run_tdnn.sh
5.12 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
#!/bin/bash
# this is a script to train the nnet3 TDNN acoustic model
stage=1
affix=
train_stage=-10
reporting_email=
common_egs_dir=
remove_egs=true
egs_stage=0
. ./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. Otherwise, call this script with --use-gpu false
EOF
fi
# do the common parts of the script.
local/nnet3/run_ivector_common.sh --stage $stage || exit 1;
ali_dir=exp/tri5a_rvb_ali
dir=exp/nnet3/tdnn
dir=$dir${affix:+_$affix}
if [ $stage -le 7 ]; then
echo "$0: creating neural net configs using the xconfig parser";
num_targets=$(tree-info $ali_dir/tree | grep num-pdfs | awk '{print $2}')
mkdir -p $dir/configs
cat <<EOF > $dir/configs/network.xconfig
input dim=100 name=ivector
input dim=40 name=input
# please note that it is important to have input layer with the name=input
# as the layer immediately preceding the fixed-affine-layer to enable
# the use of short notation for the descriptor
fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
# the first splicing is moved before the lda layer, so no splicing here
relu-renorm-layer name=tdnn1 dim=1248
relu-renorm-layer name=tdnn2 dim=1248 input=Append(-1,2)
relu-renorm-layer name=tdnn3 dim=1248 input=Append(-3,3)
relu-renorm-layer name=tdnn4 dim=1248 input=Append(-3,3)
relu-renorm-layer name=tdnn5 dim=1248 input=Append(-7,2)
relu-renorm-layer name=tdnn6 dim=1248
output-layer name=output dim=$num_targets max-change=1.5
EOF
steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
fi
if [ $stage -le 8 ]; then
if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
utils/create_split_dir.pl \
/export/b0{3,4,5,6}/$USER/kaldi-data/egs/aspire-$(date +'%m_%d_%H_%M')/s5/$dir/egs/storage $dir/egs/storage
fi
steps/nnet3/train_dnn.py --stage=$train_stage \
--cmd="$decode_cmd" \
--feat.online-ivector-dir exp/nnet3/ivectors_train \
--feat.cmvn-opts="--norm-means=false --norm-vars=false" \
--trainer.num-epochs 3 \
--trainer.optimization.num-jobs-initial 4 \
--trainer.optimization.num-jobs-final 22 \
--trainer.optimization.initial-effective-lrate 0.0017 \
--trainer.optimization.final-effective-lrate 0.00017 \
--egs.dir "$common_egs_dir" \
--egs.stage "$egs_stage" \
--cleanup.remove-egs $remove_egs \
--cleanup.preserve-model-interval 50 \
--feat-dir=data/train_rvb_hires \
--ali-dir $ali_dir \
--lang data/lang \
--reporting.email="$reporting_email" \
--dir=$dir || exit 1;
fi
#ASpIRE decodes
if [ $stage -le 9 ]; then
local/nnet3/prep_test_aspire.sh --stage 1 --decode-num-jobs 30 --affix "v7" \
--window 10 --overlap 5 \
--sub-speaker-frames 6000 --max-count 75 --ivector-scale 0.75 \
--pass2-decode-opts "--min-active 1000" \
dev_aspire data/lang exp/tri5a/graph_pp $dir
fi
exit 0;
# final result
# %WER 31.0 | 2120 27217 | 74.8 16.1 9.1 5.9 31.0 77.9 | -0.707 | exp/nnet3/tdnn/decode_dev_aspire_whole_uniformsegmented_win10_over5_v6_200jobs_iterfinal_pp_fg/score_14/penalty_0.0/ctm.filt.filt.sys
# intermediate results
#%WER 34.2 | 2120 27212 | 71.6 18.3 10.2 5.8 34.2 80.2 | -0.613 | exp/nnet3/tdnn/decode_dev_aspire_whole_uniformsegmented_win10_over5_v6_200jobs_iter100_pp_fg/score_14/penalty_0.0/ctm.filt.filt.sys
#%WER 32.8 | 2120 27212 | 73.2 17.3 9.4 6.0 32.8 79.3 | -0.657 | exp/nnet3/tdnn/decode_dev_aspire_whole_uniformsegmented_win10_over5_v6_200jobs_iter200_pp_fg/score_13/penalty_0.0/ctm.filt.filt.sys
#%WER 32.3 | 2120 27215 | 73.7 17.1 9.2 6.0 32.3 79.7 | -0.676 | exp/nnet3/tdnn/decode_dev_aspire_whole_uniformsegmented_win10_over5_v6_200jobs_iter300_pp_fg/score_13/penalty_0.0/ctm.filt.filt.sys
#%WER 31.7 | 2120 27215 | 74.3 16.8 8.9 6.0 31.7 78.9 | -0.690 | exp/nnet3/tdnn/decode_dev_aspire_whole_uniformsegmented_win10_over5_v6_200jobs_iter400_pp_fg/score_13/penalty_0.0/ctm.filt.filt.sys
#%WER 31.6 | 2120 27216 | 74.5 16.6 8.8 6.1 31.6 79.7 | -0.723 | exp/nnet3/tdnn/decode_dev_aspire_whole_uniformsegmented_win10_over5_v6_200jobs_iter500_pp_fg/score_13/penalty_0.0/ctm.filt.filt.sys
#%WER 31.3 | 2120 27216 | 74.9 16.6 8.5 6.2 31.3 78.4 | -0.737 | exp/nnet3/tdnn/decode_dev_aspire_whole_uniformsegmented_win10_over5_v6_200jobs_iter600_pp_fg/score_12/penalty_0.0/ctm.filt.filt.sys
#%WER 31.2 | 2120 27216 | 74.7 16.2 9.1 5.9 31.2 79.0 | -0.708 | exp/nnet3/tdnn/decode_dev_aspire_whole_uniformsegmented_win10_over5_v6_200jobs_iter700_pp_fg/score_14/penalty_0.0/ctm.filt.filt.sys
#%WER 31.1 | 2120 27219 | 74.7 16.4 8.9 5.9 31.1 78.4 | -0.732 | exp/nnet3/tdnn/decode_dev_aspire_whole_uniformsegmented_win10_over5_v6_200jobs_iter800_pp_fg/score_12/penalty_0.25/ctm.filt.filt.sys
#%WER 31.1 | 2120 27220 | 74.9 16.3 8.8 6.0 31.1 78.1 | -0.719 | exp/nnet3/tdnn/decode_dev_aspire_whole_uniformsegmented_win10_over5_v6_200jobs_iter1000_pp_fg/score_13/penalty_0.0/ctm.filt.filt.sys
exit 0;