run_tdnn.sh
3.62 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
#!/bin/bash
# This script is based on swbd/s5c/local/nnet3/run_tdnn.sh
# this is the standard "tdnn" system, built in nnet3; it's what we use to
# call multi-splice.
# At this script level we don't support not running on GPU, as it would be painfully slow.
# If you want to run without GPU you'd have to call train_tdnn.sh with --gpu false,
# --num-threads 16 and --minibatch-size 128.
set -e
stage=0
train_stage=-10
affix=
common_egs_dir=
# training options
initial_effective_lrate=0.0015
final_effective_lrate=0.00015
num_epochs=4
num_jobs_initial=2
num_jobs_final=8
remove_egs=false
# feature options
use_ivectors=true
# End configuration section.
. ./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
dir=exp/nnet3/tdnn_sp${affix:+_$affix}
gmm_dir=exp/tri5a
train_set=train_sp
ali_dir=${gmm_dir}_sp_ali
graph_dir=$gmm_dir/graph
local/nnet3/run_ivector_common.sh --stage $stage || exit 1;
if [ $stage -le 7 ]; then
echo "$0: creating neural net configs";
num_targets=$(tree-info $ali_dir/tree |grep num-pdfs|awk '{print $2}')
mkdir -p $dir/configs
cat <<EOF > $dir/configs/network.xconfig
input dim=100 name=ivector
input dim=43 name=input
# please note that it is important to have input layer with the name=input
# as the layer immediately preceding the fixed-affine-layer to enable
# the use of short notation for the descriptor
fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
# the first splicing is moved before the lda layer, so no splicing here
relu-batchnorm-layer name=tdnn1 dim=850
relu-batchnorm-layer name=tdnn2 dim=850 input=Append(-1,0,2)
relu-batchnorm-layer name=tdnn3 dim=850 input=Append(-3,0,3)
relu-batchnorm-layer name=tdnn4 dim=850 input=Append(-7,0,2)
relu-batchnorm-layer name=tdnn5 dim=850 input=Append(-3,0,3)
relu-batchnorm-layer name=tdnn6 dim=850
output-layer name=output input=tdnn6 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 [ $stage -le 8 ]; then
steps/nnet3/train_dnn.py --stage=$train_stage \
--cmd="$decode_cmd" \
--feat.online-ivector-dir exp/nnet3/ivectors_${train_set} \
--feat.cmvn-opts="--norm-means=false --norm-vars=false" \
--trainer.num-epochs $num_epochs \
--trainer.optimization.num-jobs-initial $num_jobs_initial \
--trainer.optimization.num-jobs-final $num_jobs_final \
--trainer.optimization.initial-effective-lrate $initial_effective_lrate \
--trainer.optimization.final-effective-lrate $final_effective_lrate \
--egs.dir "$common_egs_dir" \
--cleanup.remove-egs $remove_egs \
--cleanup.preserve-model-interval 500 \
--use-gpu wait \
--feat-dir=data/${train_set}_hires \
--ali-dir $ali_dir \
--lang data/lang \
--reporting.email="$reporting_email" \
--dir=$dir || exit 1;
fi
if [ $stage -le 9 ]; then
# this version of the decoding treats each utterance separately
# without carrying forward speaker information.
for decode_set in test eval; do
num_jobs=`cat data/${decode_set}_hires/utt2spk|cut -d' ' -f2|sort -u|wc -l`
decode_dir=${dir}/decode_$decode_set
steps/nnet3/decode.sh --nj $num_jobs --cmd "$decode_cmd" \
--online-ivector-dir exp/nnet3/ivectors_${decode_set} \
$graph_dir data/${decode_set}_hires $decode_dir || exit 1;
done
wait;
fi
exit 0;