Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
27 commits
Select commit Hold shift + click to select a range
13e8bed
[src,scripts,egs] nnet3,fast-lstm: changes to support separate per-fr…
danpovey Jan 31, 2017
863534b
[egs] Small fixes/additions in Swbd/s5c chain scripts
danpovey Jan 31, 2017
8384eae
Merge branch 'shortcut' into shortcut-dropout
danpovey Jan 31, 2017
eb0f458
[src,egs,scripts] Modifying dropout in LSTM to be on (i,f,o) gates no…
danpovey Jan 31, 2017
96d92d7
Merge remote-tracking branch 'upstream/shortcut' into shortcut-dropout
danpovey Jan 31, 2017
19af8ca
Merge remote-tracking branch 'upstream/shortcut' into shortcut-dropout
danpovey Jan 31, 2017
4d2f00e
Merge branch 'shortcut' into shortcut-dropout
danpovey Feb 2, 2017
6582acf
[scripts] Update example scripts for dropout on Tedlium s5_r2
danpovey Feb 3, 2017
a406d0f
Merge branch 'shortcut' into shortcut-dropout
danpovey Feb 3, 2017
eb94ffd
for ref
GaofengCheng Mar 21, 2017
b9c3e20
merge fast lstm dropout
GaofengCheng Apr 9, 2017
9afaf39
delete temporary tuning sdripts in tedlium
GaofengCheng Apr 9, 2017
e9ac4e2
delete irrelevant file
GaofengCheng Apr 9, 2017
638f083
delete exclusive option in fast lstm code
GaofengCheng Apr 9, 2017
49c4558
solve some cuda-kernel line mismatch problem
GaofengCheng Apr 9, 2017
05fc6d2
small bug fix
GaofengCheng Apr 9, 2017
90df5d7
small fix
GaofengCheng Apr 9, 2017
1a58236
update scripts for tdnn-(fast)lstm of AMI-IHM
GaofengCheng Apr 11, 2017
69a36e4
change scripts comment style and RESULTS
GaofengCheng Apr 11, 2017
d03be0f
adding SDM results
GaofengCheng Apr 12, 2017
07d6774
Merge branch 'master' of https://github.com/kaldi-asr/kaldi into nnet…
GaofengCheng Apr 13, 2017
936863e
adding SWBD (parts of all) scripts with dropout
GaofengCheng Apr 17, 2017
f51fb75
small fix
GaofengCheng Apr 17, 2017
139f412
update tdnn-blstm with dropout in SWBD
GaofengCheng Apr 18, 2017
9a8b81c
update tdnn+regular-LSTM(4epoch) in SWBD
GaofengCheng Apr 18, 2017
48f41a7
adding tedlium scripts
GaofengCheng Apr 20, 2017
62fee2b
small fix
GaofengCheng Apr 20, 2017
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
5 changes: 5 additions & 0 deletions egs/ami/s5b/RESULTS_ihm
Original file line number Diff line number Diff line change
Expand Up @@ -84,6 +84,11 @@
%WER 20.8 | 13098 94489 | 82.0 10.0 8.0 2.8 20.8 53.2 | -0.096 | exp/ihm/chain_cleaned/tdnn_lstm1i_sp_bi_ld5/decode_dev/ascore_11/dev_hires.ctm.filt.sys
%WER 20.7 | 12643 89980 | 81.7 11.5 6.8 2.5 20.7 51.8 | 0.015 | exp/ihm/chain_cleaned/tdnn_lstm1i_sp_bi_ld5/decode_eval/ascore_11/eval_hires.ctm.filt.sys

# local/chain/tuning/run_tdnn_lstm_1l.sh --mic ihm --train-set train_cleaned --gmm tri3_cleaned
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

At this level, please just include the results for the "recommended" system which is 1m.
You should put all the comparative results in the individual scripts inside local/chain/tuning.
Use the standard compare_wer.sh script, whatever it's called, and also include the output
of chain_dir_info.pl from each of those scripts, in a comment in that script.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@danpovey OK, will do

# same as local/chain/tuning/run_tdnn_lstm_1i.sh, except that dropout is adopted
# cleanup + chain TDNN+LSTM model + per-frame dropout
%WER 19.8 | 13098 94475 | 83.1 9.6 7.4 2.8 19.8 51.8 | -0.041 | exp/ihm/chain_cleaned/tdnn_lstm1l_sp_bi_ld5/decode_dev/ascore_10/dev_hires.ctm.filt.sys
%WER 19.2 | 12643 89964 | 83.2 10.7 6.1 2.5 19.2 49.7 | 0.079 | exp/ihm/chain_cleaned/tdnn_lstm1l_sp_bi_ld5/decode_eval/ascore_10/eval_hires.ctm.filt.sys

# local/chain/multi_condition/tuning/run_tdnn_lstm_1a.sh --mic ihm
# cleanup + chain TDNN+LSTM model + IHM reverberated data
Expand Down
5 changes: 5 additions & 0 deletions egs/ami/s5b/RESULTS_sdm
Original file line number Diff line number Diff line change
Expand Up @@ -91,6 +91,11 @@
%WER 37.6 | 15122 94495 | 66.1 18.7 15.1 3.7 37.6 63.2 | 0.646 | exp/sdm1/chain_cleaned/tdnn_lstm1i_sp_bi_ihmali_ld5/decode_dev/ascore_10/dev_hires_o4.ctm.filt.sys
%WER 40.9 | 13807 89961 | 62.4 20.0 17.6 3.3 40.9 65.7 | 0.612 | exp/sdm1/chain_cleaned/tdnn_lstm1i_sp_bi_ihmali_ld5/decode_eval/ascore_10/eval_hires_o4.ctm.filt.sys

# local/chain/tuning/run_tdnn_lstm_1l.sh --mic sdm1 --use-ihm-ali true --train-set train_cleaned --gmm tri3_cleaned
# same as local/chain/tuning/run_tdnn_lstm_1i.sh, except that dropout is adopted
# cleanup + chain TDNN+LSTM model, SDM audio + alignments from ihm data + per-frame dropout.
%WER 35.9 | 14900 94497 | 67.8 18.2 14.1 3.7 35.9 62.5 | 0.647 | exp/sdm1/chain_cleaned/tdnn_lstm1l_sp_bi_ihmali_ld5/decode_dev/ascore_9/dev_hires_o4.ctm.filt.sys
%WER 39.4 | 13223 89946 | 64.1 19.7 16.2 3.5 39.4 67.0 | 0.611 | exp/sdm1/chain_cleaned/tdnn_lstm1l_sp_bi_ihmali_ld5/decode_eval/ascore_9/eval_hires_o4.ctm.filt.sys

# local/chain/multi_condition/tuning/run_tdnn_lstm_1a.sh --mic sdm1 --use-ihm-ali true --train-set train_cleaned --gmm tri3_cleaned
# cleanup + chain TDNN+LSTM model, SDM original + IHM reverberated data, alignments from ihm data.
Expand Down
3 changes: 2 additions & 1 deletion egs/ami/s5b/local/chain/tuning/run_tdnn_lstm_1i.sh
Original file line number Diff line number Diff line change
Expand Up @@ -26,6 +26,7 @@ gmm=tri3_cleaned # the gmm for the target data
ihm_gmm=tri3 # the gmm for the IHM system (if --use-ihm-ali true).
num_threads_ubm=32
nnet3_affix=_cleaned # cleanup affix for nnet3 and chain dirs, e.g. _cleaned
num_epochs=4

chunk_width=150
chunk_left_context=40
Expand Down Expand Up @@ -242,7 +243,7 @@ if [ $stage -le 16 ]; then
--egs.chunk-right-context $chunk_right_context \
--trainer.num-chunk-per-minibatch 64 \
--trainer.frames-per-iter 1500000 \
--trainer.num-epochs 4 \
--trainer.num-epochs $num_epochs \
--trainer.optimization.shrink-value 0.99 \
--trainer.optimization.num-jobs-initial 2 \
--trainer.optimization.num-jobs-final 12 \
Expand Down
3 changes: 2 additions & 1 deletion egs/ami/s5b/local/chain/tuning/run_tdnn_lstm_1j.sh
Original file line number Diff line number Diff line change
Expand Up @@ -34,6 +34,7 @@ gmm=tri3_cleaned # the gmm for the target data
ihm_gmm=tri3 # the gmm for the IHM system (if --use-ihm-ali true).
num_threads_ubm=32
nnet3_affix=_cleaned # cleanup affix for nnet3 and chain dirs, e.g. _cleaned
num_epochs=4

chunk_width=150
chunk_left_context=40
Expand Down Expand Up @@ -254,7 +255,7 @@ if [ $stage -le 16 ]; then
--egs.chunk-right-context-final 0 \
--trainer.num-chunk-per-minibatch 64,32 \
--trainer.frames-per-iter 1500000 \
--trainer.num-epochs 4 \
--trainer.num-epochs $num_epochs \
--trainer.optimization.shrink-value 0.99 \
--trainer.optimization.num-jobs-initial 2 \
--trainer.optimization.num-jobs-final 12 \
Expand Down
344 changes: 344 additions & 0 deletions egs/ami/s5b/local/chain/tuning/run_tdnn_lstm_1l.sh

Large diffs are not rendered by default.

352 changes: 352 additions & 0 deletions egs/ami/s5b/local/chain/tuning/run_tdnn_lstm_1m.sh

Large diffs are not rendered by default.

6 changes: 6 additions & 0 deletions egs/swbd/s5c/RESULTS
Original file line number Diff line number Diff line change
Expand Up @@ -203,6 +203,12 @@ exit 0
%WER 21.2 | 2628 21594 | 81.4 12.8 5.9 2.6 21.2 56.7 | exp/chain/lstm_d_ld5_sp/decode_eval2000_sw1_fsh_fg/score_10_0.0/eval2000_hires.ctm.callhm.filt.sys
%WER 13.88 [ 6829 / 49204, 935 ins, 1690 del, 4204 sub ] exp/chain/lstm_d_ld5_sp/decode_train_dev_sw1_fsh_fg/wer_9_0.0

# current best 'chain' models with TDNN + LSTM + dropout (see local/chain/run_tdnn_lstm_1l.sh)
%WER 13.5 | 4459 42989 | 88.2 8.0 3.8 1.7 13.5 48.2 | exp/chain/tdnn_lstm_1b_dropout_ld5_sp/decode_eval2000_sw1_fsh_fg/score_10_0.0/eval2000_hires.ctm.filt.sys
%WER 8.8 | 1831 21395 | 92.3 5.2 2.5 1.1 8.8 41.9 | exp/chain/tdnn_lstm_1b_dropout_ld5_sp/decode_eval2000_sw1_fsh_fg/score_10_0.0/eval2000_hires.ctm.swbd.filt.sys
%WER 18.1 | 2628 21594 | 84.0 10.8 5.2 2.2 18.1 52.6 | exp/chain/tdnn_lstm_1b_dropout_ld5_sp/decode_eval2000_sw1_fsh_fg/score_10_1.0/eval2000_hires.ctm.callhm.filt.sys
%WER 11.59 [ 5615 / 48460, 708 ins, 1450 del, 3457 sub ] exp/chain/tdnn_lstm_1b_dropout_ld5_sp/decode_train_dev_sw1_fsh_fg/wer_9_0.0

# these are results with nnet3 LSTMs with CTC training : local/ctc/run_lstm.sh
%WER 17.4 | 1831 21395 | 85.3 10.1 4.6 2.7 17.4 57.8 | exp/ctc/lstm_sp/decode_eval2000_sw1_fsh_fg_0.15/score_12_0.0/eval2000_hires.ctm.swbd.filt.sys
%WER 19.4 | 1831 21395 | 83.5 11.2 5.2 3.0 19.4 60.7 | exp/ctc/lstm_sp/decode_eval2000_sw1_tg_0.15/score_12_0.5/eval2000_hires.ctm.swbd.filt.sys
Expand Down
248 changes: 248 additions & 0 deletions egs/swbd/s5c/local/chain/tuning/run_blstm_6l.sh
Original file line number Diff line number Diff line change
@@ -0,0 +1,248 @@
#!/bin/bash

# 6l is same as 6k, but with the per-frame dropout
# location4 as paper : http://www.danielpovey.com/files/2017_interspeech_dropout.pdf

# local/chain/compare_wer_general.sh blstm_6k_sp blstm_6l_sp
# attention: the blatm_6k_sp result here is far better than the updated
# result (14.5 vs 14.1), this may due to noise

# System blstm_6k_sp blstm_6l_sp
# WER on train_dev(tg) 13.30 13.06
# WER on train_dev(fg) 12.34 12.16
# WER on eval2000(tg) 15.5 15.2
# WER on eval2000(fg) 14.1 13.8
# Final train prob -0.052 -0.065
# Final valid prob -0.090 -0.093
# Final train prob (xent) -0.743 -0.831
# Final valid prob (xent) -0.9579 -0.9821

# exp/chain/blstm_6k_sp/: num-iters=327 nj=3..16 num-params=41.2M dim=40+100->6074 combine=-0.069->-0.069 xent:train/valid[217,326,final]=(-0.849,-0.748,-0.743/-1.04,-0.959,-0.958) logprob:train/valid[217,326,final]=(-0.065,-0.053,-0.052/-0.096,-0.090,-0.090)
# exp/chain/blstm_6l_sp/: num-iters=327 nj=3..16 num-params=41.2M dim=40+100->6074 combine=-0.084->-0.082 xent:train/valid[217,326,final]=(-1.45,-0.840,-0.831/-1.58,-0.994,-0.982) logprob:train/valid[217,326,final]=(-0.110,-0.066,-0.065/-0.132,-0.094,-0.093)
set -e

# configs for 'chain'
stage=12
train_stage=-10
get_egs_stage=-10
speed_perturb=true
dir=exp/chain/blstm_6l # Note: _sp will get added to this if $speed_perturb == true.
decode_iter=
decode_dir_affix=

# training options
leftmost_questions_truncate=-1
chunk_width=150
chunk_left_context=40
chunk_right_context=40
xent_regularize=0.025
self_repair_scale=0.00001
label_delay=0
dropout_schedule='0,0@0.20,0.1@0.50,0'

# decode options
extra_left_context=50
extra_right_context=50
frames_per_chunk=

remove_egs=false
common_egs_dir=

affix=
# 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 8" if you have already
# run those things.

suffix=
if [ "$speed_perturb" == "true" ]; then
suffix=_sp
fi

dir=$dir${affix:+_$affix}
if [ $label_delay -gt 0 ]; then dir=${dir}_ld$label_delay; fi
dir=${dir}$suffix
train_set=train_nodup$suffix
ali_dir=exp/tri4_ali_nodup$suffix
treedir=exp/chain/tri5_7d_tree$suffix
lang=data/lang_chain_2y


# if we are using the speed-perturbed data we need to generate
# alignments for it.
local/nnet3/run_ivector_common.sh --stage $stage \
--speed-perturb $speed_perturb \
--generate-alignments $speed_perturb || exit 1;


if [ $stage -le 9 ]; then
# Get the alignments as lattices (gives the CTC training more freedom).
# use the same num-jobs as the alignments
nj=$(cat exp/tri4_ali_nodup$suffix/num_jobs) || exit 1;
steps/align_fmllr_lats.sh --nj $nj --cmd "$train_cmd" data/$train_set \
data/lang exp/tri4 exp/tri4_lats_nodup$suffix
rm exp/tri4_lats_nodup$suffix/fsts.*.gz # save space
fi


if [ $stage -le 10 ]; 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;
# 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

if [ $stage -le 11 ]; then
# Build a tree using our new topology.
steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
--leftmost-questions-truncate $leftmost_questions_truncate \
--context-opts "--context-width=2 --central-position=1" \
--cmd "$train_cmd" 7000 data/$train_set $lang $ali_dir $treedir
fi

if [ $stage -le 12 ]; then
echo "$0: creating neural net configs using the xconfig parser";

num_targets=$(tree-info $treedir/tree |grep num-pdfs|awk '{print $2}')
[ -z $num_targets ] && { echo "$0: error getting num-targets"; exit 1; }
learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)

lstm_opts="decay-time=20 dropout-proportion=0.0"

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

# check steps/libs/nnet3/xconfig/lstm.py for the other options and defaults
fast-lstmp-layer name=blstm1-forward input=lda cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
fast-lstmp-layer name=blstm1-backward input=lda cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=3 $lstm_opts

fast-lstmp-layer name=blstm2-forward input=Append(blstm1-forward, blstm1-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
fast-lstmp-layer name=blstm2-backward input=Append(blstm1-forward, blstm1-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=3 $lstm_opts

fast-lstmp-layer name=blstm3-forward input=Append(blstm2-forward, blstm2-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 $lstm_opts
fast-lstmp-layer name=blstm3-backward input=Append(blstm2-forward, blstm2-backward) cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=3 $lstm_opts

## adding the layers for chain branch
output-layer name=output input=Append(blstm3-forward, blstm3-backward) output-delay=$label_delay include-log-softmax=false dim=$num_targets max-change=1.5

# adding the layers for xent branch
# This block prints the configs for a separate output that will be
# trained with a cross-entropy objective in the 'chain' models... this
# has the effect of regularizing the hidden parts of the model. we use
# 0.5 / args.xent_regularize as the learning rate factor- the factor of
# 0.5 / args.xent_regularize is suitable as it means the xent
# final-layer learns at a rate independent of the regularization
# constant; and the 0.5 was tuned so as to make the relative progress
# similar in the xent and regular final layers.
output-layer name=output-xent input=Append(blstm3-forward, blstm3-backward) output-delay=$label_delay dim=$num_targets learning-rate-factor=$learning_rate_factor 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 13 ]; then
if [[ $(hostname -f) == *.clsp.jhu.edu ]] && [ ! -d $dir/egs/storage ]; then
utils/create_split_dir.pl \
/export/b0{5,6,7,8}/$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 exp/nnet3/ivectors_${train_set} \
--feat.cmvn-opts "--norm-means=false --norm-vars=false" \
--chain.xent-regularize $xent_regularize \
--chain.leaky-hmm-coefficient 0.1 \
--chain.l2-regularize 0.00005 \
--chain.apply-deriv-weights false \
--chain.lm-opts="--num-extra-lm-states=2000" \
--trainer.num-chunk-per-minibatch 64 \
--trainer.frames-per-iter 1200000 \
--trainer.max-param-change 2.0 \
--trainer.num-epochs 4 \
--trainer.optimization.shrink-value 0.99 \
--trainer.optimization.num-jobs-initial 3 \
--trainer.optimization.num-jobs-final 16 \
--trainer.optimization.initial-effective-lrate 0.001 \
--trainer.optimization.final-effective-lrate 0.0001 \
--trainer.optimization.momentum 0.0 \
--trainer.deriv-truncate-margin 8 \
--egs.stage $get_egs_stage \
--egs.opts "--frames-overlap-per-eg 0" \
--egs.chunk-width $chunk_width \
--egs.chunk-left-context $chunk_left_context \
--egs.chunk-right-context $chunk_right_context \
--trainer.dropout-schedule $dropout_schedule \
--egs.dir "$common_egs_dir" \
--cleanup.remove-egs $remove_egs \
--feat-dir data/${train_set}_hires \
--tree-dir $treedir \
--lat-dir exp/tri4_lats_nodup$suffix \
--dir $dir || exit 1;
fi

if [ $stage -le 14 ]; 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 data/lang_sw1_tg $dir $dir/graph_sw1_tg
fi

decode_suff=sw1_tg
graph_dir=$dir/graph_sw1_tg
if [ $stage -le 15 ]; then
[ -z $extra_left_context ] && extra_left_context=$chunk_left_context;
[ -z $extra_right_context ] && extra_right_context=$chunk_right_context;
[ -z $frames_per_chunk ] && frames_per_chunk=$chunk_width;
iter_opts=
if [ ! -z $decode_iter ]; then
iter_opts=" --iter $decode_iter "
fi
for decode_set in train_dev eval2000; do
(
steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
--nj 50 --cmd "$decode_cmd" $iter_opts \
--extra-left-context $extra_left_context \
--extra-right-context $extra_right_context \
--frames-per-chunk "$frames_per_chunk" \
--online-ivector-dir exp/nnet3/ivectors_${decode_set} \
$graph_dir data/${decode_set}_hires \
$dir/decode_${decode_set}${decode_dir_affix:+_$decode_dir_affix}_${decode_suff} || exit 1;
if $has_fisher; then
steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" \
data/lang_sw1_{tg,fsh_fg} data/${decode_set}_hires \
$dir/decode_${decode_set}${decode_dir_affix:+_$decode_dir_affix}_sw1_{tg,fsh_fg} || exit 1;
fi
) &
done
fi
wait;
exit 0;
Loading