_size, num_hidden_proj))
for l in range(num_rnn_layer)]
backward_init_h = [('backward_l%d_init_h' % l, (batch_size, num_hidden_proj))
for l in range(num_rnn_layer)]
else:
forward_init_h = [('forward_l%d_init_h' % l, (batch_size, num_hidden_rnn_list[l]))
for l in range(num_rnn_layer)]
backward_init_h = [('backward_l%d_init_h' % l, (batch_size, num_hidden_rnn_list[l]))
for l in range(num_rnn_layer)]
init_h = forward_init_h + backward_init_h
elif rnn_type == 'gru':
init_h = [('l%d_init_h' % l, (batch_size, num_hidden_rnn_list[l]))
for l in range(num_rnn_layer)]
elif rnn_type == 'bigru':
forward_init_h = [('forward_l%d_init_h' % l, (batch_size, num_hidden_rnn_list[l]))
for l in range(num_rnn_layer)]
backward_init_h = [('backward_l%d_init_h' % l, (batch_size, num_hidden_rnn_list[l]))
for l in range(num_rnn_layer)]
init_h = forward_init_h + backward_init_h
else:
raise Exception('network type should be one of the lstm,bilstm,gru,bigru')
if rnn_type == 'lstm' or rnn_type == 'bilstm':
init_states = init_c + init_h
elif rnn_type == 'gru' or rnn_type == 'bigru':
init_states = init_h
return init_states
def get_feature(internel_layer, layers, filters, batch_norm=False, **kwargs):
for i, num in enumerate(layers):
for j in range(num):
internel_layer = mx.sym.Convolution(data=internel_layer, kernel=(3, 3), pad=(1, 1), num_filter=filters[i],
name="conv%s_%s" % (i + 1, j + 1))
if batch_norm:
internel_layer = mx.symbol.BatchNorm(data=internel_layer, name="bn%s_%s" % (i + 1, j + 1))
internel_layer = mx.sym.Activation(data=internel_layer, act_type="relu", name="relu%s_%s" % (i + 1, j + 1))
internel_layer = mx.sym.Pooling(data=internel_layer, pool_type="max", kernel=(2, 2), stride=(2, 2),
name="pool%s" % (i + 1))
return internel_layer
def arch(args, seq_len=None):
"""
define deep speech 2 network
"""
if isinstance(args, argparse.Namespace):
mode = args.config.get("common", "mode")
is_bucketing = args.config.getboolean("arch", "is_bucketing")
if mode == "train" or is_bucketing:
# channel_num = args.config.getint("arch", "channel_num")
conv_layer1_filter_dim = \
tuple(json.loads(args.config.get("arch", "conv_layer1_filter_dim")))
conv_layer1_stride = tuple(json.loads(args.config.get("arch", "conv_layer1_stride")))
conv_layer2_filter_dim = \
tuple(json.loads(args.config.get("arch", "conv_layer2_filter_dim")))
conv_layer2_stride = tuple(json.loads(args.config.get("arch", "conv_layer2_stride")))
rnn_type = args.config.get("arch", "rnn_type")
num_rnn_layer = args.config.getint("arch", "num_rnn_layer")
num_hidden_proj = args.config.getint("arch", "num_hidden_proj")
num_hidden_rnn_list = json.loads(args.config.get("arch", "num_hidden_rnn_list"))
is_batchnorm = args.config.getboolean("arch", "is_batchnorm")
fbank = args.config.getboolean("data", "fbank")
if seq_len is None:
seq_len = args.config.getint('arch', 'max_t_count')
num_label = args.config.getint('arch', 'max_label_length')
num_rear_fc_layers = args.config.getint("arch", "num_rear_fc_layers")
num_hidden_rear_fc_list = json.loads(args.config.get("arch", "num_hidden_rear_fc_list"))
act_type_rear_fc_list = json.loads(args.config.get("arch", "act_type_rear_fc_list"))
# model symbol generation
# input preparation
data = mx.sym.Variable('data')
label = mx.sym.Variable('label')
if not fbank:
net = mx.sym.Reshape(data=data, shape=(-4, -1, 1, 0, 0))
else:
net = data
net = mx.sym.Convolution(data=net, kernel=(3, 3), pad=(1, 1), num_filter=64, no_bias=is_batchnorm,
name="conv%s_%s" % (1, 1))
if is_batchnorm:
net = mx.symbol.BatchNorm(data=net, name="bn%s_%s" % (1, 1))
net = mx.sym.Activation(data=net, act_type="relu", name="relu%s_%s" %(1, 1))
net = mx.sym.Convolution(data=net, kernel=(3, 3), pad=(1, 1), num_filter=64, no_bias=is_batchnorm,
name="conv%s_%s" % (1, 2))
if is_batchnorm:
net = mx.symbol.BatchNorm(data=net, name="bn%s_%s" % (1, 2))
net = mx.sym.Activation(data=net, act_type="relu", name="relu%s_%s" %(1, 2))
net = mx.sym.Pooling(data=net, pool_type="max", kernel=(3, 3), stride=(2, 2),
name="pool%s" % (1))
net = mx.sym.Convolution(data=net, kernel=(3, 3), pad=(1, 1), num_filter=128, no_bias=is_batchnorm,
name="conv%s_%s" % (2, 1))
if is_batchnorm:
net = mx.symbol.BatchNorm(data=net, name="bn%s_%s" % (2, 1))
net = mx.sym.Activation(data=net, act_type="relu", name="relu%s_%s" %(2, 1))
net = mx.sym.Convolution(data=net, kernel=(3, 3), pad=(1, 1), num_filter=128, no_bias=is_batchnorm,
name="conv%s_%s" % (2, 2))
if is_batchnorm:
net = mx.symbol.BatchNorm(data=net, name="bn%s_%s" % (2, 2))
net = mx.sym.Activation(data=net, act_type="relu", name="relu%s_%s" %(2, 2))
net = mx.sym.Pooling(data=net, pool_type="max", kernel=(3, 3), stride=(2, 2),
name="pool%s" % (2))
# net = conv(net=net,
# channels=channel_num,
# filter_dimension=conv_layer1_filter_dim,
# stride=conv_layer1_stride,
# no_bias=is_batchnorm,
# name='conv1')
# if is_batchnorm:
# # batch norm normalizes axis 1
# net = batchnorm(net, name="conv1_batchnorm")
#
# net = conv(net=net,
# channels=channel_num,
# filter_dimension=conv_layer2_filter_dim,
# stride=conv_layer2_stride,
# no_bias=is_batchnorm,
# name='conv2')
# if is_batchnorm:
# # batch norm normalizes axis 1
# net = batchnorm(net, name="conv2_batchnorm")
#
# net = conv(net=net,
# channels=96,
# filter_dimension=conv_layer3_filter_dim,
# stride=conv_layer3_stride,
# no_bias=is_batchnorm,
# name='conv3')
# if is_batchnorm:
# # batch norm normalizes axis 1
# net = batchnorm(net, name="conv3_batchnorm")
net = mx.sym.transpose(data=net, axes=(0, 2, 1, 3))
net = mx.sym.Reshape(data=net, shape=(0, 0, -3))
seq_len_after_conv_layer1 = int(math.floor((seq_len - 3) / 2)) + 1
seq_len_after_conv = int(math.floor((seq_len_after_conv_layer1 - 3) / 2)) + 1
net = slice_symbol_to_seq_symobls(net=net, seq_len=seq_len_after_conv, axis=1)
if rnn_type == "bilstm":
net, f_states, b_states = bi_lstm_unroll(net=net,
seq_len=seq_len_after_conv,
num_hidden_lstm_list=num_hidden_rnn_list,
num_lstm_layer=num_rnn_layer,
dropout=0.,
num_hidden_proj=num_hidden_proj,
is_batchnorm=is_batchnorm,
is_bucketing=is_bucketing)
elif rnn_type == "gru":
net = gru_unroll(net=net,
seq_len=seq_len_after_conv,
num_hidden_gru_list=num_hidden_rnn_list,
num_gru_layer=num_rnn_layer,
dropout=0.,
is_batchnorm=is_batchnorm,
is_bucketing=is_bucketing)
elif rnn_type == "bigru":
net = bi_gru_unroll(net=net,
seq_len=seq_len_after_conv,
num_hidden_gru_list=num_hidden_rnn_list,
num_gru_layer=num_rnn_layer,
dropout=0.,
is_batchnorm=is_batchnorm,
is_bucketing=is_bucketing)
else:
raise Exception('rnn_type should be one of the followings, bilstm,gru,bigru')
# rear fc layers
net = sequence_fc(net=net, seq_len=seq_len_after_conv,
num_layer=num_rear_fc_layers, prefix="rear",
num_hidden_list=num_hidden_rear_fc_list,
act_type_list=act_type_rear_fc_list,
is_batchnorm=is_batchnorm)
cls_weight = mx.sym.Variable("cls_weight")
cls_bias = mx.sym.Variable("cls_bias")
fc_seq = []
character_classes_count = args.config.getint('arch', 'n_classes') + 1
for seqidx in range(seq_len_after_conv):
hidden = net[seqidx]
hidden = mx.sym.FullyConnected(data=hidden,
num_hidden=character_classes_count,
weight=cls_weight,
bias=cls_bias)
fc_seq.append(hidden)
net = mx.sym.Concat(*fc_seq, dim=0, name="warpctc_layer_concat")
# warpctc layer
net = warpctc_layer(net=net,
seq_len=seq_len_after_conv,
label=label,
num_label=num_label,
character_classes_count=
(args.config.getint('arch', 'n_classes') + 1))
args.config.set('arch', 'max_t_count', str(seq_len_after_conv))
return net
elif mode == 'load' or mode == 'predict':
conv_layer1_filter_dim = \
tuple(json.loads(args.config.get("arch", "conv_layer1_filter_dim")))
conv_layer1_stride = tuple(json.loads(args.config.get("arch", "conv_layer1_stride")))
conv_layer2_filter_dim = \
tuple(json.loads(args.config.get("arch", "conv_layer2_filter_dim")))
conv_layer2_stride = tuple(json.loads(args.config.get("arch", "conv_layer2_stride")))
conv_layer3_filter_dim = \
tuple(json.loads(args.config.get("arch", "conv_layer3_filter_dim")))
conv_layer3_stride = tuple(json.loads(args.config.get("arch", "conv_layer3_stride")))
if seq_len is None:
seq_len = args.config.getint('arch', 'max_t_count')
seq_len_after_conv_layer1 = int(
math.floor((seq_len - conv_layer1_filter_dim[0]) / conv_layer1_stride[0])) + 1
seq_len_after_conv_layer2 = int(
math.floor((seq_len_after_conv_layer1 - conv_layer2_filter_dim[0])
/ conv_layer2_stride[0])) + 1
seq_len_after_conv_layer3 = int(
math.floor((seq_len_after_conv_layer2 - conv_layer3_filter_dim[0])
/ conv_layer3_stride[0])) + 1
args.config.set('arch', 'max_t_count', str(seq_len_after_conv_layer3))
else:
raise Exception('mode must be the one of the followings - train,predict,load')
class BucketingArch(object):
def __init__(self, args):
self.args = args
def sym_gen(self, seq_len):
args = self.args
net = arch(args, seq_len)
init_states = prepare_data(args)
init_state_names = [x[0] for x in init_states]
init_state_names.insert(0, 'data')
return net, init_state_names, ('label',)
def get_sym_gen(self):
return self.sym_gen