run_tdnn_1g.sh
11 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
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
#!/bin/bash
# 1g is like 1f but upgrading to a "resnet-style TDNN-F model", i.e.
# with bypass resnet connections, and re-tuned.
# compute-wer --text --mode=present ark:exp/chain/multipsplice_tdnn/decode_fsp_train_test/scoring_kaldi/test_filt.txt ark,p:-
# %WER 22.21 [ 8847 / 39831, 1965 ins, 2127 del, 4755 sub ]
# %SER 56.98 [ 3577 / 6278 ]
# Scored 6278 sentences, 0 not present in hyp.
# steps/info/chain_dir_info.pl exp/chain/multipsplice_tdnn
# exp/chain/multipsplice_tdnn: num-iters=296 nj=1..2 num-params=8.2M dim=40+100->2489 combine=-0.170->-0.165 (over 8) xent:train/valid[196,295,final]=(-2.30,-1.93,-1.83/-2.24,-1.96,-1.86) logprob:train/valid[196,295,final]=(-0.208,-0.169,-0.164/-0.189,-0.161,-0.158)
set -e -o pipefail
# First the options that are passed through to run_ivector_common.sh
# (some of which are also used in this script directly).
stage=0
nj=30
train_set=train
test_sets="test dev"
gmm=tri5a # this is the source gmm-dir that we'll use for alignments; it
# should have alignments for the specified training data.
num_threads_ubm=32
nnet3_affix= # affix for exp dirs, e.g. it was _cleaned in tedlium.
# Options which are not passed through to run_ivector_common.sh
affix=1g #affix for TDNN+LSTM directory e.g. "1a" or "1b", in case we change the configuration.
common_egs_dir=
reporting_email=
# LSTM/chain options
train_stage=-10
xent_regularize=0.1
dropout_schedule='0,0@0.20,0.3@0.50,0'
# training chunk-options
chunk_width=140,100,160
# we don't need extra left/right context for TDNN systems.
chunk_left_context=0
chunk_right_context=0
# training options
srand=0
remove_egs=true
#decode options
test_online_decoding=false # if true, it will run the last decoding stage.
# 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
if [ $stage -le 15 ]; then
echo "local/nnet3/run_ivector_common.sh \
--stage $stage --nj $nj \
--train-set $train_set --gmm $gmm \
--num-threads-ubm $num_threads_ubm \
--nnet3-affix "$nnet3_affix""
local/nnet3/run_ivector_common.sh \
--stage $stage --nj $nj \
--train-set $train_set --gmm $gmm \
--num-threads-ubm $num_threads_ubm \
--nnet3-affix "$nnet3_affix"
fi
gmm_dir=exp/${gmm}
ali_dir=exp/${gmm}_ali_${train_set}_sp
lat_dir=exp/tri5a_lats_nodup_sp
dir=exp/chain/multipsplice_tdnn
train_data_dir=data/${train_set}_sp_hires
train_ivector_dir=exp/nnet3/ivectors_train_sp_hires
lores_train_data_dir=data/${train_set}_sp
# note: you don't necessarily have to change the treedir name
# each time you do a new experiment-- only if you change the
# configuration in a way that affects the tree.
tree_dir=exp/chain/${gmm}_tree
# the 'lang' directory is created by this script.
# If you create such a directory with a non-standard topology
# you should probably name it differently.
lang=data/lang_${gmm}_chain
#for f in $train_data_dir/feats.scp $train_ivector_dir/ivector_online.scp \
# $lores_train_data_dir/feats.scp $gmm_dir/final.mdl \
# $ali_dir/ali.1.gz $gmm_dir/final.mdl; do
# [ ! -f $f ] && echo "$0: expected file $f to exist" && exit 1
#done
if [ $stage -le 16 ]; then
echo "$0: creating lang directory $lang with chain-type topology"
# 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.]
if [ -d $lang ]; then
if [ $lang/L.fst -nt data/lang/L.fst ]; then
echo "$0: $lang already exists, not overwriting it; continuing"
else
echo "$0: $lang already exists and seems to be older than data/lang..."
echo " ... not sure what to do. Exiting."
exit 1;
fi
else
cp -r data/lang $lang
silphonelist=$(cat $lang/phones/silence.csl) || exit 1;
nonsilphonelist=$(cat $lang/phones/nonsilence.csl) || exit 1;
# Use our special topology... note that later on may have to tune this
# topology.
steps/nnet3/chain/gen_topo.py $nonsilphonelist $silphonelist >$lang/topo
fi
fi
if [ $stage -le 17 ]; then
# Get the alignments as lattices (gives the chain training more freedom).
# use the same num-jobs as the alignments
steps/align_fmllr_lats.sh --nj 100 --cmd "$train_cmd" ${lores_train_data_dir} \
data/lang $gmm_dir $lat_dir
rm $lat_dir/fsts.*.gz # save space
fi
if [ $stage -le 18 ]; then
# Build a tree using our new topology. We know we have alignments for the
# speed-perturbed data (local/nnet3/run_ivector_common.sh made them), so use
# those. The num-leaves is always somewhat less than the num-leaves from
# the GMM baseline.
if [ -f $tree_dir/final.mdl ]; then
echo "$0: $tree_dir/final.mdl already exists, refusing to overwrite it."
exit 1;
fi
steps/nnet3/chain/build_tree.sh \
--frame-subsampling-factor 3 \
--context-opts "--context-width=2 --central-position=1" \
--cmd "$train_cmd" 3500 ${lores_train_data_dir} \
$lang $ali_dir $tree_dir
fi
if [ $stage -le 19 ]; then
mkdir -p $dir
echo "$0: creating neural net configs using the xconfig parser";
num_targets=$(tree-info $tree_dir/tree |grep num-pdfs|awk '{print $2}')
learning_rate_factor=$(echo "print (0.5/$xent_regularize)" | python)
tdnn_opts="l2-regularize=0.01 dropout-proportion=0.0 dropout-per-dim-continuous=true"
tdnnf_opts="l2-regularize=0.01 dropout-proportion=0.0 bypass-scale=0.66"
linear_opts="l2-regularize=0.01 orthonormal-constraint=-1.0"
prefinal_opts="l2-regularize=0.01"
output_opts="l2-regularize=0.005"
mkdir -p $dir/configs
cat <<EOF > $dir/configs/network.xconfig
input dim=100 name=ivector
input dim=40 name=input
# please note that it is important to have input layer with the name=input
# as the layer immediately preceding the fixed-affine-layer to enable
# the use of short notation for the descriptor
fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
# the first splicing is moved before the lda layer, so no splicing here
relu-batchnorm-dropout-layer name=tdnn1 $tdnn_opts dim=1024
tdnnf-layer name=tdnnf2 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=1
tdnnf-layer name=tdnnf3 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=1
tdnnf-layer name=tdnnf4 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=1
tdnnf-layer name=tdnnf5 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=0
tdnnf-layer name=tdnnf6 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=3
tdnnf-layer name=tdnnf7 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=3
tdnnf-layer name=tdnnf8 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=3
tdnnf-layer name=tdnnf9 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=3
tdnnf-layer name=tdnnf10 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=3
tdnnf-layer name=tdnnf11 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=3
tdnnf-layer name=tdnnf12 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=3
tdnnf-layer name=tdnnf13 $tdnnf_opts dim=1024 bottleneck-dim=128 time-stride=3
linear-component name=prefinal-l dim=192 $linear_opts
prefinal-layer name=prefinal-chain input=prefinal-l $prefinal_opts big-dim=1024 small-dim=192
output-layer name=output include-log-softmax=false dim=$num_targets $output_opts
prefinal-layer name=prefinal-xent input=prefinal-l $prefinal_opts big-dim=1024 small-dim=192
output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor $output_opts
EOF
steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
fi
if [ $stage -le 20 ]; 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/wsj-$(date +'%m_%d_%H_%M')/s5/$dir/egs/storage $dir/egs/storage
fi
steps/nnet3/chain/train.py --stage=$train_stage \
--cmd "$decode_cmd" \
--feat.online-ivector-dir $train_ivector_dir \
--feat.cmvn-opts "--norm-means=false --norm-vars=false" \
--chain.xent-regularize $xent_regularize \
--chain.leaky-hmm-coefficient 0.1 \
--chain.l2-regularize 0.0 \
--chain.apply-deriv-weights false \
--chain.lm-opts="--num-extra-lm-states=2000" \
--trainer.dropout-schedule $dropout_schedule \
--trainer.srand $srand \
--trainer.max-param-change 2.0 \
--trainer.num-epochs 4 \
--trainer.frames-per-iter 5000000 \
--trainer.optimization.num-jobs-initial 1 \
--trainer.optimization.num-jobs-final=2 \
--trainer.optimization.initial-effective-lrate 0.0005 \
--trainer.optimization.final-effective-lrate 0.00005 \
--trainer.num-chunk-per-minibatch 128,64 \
--trainer.optimization.momentum 0.0 \
--egs.chunk-width $chunk_width \
--egs.chunk-left-context 0 \
--egs.chunk-right-context 0 \
--egs.dir "$common_egs_dir" \
--egs.opts "--frames-overlap-per-eg 0" \
--cleanup.remove-egs $remove_egs \
--use-gpu true \
--feat-dir $train_data_dir \
--tree-dir $tree_dir \
--lat-dir exp/tri5a_lats_nodup_sp \
--dir $dir || exit 1;
fi
if [ $stage -le 21 ]; then
# The reason we are using data/lang_test 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.
#LM was trained only on Fisher Spanish train subset.
utils/mkgraph.sh \
--self-loop-scale 1.0 data/lang_test \
$tree_dir $tree_dir/graph_fsp_train || exit 1;
fi
rnnlmdir=exp/rnnlm_lstm_tdnn_1b
if [ $stage -le 22 ]; then
local/rnnlm/train_rnnlm.sh --dir $rnnlmdir || exit 1;
fi
if [ $stage -le 23 ]; then
frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
rm $dir/.error 2>/dev/null || true
for data in $test_sets; do
(
nspk=$(wc -l <data/${data}_hires/spk2utt)
for lmtype in fsp_train; do
steps/nnet3/decode.sh \
--acwt 1.0 --post-decode-acwt 10.0 \
--extra-left-context 0 --extra-right-context 0 \
--extra-left-context-initial 0 \
--extra-right-context-final 0 \
--frames-per-chunk $frames_per_chunk \
--nj $nspk --cmd "$decode_cmd" --num-threads 4 \
--online-ivector-dir exp/nnet3/ivectors_${data}_hires \
$tree_dir/graph_${lmtype} data/${data}_hires ${dir}/decode_${lmtype}_${data} || exit 1;
done
bash local/rnnlm/lmrescore_nbest.sh 1.0 data/lang_test $rnnlmdir data/${data}_hires/ \
${dir}/decode_${lmtype}_${data} $dir/decode_rnnLM_${lmtype}_${data} || exit 1;
) || touch $dir/.error &
done
wait
[ -f $dir/.error ] && echo "$0: there was a problem while decoding" && exit 1
fi
exit 0;