run_e2e_cnn.sh
6.52 KB
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
#!/bin/bash
# Copyright 2017 Hossein Hadian
# This script does end2end chain training (i.e. from scratch)
# local/chain/compare_wer.sh exp/chain/e2e_cnn_1a/
# System e2e_cnn_1a
# score_basic score_nomalized
# WER 13.64 10.6
# WER (rescored) 13.13 10.2
# CER 2.99 3.0
# CER (rescored) 2.88 2.9
# Final train prob 0.0113
# Final valid prob 0.0152
# steps/info/chain_dir_info.pl exp/chain/e2e_cnn_1a
# exp/chain/e2e_cnn_1a: num-iters=48 nj=5..8 num-params=3.0M dim=40->352 combine=0.047->0.047 (over 2) logprob:train/valid[31,47,final]=(0.002,0.008,0.011/0.008,0.013,0.015)
set -e
# configs for 'chain'
stage=0
nj=30
train_stage=-10
get_egs_stage=-10
affix=1a
# training options
tdnn_dim=450
minibatch_size=150=64,32/300=32,16/600=16,8/1200=8,4
cmvn_opts="--norm-means=false --norm-vars=false"
train_set=train
lang_decode=data/lang
lang_rescore=data/lang_rescore_6g
decode_e2e=true
# End configuration section.
echo "$0 $@" # Print the command line for logging
. ./cmd.sh
. ./path.sh
. ./utils/parse_options.sh
if ! cuda-compiled; then
cat <<EOF && exit 1
This script is intended to be used with GPUs but you have not compiled Kaldi with CUDA
If you want to use GPUs (and have them), go to src/, and configure and make on a machine
where "nvcc" is installed.
EOF
fi
lang=data/lang_e2e
treedir=exp/chain/e2e_monotree # it's actually just a trivial tree (no tree building)
dir=exp/chain/e2e_cnn_${affix}
if [ $stage -le 0 ]; then
# Create a version of the lang/ directory that has one state per phone in the
# topo file. [note, it really has two states.. the first one is only repeated
# once, the second one has zero or more repeats.]
rm -rf $lang
cp -r data/lang $lang
silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
fi
if [ $stage -le 1 ]; then
steps/nnet3/chain/e2e/prepare_e2e.sh --nj $nj --cmd "$cmd" \
--shared-phones true \
--type mono \
data/$train_set $lang $treedir
$cmd $treedir/log/make_phone_lm.log \
cat data/$train_set/text \| \
steps/nnet3/chain/e2e/text_to_phones.py data/lang \| \
utils/sym2int.pl -f 2- data/lang/phones.txt \| \
chain-est-phone-lm --num-extra-lm-states=500 \
ark:- $treedir/phone_lm.fst
fi
if [ $stage -le 2 ]; then
echo "$0: creating neural net configs using the xconfig parser";
num_targets=$(tree-info $treedir/tree | grep num-pdfs | awk '{print $2}')
cnn_opts="l2-regularize=0.075"
tdnn_opts="l2-regularize=0.075"
output_opts="l2-regularize=0.1"
common1="$cnn_opts required-time-offsets= height-offsets=-2,-1,0,1,2 num-filters-out=36"
common2="$cnn_opts required-time-offsets= height-offsets=-2,-1,0,1,2 num-filters-out=70"
common3="$cnn_opts required-time-offsets= height-offsets=-1,0,1 num-filters-out=70"
mkdir -p $dir/configs
cat <<EOF > $dir/configs/network.xconfig
input dim=40 name=input
conv-relu-batchnorm-layer name=cnn1 height-in=40 height-out=40 time-offsets=-3,-2,-1,0,1,2,3 $common1
conv-relu-batchnorm-layer name=cnn2 height-in=40 height-out=20 time-offsets=-2,-1,0,1,2 $common1 height-subsample-out=2
conv-relu-batchnorm-layer name=cnn3 height-in=20 height-out=20 time-offsets=-4,-2,0,2,4 $common2
conv-relu-batchnorm-layer name=cnn4 height-in=20 height-out=20 time-offsets=-4,-2,0,2,4 $common2
conv-relu-batchnorm-layer name=cnn5 height-in=20 height-out=10 time-offsets=-4,-2,0,2,4 $common2 height-subsample-out=2
conv-relu-batchnorm-layer name=cnn6 height-in=10 height-out=10 time-offsets=-4,0,4 $common3
conv-relu-batchnorm-layer name=cnn7 height-in=10 height-out=10 time-offsets=-4,0,4 $common3
relu-batchnorm-layer name=tdnn1 input=Append(-4,0,4) dim=$tdnn_dim $tdnn_opts
relu-batchnorm-layer name=tdnn2 input=Append(-4,0,4) dim=$tdnn_dim $tdnn_opts
relu-batchnorm-layer name=tdnn3 input=Append(-4,0,4) dim=$tdnn_dim $tdnn_opts
## adding the layers for chain branch
relu-batchnorm-layer name=prefinal-chain dim=$tdnn_dim target-rms=0.5 $output_opts
output-layer name=output include-log-softmax=false dim=$num_targets max-change=1.5 $output_opts
EOF
steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs
fi
if [ $stage -le 3 ]; then
steps/nnet3/chain/e2e/train_e2e.py --stage $train_stage \
--cmd "$cmd" \
--feat.cmvn-opts "$cmvn_opts" \
--chain.leaky-hmm-coefficient 0.1 \
--chain.apply-deriv-weights true \
--egs.stage $get_egs_stage \
--egs.opts "--num_egs_diagnostic 100 --num_utts_subset 400" \
--chain.frame-subsampling-factor 4 \
--chain.alignment-subsampling-factor 4 \
--trainer.add-option="--optimization.memory-compression-level=2" \
--trainer.num-chunk-per-minibatch $minibatch_size \
--trainer.frames-per-iter 1500000 \
--trainer.num-epochs 3 \
--trainer.optimization.momentum 0 \
--trainer.optimization.num-jobs-initial 5 \
--trainer.optimization.num-jobs-final 8 \
--trainer.optimization.initial-effective-lrate 0.001 \
--trainer.optimization.final-effective-lrate 0.0001 \
--trainer.optimization.shrink-value 1.0 \
--trainer.max-param-change 2.0 \
--cleanup.remove-egs true \
--feat-dir data/${train_set} \
--tree-dir $treedir \
--dir $dir || exit 1;
fi
if [ $stage -le 4 ] && $decode_e2e; then
# The reason we are using data/lang here, instead of $lang, is just to
# emphasize that it's not actually important to give mkgraph.sh the
# lang directory with the matched topology (since it gets the
# topology file from the model). So you could give it a different
# lang directory, one that contained a wordlist and LM of your choice,
# as long as phones.txt was compatible.
utils/mkgraph.sh \
--self-loop-scale 1.0 $lang_decode \
$dir $dir/graph || exit 1;
fi
if [ $stage -le 5 ] && $decode_e2e; then
frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
--nj 30 --cmd "$cmd" --beam 12 \
$dir/graph data/test $dir/decode_test || exit 1;
steps/lmrescore_const_arpa.sh --cmd "$cmd" $lang_decode $lang_rescore \
data/test $dir/decode_test{,_rescored} || exit 1
echo "Done. Date: $(date). Results:"
local/chain/compare_wer.sh $dir
fi