Skip to content
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

Jump to bottom

Simplifying AMI recipe #2800

Merged
merged 1 commit into from
Oct 27, 2018
Merged

Simplifying AMI recipe #2800

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
4 changes: 4 additions & 0 deletions egs/ami/s5b/RESULTS_ihm
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -88,6 +88,10 @@

# local/chain/multi_condition/tuning/run_tdnn_lstm_1a.sh --mic ihm
# cleanup + chain TDNN+LSTM model + IHM reverberated data
# Old results:
%WER 19.4 | 13098 94479 | 83.8 10.0 6.1 3.2 19.4 51.8 | -0.168 | exp/ihm/chain_cleaned_rvb/tdnn_lstm1i_sp_rvb_bi/decode_dev/ascore_10/dev_hires.ctm.filt.sys
%WER 19.3 | 12643 89977 | 83.3 11.0 5.7 2.6 19.3 49.6 | -0.046 | exp/ihm/chain_cleaned_rvb/tdnn_lstm1i_sp_rvb_bi/decode_eval/ascore_10/eval_hires.ctm.filt.sys

# New results after simplifying scripts to remove combining short segments etc.:
%WER 19.4 | 12643 89979 | 83.1 10.9 6.0 2.5 19.4 50.7 | 0.010 | exp/ihm/chain_cleaned_rvb/tdnn_lstm1a_sp_rvb_bi/decode_eval/ascore_11/eval_hires.ctm.filt.sys
%WER 19.4 | 13098 94484 | 83.7 10.2 6.1 3.1 19.4 52.0 | -0.119 | exp/ihm/chain_cleaned_rvb/tdnn_lstm1a_sp_rvb_bi/decode_dev/ascore_11/dev_hires.ctm.filt.sys
4 changes: 4 additions & 0 deletions egs/ami/s5b/RESULTS_sdm
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -96,6 +96,10 @@
# 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.
# *** best system ***
# Old results:
%WER 34.0 | 14455 94497 | 69.8 17.7 12.5 3.8 34.0 63.9 | 0.675 | exp/sdm1/chain_cleaned_rvb/tdnn_lstm1i_sp_rvb_bi_ihmali/decode_dev/ascore_10/dev_hires_o4.ctm.filt.sys
%WER 37.5 | 13261 89982 | 65.9 19.3 14.7 3.5 37.5 66.2 | 0.642 | exp/sdm1/chain_cleaned_rvb/tdnn_lstm1i_sp_rvb_bi_ihmali/decode_eval/ascore_10/eval_hires_o4.ctm.filt.sys

# New results after simplifying scripts to remove combining short segments etc.:
%WER 34.6 | 14604 94498 | 69.6 18.8 11.6 4.2 34.6 64.4 | 0.652 | exp/sdm1/chain_cleaned_rvb/tdnn_lstm1a_sp_rvb_bi_ihmali/decode_dev/ascore_11/dev_hires_o4.ctm.filt.sys
%WER 37.6 | 13606 89636 | 66.1 21.0 12.9 3.7 37.6 65.1 | 0.613 | exp/sdm1/chain_cleaned_rvb/tdnn_lstm1a_sp_rvb_bi_ihmali/decode_eval/ascore_11/eval_hires_o4.ctm.filt.sys
87 changes: 41 additions & 46 deletions egs/ami/s5b/local/chain/multi_condition/tuning/run_tdnn_lstm_1a.sh
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,29 +19,28 @@ set -e -o pipefail
stage=0
mic=ihm
nj=30
min_seg_len=1.55
use_ihm_ali=false
train_set=train_cleaned
gmm=tri3_cleaned # the gmm for the target data
ihm_gmm=tri3_cleaned # the gmm for the IHM system (if --use-ihm-ali true).
num_threads_ubm=32
num_data_reps=1

chunk_width=150
chunk_width=160,140,110,80
chunk_left_context=40
chunk_right_context=0
label_delay=5
# The rest are configs specific to this script. Most of the parameters
# are just hardcoded at this level, in the commands below.
train_stage=-10
tree_affix= # affix for tree directory, e.g. "a" or "b", in case we change the configuration.
tlstm_affix=1i #affix for TDNN-LSTM directory, e.g. "a" or "b", in case we change the configuration.
tlstm_affix=1a #affix for TDNN-LSTM directory, e.g. "a" or "b", in case we change the configuration.
common_egs_dir= # you can set this to use previously dumped egs.


# decode options
extra_left_context=50
frames_per_chunk=
frames_per_chunk=160


# End configuration section.
Expand Down Expand Up @@ -75,31 +74,27 @@ rvb_affix=_rvb

if $use_ihm_ali; then
gmm_dir=exp/ihm/${ihm_gmm}
ali_dir=exp/${mic}/${ihm_gmm}_ali_${train_set}_sp_comb_ihmdata
lores_train_data_dir=data/$mic/${train_set}_ihmdata_sp_comb
lores_train_data_dir=data/$mic/${train_set}_ihmdata_sp
tree_dir=exp/$mic/chain${nnet3_affix}/tree_bi${tree_affix}_ihmdata
original_lat_dir=exp/$mic/chain${nnet3_affix}/${ihm_gmm}_${train_set}_sp_comb_lats_ihmdata
lat_dir=exp/$mic/chain${nnet3_affix}${rvb_affix}/${ihm_gmm}_${train_set}_sp${rvb_affix}_comb_lats_ihmdata
original_lat_dir=exp/$mic/chain${nnet3_affix}/${ihm_gmm}_${train_set}_sp_lats_ihmdata
lat_dir=exp/$mic/chain${nnet3_affix}${rvb_affix}/${ihm_gmm}_${train_set}_sp${rvb_affix}_lats_ihmdata
dir=exp/$mic/chain${nnet3_affix}${rvb_affix}/tdnn_lstm${tlstm_affix}_sp${rvb_affix}_bi_ihmali
# note: the distinction between when we use the 'ihmdata' suffix versus
# 'ihmali' is pretty arbitrary.
else
gmm_dir=exp/${mic}/$gmm
ali_dir=exp/${mic}/${gmm}_ali_${train_set}_sp_comb
lores_train_data_dir=data/$mic/${train_set}_sp_comb
lores_train_data_dir=data/$mic/${train_set}_sp
tree_dir=exp/$mic/chain${nnet3_affix}/tree_bi${tree_affix}
original_lat_dir=exp/$mic/chain${nnet3_affix}/${gmm}_${train_set}_sp_comb_lats
lat_dir=exp/$mic/chain${nnet3_affix}${rvb_affix}/${gmm}_${train_set}_sp${rvb_affix}_comb_lats
original_lat_dir=exp/$mic/chain${nnet3_affix}/${gmm}_${train_set}_sp_lats
lat_dir=exp/$mic/chain${nnet3_affix}${rvb_affix}/${gmm}_${train_set}_sp${rvb_affix}_lats
dir=exp/$mic/chain${nnet3_affix}${rvb_affix}/tdnn_lstm${tlstm_affix}_sp${rvb_affix}_bi
fi


local/nnet3/multi_condition/run_ivector_common.sh --stage $stage \
--mic $mic \
--nj $nj \
--min-seg-len $min_seg_len \
--train-set $train_set \
--gmm $gmm \
--num-threads-ubm $num_threads_ubm \
--num-data-reps $num_data_reps \
--nnet3-affix "$nnet3_affix"
Expand All @@ -109,13 +104,13 @@ local/nnet3/multi_condition/run_ivector_common.sh --stage $stage \
local/nnet3/prepare_lores_feats.sh --stage $stage \
--mic $mic \
--nj $nj \
--min-seg-len $min_seg_len \
--min-seg-len "" \
Copy link
Contributor

Choose a reason for hiding this comment

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

wouldn't it be cleaner to remove the min-seg-len option from this script entirely? The run_ivector_common.sh script already doesn't seem to support it.

But I'm concerned that this PR will break the other chain example scripts and the other nnet3 example scripts.

--use-ihm-ali $use_ihm_ali \
--train-set $train_set


train_data_dir=data/$mic/${train_set}_sp${rvb_affix}_hires_comb
train_ivector_dir=exp/$mic/nnet3${nnet3_affix}${rvb_affix}/ivectors_${train_set}_sp${rvb_affix}_hires_comb
train_data_dir=data/$mic/${train_set}_sp${rvb_affix}_hires
train_ivector_dir=exp/$mic/nnet3${nnet3_affix}${rvb_affix}/ivectors_${train_set}_sp${rvb_affix}_hires
final_lm=`cat data/local/lm/final_lm`
LM=$final_lm.pr1-7

Expand All @@ -126,19 +121,6 @@ for f in $gmm_dir/final.mdl $lores_train_data_dir/feats.scp \
done


if [ $stage -le 11 ]; then
if [ -f $ali_dir/ali.1.gz ]; then
echo "$0: alignments in $ali_dir appear to already exist. Please either remove them "
echo " ... or use a later --stage option."
exit 1
fi
echo "$0: aligning perturbed, short-segment-combined ${maybe_ihm}data"
steps/align_fmllr.sh --nj $nj --cmd "$train_cmd" \
${lores_train_data_dir} data/lang $gmm_dir $ali_dir
fi

[ ! -f $ali_dir/ali.1.gz ] && echo "$0: expected $ali_dir/ali.1.gz to exist" && exit 1

if [ $stage -le 12 ]; then
echo "$0: creating lang directory with one state per phone."
# Create a version of the lang/ directory that has one state per phone in the
Expand All @@ -165,28 +147,42 @@ fi
if [ $stage -le 13 ]; then
# Get the alignments as lattices (gives the chain training more freedom).
# use the same num-jobs as the alignments
steps/align_fmllr_lats.sh --nj 100 --cmd "$train_cmd" ${lores_train_data_dir} \
steps/align_fmllr_lats.sh --nj 100 --cmd "$train_cmd" \
--generate-ali-from-lats true ${lores_train_data_dir} \
data/lang $gmm_dir $original_lat_dir
rm $original_lat_dir/fsts.*.gz # save space

lat_dir_ihmdata=exp/ihm/chain${nnet3_affix}/${gmm}_${train_set}_sp_comb_lats
lat_dir_ihmdata=exp/ihm/chain${nnet3_affix}/${gmm}_${train_set}_sp_lats

original_lat_nj=$(cat $original_lat_dir/num_jobs)
ihm_lat_nj=$(cat $lat_dir_ihmdata/num_jobs)

mkdir -p $lat_dir/temp/
mkdir -p $lat_dir/temp2/
lattice-copy "ark:gunzip -c $original_lat_dir/lat.*.gz |" ark,scp:$lat_dir/temp/lats.ark,$lat_dir/temp/lats.scp
lattice-copy "ark:gunzip -c $lat_dir_ihmdata/lat.*.gz |" ark,scp:$lat_dir/temp2/lats.ark,$lat_dir/temp2/lats.scp
$train_cmd --max-jobs-run 10 JOB=1:$original_lat_nj $lat_dir/temp/log/copy_original_lats.JOB.log \
lattice-copy "ark:gunzip -c $original_lat_dir/lat.JOB.gz |" ark,scp:$lat_dir/temp/lats.JOB.ark,$lat_dir/temp/lats.JOB.scp

$train_cmd --max-jobs-run 10 JOB=1:$ihm_lat_nj $lat_dir/temp2/log/copy_ihm_lats.JOB.log \
lattice-copy "ark:gunzip -c $lat_dir_ihmdata/lat.JOB.gz |" ark,scp:$lat_dir/temp2/lats.JOB.ark,$lat_dir/temp2/lats.JOB.scp

for n in $(seq $original_lat_nj); do
cat $lat_dir/temp/lats.$n.scp
done > $lat_dir/temp/combined_lats.scp

# copy the lattices for the reverberated data
rm -f $lat_dir/temp/combined_lats.scp
touch $lat_dir/temp/combined_lats.scp
cat $lat_dir/temp/lats.scp >> $lat_dir/temp/combined_lats.scp
for i in `seq 1 $num_data_reps`; do
cat $lat_dir/temp2/lats.scp | sed -e "s/^/rev${i}_/" >> $lat_dir/temp/combined_lats.scp
done
for n in $(seq $ihm_lat_nj); do
cat $lat_dir/temp2/lats.$n.scp
done | sed -e "s/^/rev${i}_/"
done >> $lat_dir/temp/combined_lats.scp

sort -u $lat_dir/temp/combined_lats.scp > $lat_dir/temp/combined_lats_sorted.scp

lattice-copy scp:$lat_dir/temp/combined_lats_sorted.scp "ark:|gzip -c >$lat_dir/lat.1.gz" || exit 1;
echo "1" > $lat_dir/num_jobs
utils/split_data.sh $train_data_dir $nj

$train_cmd --max-jobs-run 10 JOB=1:$nj $lat_dir/copy_combined_lats.JOB.log \
lattice-copy --include=$train_data_dir/split$nj/JOB/utt2spk \
scp:$lat_dir/temp/combined_lats_sorted.scp \
"ark:|gzip -c >$lat_dir/lat.JOB.gz" || exit 1;

echo $nj > $lat_dir/num_jobs

# copy other files from original lattice dir
for f in cmvn_opts final.mdl splice_opts tree; do
Expand All @@ -206,7 +202,7 @@ if [ $stage -le 14 ]; then
steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
--context-opts "--context-width=2 --central-position=1" \
--leftmost-questions-truncate -1 \
--cmd "$train_cmd" 4200 ${lores_train_data_dir} data/lang_chain $ali_dir $tree_dir
--cmd "$train_cmd" 4200 ${lores_train_data_dir} data/lang_chain $original_lat_dir $tree_dir
fi

xent_regularize=0.1
Expand Down Expand Up @@ -312,7 +308,6 @@ if [ $stage -le 18 ]; then
rm $dir/.error 2>/dev/null || true

[ -z $extra_left_context ] && extra_left_context=$chunk_left_context;
[ -z $frames_per_chunk ] && frames_per_chunk=$chunk_width;

for decode_set in dev eval; do
(
Expand Down
Loading