Skip to content

[egs] Update TDNN-F script for Librispeech #2574

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
danpovey merged 1 commit into from
Jul 26, 2018
Merged

[egs] Update TDNN-F script for Librispeech #2574

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
2 changes: 1 addition & 1 deletion egs/librispeech/s5/local/chain/run_tdnn.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -1 +1 @@
tuning/run_tdnn_1c.sh
tuning/run_tdnn_1d.sh
363 changes: 363 additions & 0 deletions egs/librispeech/s5/local/chain/tuning/run_tdnn_1d.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,363 @@
#!/bin/bash
set -e

# 1d is as 1c but a recipe based on the newer, more compact configs, and with
# various configuration changes; it also includes dropout (although I'm not
# sure whether dropout was actually helpful, that needs to be tested).
#
# local/chain/compare_wer.sh exp/chain_cleaned/tdnn_1c_sp exp/chain_cleaned/tdnn_1d_sp
# System tdnn_1c_sp tdnn_1d_sp
# WER on dev(fglarge) 3.31 3.29
# WER on dev(tglarge) 3.41 3.44
# WER on dev(tgmed) 4.30 4.22
# WER on dev(tgsmall) 4.81 4.72
# WER on dev_other(fglarge) 8.73 8.71
# WER on dev_other(tglarge) 9.22 9.05
# WER on dev_other(tgmed) 11.24 11.09
# WER on dev_other(tgsmall) 12.29 12.13
# WER on test(fglarge) 3.88 3.80
# WER on test(tglarge) 4.05 3.89
# WER on test(tgmed) 4.86 4.72
# WER on test(tgsmall) 5.30 5.19
# WER on test_other(fglarge) 9.09 8.76
# WER on test_other(tglarge) 9.54 9.19
# WER on test_other(tgmed) 11.65 11.22
# WER on test_other(tgsmall) 12.77 12.24
# Final train prob -0.0510 -0.0378
# Final valid prob -0.0619 -0.0374
# Final train prob (xent) -0.7499 -0.6099
# Final valid prob (xent) -0.8118 -0.6353
# Num-parameters 20093920 22623456


#
# 1c23 is as 1c22 but with bypass-scale increased to 0.75 Better!
# 1c22 is as 1c21 but with bottleneck-dim reduced from 192 to 160.
# 1c21 is as 1c19 but with 2.5 million, instead of 5 million, frames-per-iter.
# 1c19 is a rerun of 1c{14,16} but with --constrained false in the egs.opts,
# and upgrading to new-style configs.
# 1c16 is (by mistake) a rerun of 1c14.

# local/chain/compare_wer.sh exp/chain_cleaned/tdnn_1c14_sp exp/chain_cleaned/tdnn_1c16_sp
# System tdnn_1c14_sp tdnn_1c16_sp
# WER on dev(fglarge) 3.38 3.34
# WER on dev(tglarge) 3.44 3.40
# WER on dev(tgmed) 4.33 4.34
# WER on dev(tgsmall) 4.80 4.79
# WER on dev_other(fglarge) 8.63 8.66
# WER on dev_other(tglarge) 9.04 9.11
# WER on dev_other(tgmed) 11.03 11.21
# WER on dev_other(tgsmall) 12.21 12.26
# WER on test(fglarge) 3.79 3.77
# WER on test(tglarge) 3.92 3.96
# WER on test(tgmed) 4.80 4.79
# WER on test(tgsmall) 5.34 5.31
# WER on test_other(fglarge) 8.94 8.94
# WER on test_other(tglarge) 9.35 9.28
# WER on test_other(tgmed) 11.32 11.28
# WER on test_other(tgsmall) 12.43 12.39
# Final train prob -0.0491 -0.0486
# Final valid prob -0.0465 -0.0465
# Final train prob (xent) -0.6463 -0.6371
# Final valid prob (xent) -0.6668 -0.6593
# Num-parameters 23701728 23701728

# 1c14 is as 1c13 but with two more layers.
# A bit better! Overfits slightly.
# local/chain/compare_wer.sh exp/chain_cleaned/tdnn_1c_sp exp/chain_cleaned/tdnn_1c10_sp exp/chain_cleaned/tdnn_1c11_sp exp/chain_cleaned/tdnn_1c12_sp exp/chain_cleaned/tdnn_1c13_sp exp/chain_cleaned/tdnn_1c14_sp
# System tdnn_1c_sp tdnn_1c10_sp tdnn_1c11_sp tdnn_1c12_sp tdnn_1c13_sp tdnn_1c14_sp
# WER on dev(fglarge) 3.31 3.43 3.37 3.36 3.33 3.38
# WER on dev(tglarge) 3.41 3.50 3.45 3.43 3.40 3.44
# WER on dev(tgmed) 4.30 4.37 4.30 4.40 4.25 4.33
# WER on dev(tgsmall) 4.81 4.79 4.82 4.86 4.74 4.80
# WER on dev_other(fglarge) 8.73 9.10 8.61 8.49 8.78 8.63
# WER on dev_other(tglarge) 9.22 9.46 9.11 8.92 9.23 9.04
# WER on dev_other(tgmed) 11.24 11.33 11.23 10.91 11.10 11.03
# WER on dev_other(tgsmall) 12.29 12.58 12.23 12.07 12.33 12.21
# WER on test(fglarge) 3.88 3.86 3.83 3.78 3.84 3.79
# WER on test(tglarge) 4.05 4.01 3.96 3.93 3.96 3.92
# WER on test(tgmed) 4.86 4.80 4.83 4.81 4.77 4.80
# WER on test(tgsmall) 5.30 5.31 5.24 5.24 5.22 5.34
# WER on test_other(fglarge) 9.09 9.02 9.05 8.88 9.02 8.94
# WER on test_other(tglarge) 9.54 9.58 9.47 9.20 9.42 9.35
# WER on test_other(tgmed) 11.65 11.63 11.35 11.28 11.46 11.32
# WER on test_other(tgsmall) 12.77 12.69 12.51 12.38 12.60 12.43
# Final train prob -0.0510 -0.0423 -0.0449 -0.0517 -0.0460 -0.0491
# Final valid prob -0.0619 -0.0446 -0.0456 -0.0503 -0.0460 -0.0465
# Final train prob (xent) -0.7499 -0.5974 -0.6351 -0.6660 -0.6329 -0.6463
# Final valid prob (xent) -0.8118 -0.6331 -0.6612 -0.6854 -0.6588 -0.6668
# Num-parameters 20093920 21339360 21339360 22297824 21339360 23701728

# 1c13 is as 1c12 but changing tdnnf5-layer back to tdnnf6-layer.
# 1c12 is as 1c11 but with changes to the learning rates (reduced) and l2
# (doubled for non-final layers), a larger frames-per-iter, and
# changing to tdnnf5-layer, i.e. keeping the extra splicing.
# 1c11 is as 1c10 but with double the l2-regularize.
# 1c10 is as 1c but using a newer type of setup based on the Swbd
# setup I'm working on, with tdnnf6-layers.
# Basing it on 7p10m. Making it 4 epochs, for speed.

# 7n is a kind of factorized TDNN, with skip connections

# steps/info/chain_dir_info.pl exp/chain_cleaned/tdnn_1c_sp
# exp/chain_cleaned/tdnn_1c_sp: num-iters=1307 nj=3..16 num-params=20.1M dim=40+100->6024 combine=-0.051->-0.050 (over 23) xent:train/valid[869,1306,final]=(-0.808,-0.767,-0.771/-0.828,-0.780,-0.787) logprob:train/valid[869,1306,final]=(-0.051,-0.049,-0.047/-0.059,-0.056,-0.056)

# local/chain/compare_wer.sh exp/chain_cleaned/tdnn_1b_sp exp/chain_cleaned/tdnn_1c_sp
# System tdnn_1b_sp tdnn_1c_sp
# WER on dev(fglarge) 3.77 3.35
# WER on dev(tglarge) 3.90 3.49
# WER on dev(tgmed) 4.89 4.30
# WER on dev(tgsmall) 5.47 4.78
# WER on dev_other(fglarge) 10.05 8.76
# WER on dev_other(tglarge) 10.80 9.26
# WER on dev_other(tgmed) 13.07 11.21
# WER on dev_other(tgsmall) 14.46 12.47
# WER on test(fglarge) 4.20 3.87
# WER on test(tglarge) 4.28 4.08
# WER on test(tgmed) 5.31 4.80
# WER on test(tgsmall) 5.97 5.25
# WER on test_other(fglarge) 10.44 8.95
# WER on test_other(tglarge) 11.05 9.41
# WER on test_other(tgmed) 13.36 11.52
# WER on test_other(tgsmall) 14.90 12.66
# Final train prob -0.0670 -0.0475
# Final valid prob -0.0704 -0.0555
# Final train prob (xent) -1.0502 -0.7708
# Final valid prob (xent) -1.0441 -0.7874

# configs for 'chain'
stage=0
decode_nj=50
train_set=train_960_cleaned
gmm=tri6b_cleaned
nnet3_affix=_cleaned

# The rest are configs specific to this script. Most of the parameters
# are just hardcoded at this level, in the commands below.
affix=1d
tree_affix=
train_stage=-10
get_egs_stage=-10
decode_iter=

# TDNN options
frames_per_eg=150,110,100
remove_egs=true
common_egs_dir=
xent_regularize=0.1
dropout_schedule='0,0@0.20,0.5@0.50,0'

test_online_decoding=true # 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

# The iVector-extraction and feature-dumping parts are the same as the standard
# nnet3 setup, and you can skip them by setting "--stage 11" if you have already
# run those things.

local/nnet3/run_ivector_common.sh --stage $stage \
--train-set $train_set \
--gmm $gmm \
--num-threads-ubm 6 --num-processes 3 \
--nnet3-affix "$nnet3_affix" || exit 1;

gmm_dir=exp/$gmm
ali_dir=exp/${gmm}_ali_${train_set}_sp
tree_dir=exp/chain${nnet3_affix}/tree_sp${tree_affix:+_$tree_affix}
lang=data/lang_chain
lat_dir=exp/chain${nnet3_affix}/${gmm}_${train_set}_sp_lats
dir=exp/chain${nnet3_affix}/tdnn${affix:+_$affix}_sp
train_data_dir=data/${train_set}_sp_hires
lores_train_data_dir=data/${train_set}_sp
train_ivector_dir=exp/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires

# if we are using the speed-perturbed data we need to generate
# alignments for it.

for f in $gmm_dir/final.mdl $train_data_dir/feats.scp $train_ivector_dir/ivector_online.scp \
$lores_train_data_dir/feats.scp $ali_dir/ali.1.gz; do
[ ! -f $f ] && echo "$0: expected file $f to exist" && exit 1
done

# Please take this as a reference on how to specify all the options of
# local/chain/run_chain_common.sh
local/chain/run_chain_common.sh --stage $stage \
--gmm-dir $gmm_dir \
--ali-dir $ali_dir \
--lores-train-data-dir ${lores_train_data_dir} \
--lang $lang \
--lat-dir $lat_dir \
--num-leaves 7000 \
--tree-dir $tree_dir || exit 1;

if [ $stage -le 14 ]; then
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)
affine_opts="l2-regularize=0.008 dropout-proportion=0.0 dropout-per-dim=true dropout-per-dim-continuous=true"
tdnnf_opts="l2-regularize=0.008 dropout-proportion=0.0 bypass-scale=0.75"
linear_opts="l2-regularize=0.008 orthonormal-constraint=-1.0"
prefinal_opts="l2-regularize=0.008"
output_opts="l2-regularize=0.002"

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 $affine_opts dim=1536
tdnnf-layer name=tdnnf2 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=1
tdnnf-layer name=tdnnf3 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=1
tdnnf-layer name=tdnnf4 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=1
tdnnf-layer name=tdnnf5 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=0
tdnnf-layer name=tdnnf6 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf7 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf8 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf9 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf10 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf11 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf12 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf13 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf14 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf15 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf16 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf17 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
linear-component name=prefinal-l dim=256 $linear_opts

prefinal-layer name=prefinal-chain input=prefinal-l $prefinal_opts big-dim=1536 small-dim=256
output-layer name=output include-log-softmax=false dim=$num_targets $output_opts

prefinal-layer name=prefinal-xent input=prefinal-l $prefinal_opts big-dim=1536 small-dim=256
output-layer name=output-xent dim=$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 15 ]; then
if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
utils/create_split_dir.pl \
/export/b{09,10,11,12}/$USER/kaldi-data/egs/swbd-$(date +'%m_%d_%H_%M')/s5c/$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 $xent_regularize \
--chain.leaky-hmm-coefficient 0.1 \
--chain.l2-regularize 0.0 \
--chain.apply-deriv-weights false \
--chain.lm-opts="--num-extra-lm-states=2000" \
--egs.dir "$common_egs_dir" \
--egs.stage $get_egs_stage \
--egs.opts "--frames-overlap-per-eg 0 --constrained false" \
--egs.chunk-width $frames_per_eg \
--trainer.dropout-schedule $dropout_schedule \
--trainer.add-option="--optimization.memory-compression-level=2" \
--trainer.num-chunk-per-minibatch 64 \
--trainer.frames-per-iter 2500000 \
--trainer.num-epochs 4 \
--trainer.optimization.num-jobs-initial 3 \
--trainer.optimization.num-jobs-final 16 \
--trainer.optimization.initial-effective-lrate 0.00015 \
--trainer.optimization.final-effective-lrate 0.000015 \
--trainer.max-param-change 2.0 \
--cleanup.remove-egs $remove_egs \
--feat-dir $train_data_dir \
--tree-dir $tree_dir \
--lat-dir $lat_dir \
--dir $dir || exit 1;

fi

graph_dir=$dir/graph_tgsmall
if [ $stage -le 16 ]; then
# Note: it might appear that this $lang directory is mismatched, and it is as
# far as the 'topo' is concerned, but this script doesn't read the 'topo' from
# the lang directory.
utils/mkgraph.sh --self-loop-scale 1.0 --remove-oov data/lang_test_tgsmall $dir $graph_dir
# remove <UNK> from the graph, and convert back to const-FST.
fstrmsymbols --apply-to-output=true --remove-arcs=true "echo 3|" $graph_dir/HCLG.fst - | \
fstconvert --fst_type=const > $graph_dir/temp.fst
mv $graph_dir/temp.fst $graph_dir/HCLG.fst
fi

iter_opts=
if [ ! -z $decode_iter ]; then
iter_opts=" --iter $decode_iter "
fi
if [ $stage -le 17 ]; then
rm $dir/.error 2>/dev/null || true
for decode_set in test_clean test_other dev_clean dev_other; do
(
steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
--nj $decode_nj --cmd "$decode_cmd" $iter_opts \
--online-ivector-dir exp/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
$graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}_tgsmall || exit 1
steps/lmrescore.sh --cmd "$decode_cmd" --self-loop-scale 1.0 data/lang_test_{tgsmall,tgmed} \
data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}_{tgsmall,tgmed} || exit 1
steps/lmrescore_const_arpa.sh \
--cmd "$decode_cmd" data/lang_test_{tgsmall,tglarge} \
data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}_{tgsmall,tglarge} || exit 1
steps/lmrescore_const_arpa.sh \
--cmd "$decode_cmd" data/lang_test_{tgsmall,fglarge} \
data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}_{tgsmall,fglarge} || exit 1
) || touch $dir/.error &
done
wait
if [ -f $dir/.error ]; then
echo "$0: something went wrong in decoding"
exit 1
fi
fi

if $test_online_decoding && [ $stage -le 18 ]; then
# note: if the features change (e.g. you add pitch features), you will have to
# change the options of the following command line.
steps/online/nnet3/prepare_online_decoding.sh \
--mfcc-config conf/mfcc_hires.conf \
$lang exp/nnet3${nnet3_affix}/extractor $dir ${dir}_online

rm $dir/.error 2>/dev/null || true
for data in test_clean test_other dev_clean dev_other; do
(
nspk=$(wc -l <data/${data}_hires/spk2utt)
# note: we just give it "data/${data}" as it only uses the wav.scp, the
# feature type does not matter.
steps/online/nnet3/decode.sh \
--acwt 1.0 --post-decode-acwt 10.0 \
--nj $nspk --cmd "$decode_cmd" \
$graph_dir data/${data} ${dir}_online/decode_${data}_tgsmall || exit 1

) || touch $dir/.error &
done
wait
if [ -f $dir/.error ]; then
echo "$0: something went wrong in decoding"
exit 1
fi
fi


exit 0;