Specaug

egs/ami/s5b/conf/mfcc_hires.conf

@@ -3,8 +3,8 @@
 # but MFCC is more easily compressible (because less correlated) which is why 
 # we prefer this method.
 --use-energy=false   # use average of log energy, not energy.
---num-mel-bins=40     # similar to Google's setup.
---num-ceps=40     # there is no dimensionality reduction.
+--num-mel-bins=80     # similar to Google's setup.
+--num-ceps=80     # there is no dimensionality reduction.
 --low-freq=20     # low cutoff frequency for mel bins... this is high-bandwidth data, so
                   # there might be some information at the low end.
 --high-freq=-400 # high cutoff frequently, relative to Nyquist of 8000 (=7600) 

egs/ami/s5b/conf/queue.conf

@@ -0,0 +1,9 @@
+command qsub -v PATH -cwd -S /bin/bash -j y -l arch=*64*
+option mem=* -l mem_free=$0,ram_free=$0
+option mem=0          # Do not add anything to qsub_opts
+option num_threads=* -pe smp $0
+option num_threads=1  # Do not add anything to qsub_opts
+option max_jobs_run=* -tc $0
+default gpu=0
+option gpu=0 -l 'hostname=!a10*&!a18*&!b05*&!b06*'
+option gpu=* -l gpu=$0 -q g.q -l 'hostname=!b05*&!b06*'

egs/ami/s5b/local/ami_mdm_scoring_data_prep.sh

@@ -99,15 +99,19 @@ awk '{print $1}' $tmpdir/segments | \
 join $tmpdir/utt2spk_stm $tmpdir/segments | \
   awk '{ utt=$1; spk=$2; wav=$3; t_beg=$4; t_end=$5;
          if(spk_prev == spk && t_end_prev > t_beg) {
-           print "s/^"utt, wav, t_beg, t_end"$/"utt, wav, t_end_prev, t_end"/;";
+           print utt, wav, t_beg, t_end">"utt, wav, t_end_prev, t_end;
          }
          spk_prev=spk; t_end_prev=t_end;
        }' > $tmpdir/segments_to_fix
 
-if [ -s $tmpdir/segments_to_fix ]; then
+if [ `cat $tmpdir/segments_to_fix | wc -l` -gt 0 ]; then
   echo "$0. Applying following fixes to segments"
   cat $tmpdir/segments_to_fix
-  perl -i -pf $tmpdir/segments_to_fix $tmpdir/segments
+  while read line; do
+     p1=`echo $line | awk -F'>' '{print $1}'`
+     p2=`echo $line | awk -F'>' '{print $2}'`
+     sed -ir "s:$p1:$p2:" $tmpdir/segments
+  done < $tmpdir/segments_to_fix
 fi
 
 # Copy stuff into its final locations [this has been moved from the format_data
@@ -125,4 +129,3 @@ local/convert2stm.pl $dir utt2spk_stm > $dir/stm
 utils/validate_data_dir.sh --no-feats $dir
 
 echo AMI $SET set data preparation succeeded.
-

egs/ami/s5b/local/ami_sdm_scoring_data_prep.sh

@@ -111,21 +111,25 @@ awk '{print $1}' $tmpdir/segments | \
 join $tmpdir/utt2spk_stm $tmpdir/segments | \
   awk '{ utt=$1; spk=$2; wav=$3; t_beg=$4; t_end=$5;
          if(spk_prev == spk && t_end_prev > t_beg) {
-           print "s/^"utt, wav, t_beg, t_end"$/"utt, wav, t_end_prev, t_end"/;";
+           print utt, wav, t_beg, t_end">"utt, wav, t_end_prev, t_end;
          }
          spk_prev=spk; t_end_prev=t_end;
        }' > $tmpdir/segments_to_fix
 
-if [ -s $tmpdir/segments_to_fix ]; then
+if [ `cat $tmpdir/segments_to_fix | wc -l` -gt 0 ]; then
   echo "$0. Applying following fixes to segments"
   cat $tmpdir/segments_to_fix
-  perl -i -pf $tmpdir/segments_to_fix $tmpdir/segments
+  while read line; do
+     p1=`echo $line | awk -F'>' '{print $1}'`
+     p2=`echo $line | awk -F'>' '{print $2}'`
+     sed -ir "s:$p1:$p2:" $tmpdir/segments
+  done < $tmpdir/segments_to_fix
 fi
 
 # Copy stuff into its final locations [this has been moved from the format_data
 # script]
 mkdir -p $dir
-for f in segments_to_fix spk2utt utt2spk utt2spk_stm wav.scp text segments reco2file_and_channel; do
+for f in spk2utt utt2spk utt2spk_stm wav.scp text segments reco2file_and_channel; do
   cp $tmpdir/$f $dir/$f || exit 1;
 done
 
@@ -135,4 +139,3 @@ cp local/english.glm $dir/glm
 utils/validate_data_dir.sh --no-feats $dir
 
 echo AMI $DSET scenario and $SET set data preparation succeeded.
-

egs/ami/s5b/local/chain/run_tdnn.sh

@@ -1 +1 @@
-tuning/run_tdnn_1i.sh
+tuning/run_tdnn_1j.sh

egs/ami/s5b/local/chain/tuning/run_tdnn_1i.sh

@@ -29,14 +29,15 @@ ihm_gmm=tri3  # the gmm for the IHM system (if --use-ihm-ali true).
 num_threads_ubm=32
 ivector_transform_type=pca
 nnet3_affix=_cleaned  # cleanup affix for nnet3 and chain dirs, e.g. _cleaned
-num_epochs=9
+num_epochs=15
 remove_egs=true
-
+decode_iter=
 # 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.
-tdnn_affix=1i  #affix for TDNN directory, e.g. "a" or "b", in case we change the configuration.
+#tdnn_affix=1i_swbd_wide_ep15  #affix for TDNN direory, e.g. "a" or "b", in case we change the configuration.
+tdnn_affix=1j_34M_woaug  #affix for TDNN directory, e.g. "a" or "b", in case we change the configuration.
 common_egs_dir=  # you can set this to use previously dumped egs.
 
 
@@ -56,15 +57,15 @@ where "nvcc" is installed.
 EOF
 fi
 
-local/nnet3/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 $num_threads_ubm \
-                                  --ivector-transform-type "$ivector_transform_type" \
-                                  --nnet3-affix "$nnet3_affix"
+#local/nnet3/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 $num_threads_ubm \
+#                                  --ivector-transform-type "$ivector_transform_type" \
+#                                  --nnet3-affix "$nnet3_affix"
 
 # Note: the first stage of the following script is stage 8.
 local/nnet3/prepare_lores_feats.sh --stage $stage \
@@ -169,45 +170,47 @@ if [ $stage -le 15 ]; then
 
   num_targets=$(tree-info $tree_dir/tree |grep num-pdfs|awk '{print $2}')
   learning_rate_factor=$(echo "print (0.5/$xent_regularize)" | python)
-  opts="l2-regularize=0.02"
-  output_opts="l2-regularize=0.004"
+  affine_opts="l2-regularize=0.01"
+  tdnnf_opts="l2-regularize=0.01 bypass-scale=0.66"
+  linear_opts="l2-regularize=0.01 orthonormal-constraint=-1.0"
+  prefinal_opts="l2-regularize=0.01"
+  output_opts="l2-regularize=0.002"
 
   mkdir -p $dir/configs
   cat <<EOF > $dir/configs/network.xconfig
   input dim=100 name=ivector
-  input dim=40 name=input
-
-  # please note that it is important to have input layer with the name=input
-  # as the layer immediately preceding the fixed-affine-layer to enable
-  # the use of short notation for the descriptor
+  input dim=80 name=input
+
+  # this takes the MFCCs and generates filterbank coefficients.  The MFCCs
+  # are more compressible so we prefer to dump the MFCCs to disk rather
+  # than filterbanks.
+  idct-layer name=idct input=input dim=80 cepstral-lifter=22 affine-transform-file=$dir/configs/idct.mat include-in-init=true
+  batchnorm-component name=batchnorm0 input=idct include-in-init=true
+  spec-augment-layer name=spec-augment freq-max-proportion=0.5 time-zeroed-proportion=0.2 time-mask-max-frames=20 include-in-init=true
   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-layer name=tdnn1 dim=450 $opts
-  relu-batchnorm-layer name=tdnn2 input=Append(-1,0,1) dim=450 $opts
-  relu-batchnorm-layer name=tdnn3 dim=450 $opts
-  relu-batchnorm-layer name=tdnn4 input=Append(-1,0,1) dim=450 $opts
-  relu-batchnorm-layer name=tdnn5 dim=450 $opts
-  relu-batchnorm-layer name=tdnn6 input=Append(-3,0,3) dim=450 $opts
-  relu-batchnorm-layer name=tdnn7 input=Append(-3,0,3) dim=450 $opts
-  relu-batchnorm-layer name=tdnn8 input=Append(-3,0,3) dim=450 $opts
-  relu-batchnorm-layer name=tdnn9 input=Append(-3,0,3) dim=450 $opts
-
-  ## adding the layers for chain branch
-  relu-batchnorm-layer name=prefinal-chain input=tdnn9 dim=450 target-rms=0.5 $opts
-  output-layer name=output include-log-softmax=false dim=$num_targets max-change=1.5 $output_opts
-
-  # 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.
-  relu-batchnorm-layer name=prefinal-xent input=tdnn9 dim=450 target-rms=0.5 $opts
-  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor max-change=1.5 $output_opts
+  relu-batchnorm-layer name=tdnn1 $affine_opts dim=2136
+  tdnnf-layer name=tdnnf2 $tdnnf_opts dim=2136 bottleneck-dim=210 time-stride=1
+  tdnnf-layer name=tdnnf3 $tdnnf_opts dim=2136 bottleneck-dim=210 time-stride=1
+  tdnnf-layer name=tdnnf4 $tdnnf_opts dim=2136 bottleneck-dim=210 time-stride=1
+  tdnnf-layer name=tdnnf5 $tdnnf_opts dim=2136 bottleneck-dim=210 time-stride=0
+  tdnnf-layer name=tdnnf6 $tdnnf_opts dim=2136 bottleneck-dim=210 time-stride=3
+  tdnnf-layer name=tdnnf7 $tdnnf_opts dim=2136 bottleneck-dim=210 time-stride=3
+  tdnnf-layer name=tdnnf8 $tdnnf_opts dim=2136 bottleneck-dim=210 time-stride=3
+  tdnnf-layer name=tdnnf9 $tdnnf_opts dim=2136 bottleneck-dim=210 time-stride=3
+  tdnnf-layer name=tdnnf10 $tdnnf_opts dim=2136 bottleneck-dim=210 time-stride=3
+  tdnnf-layer name=tdnnf11 $tdnnf_opts dim=2136 bottleneck-dim=210 time-stride=3
+  tdnnf-layer name=tdnnf12 $tdnnf_opts dim=2136 bottleneck-dim=210 time-stride=3
+  tdnnf-layer name=tdnnf13 $tdnnf_opts dim=2136 bottleneck-dim=210 time-stride=3
+  tdnnf-layer name=tdnnf14 $tdnnf_opts dim=2136 bottleneck-dim=210 time-stride=3
+  tdnnf-layer name=tdnnf15 $tdnnf_opts dim=2136 bottleneck-dim=210 time-stride=3
+  linear-component name=prefinal-l dim=512 $linear_opts
+
+  prefinal-layer name=prefinal-chain input=prefinal-l $prefinal_opts big-dim=2136 small-dim=512
+  output-layer name=output include-log-softmax=false dim=$num_targets $output_opts
+
+  prefinal-layer name=prefinal-xent input=prefinal-l $prefinal_opts big-dim=2136 small-dim=512
+  output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor $output_opts
 
 EOF
 
@@ -232,7 +235,7 @@ if [ $stage -le 16 ]; then
     --egs.dir "$common_egs_dir" \
     --egs.opts "--frames-overlap-per-eg 0" \
     --egs.chunk-width 150 \
-    --trainer.num-chunk-per-minibatch 128 \
+    --trainer.num-chunk-per-minibatch 64 \
     --trainer.frames-per-iter 1500000 \
     --trainer.num-epochs $num_epochs \
     --trainer.optimization.num-jobs-initial 2 \
@@ -256,15 +259,21 @@ if [ $stage -le 17 ]; then
   utils/mkgraph.sh --self-loop-scale 1.0 data/lang_${LM} $dir $graph_dir
 fi
 
+iter_opts=
+if [ ! -z $decode_iter ]; then
+  iter_opts=" --iter $decode_iter "
+fi
+
 if [ $stage -le 18 ]; then
   rm $dir/.error 2>/dev/null || true
   for decode_set in dev eval; do
       (
       steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-          --nj $nj --cmd "$decode_cmd" \
+          --nj $nj --cmd "$decode_cmd" $iter_opts \
           --online-ivector-dir exp/$mic/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
           --scoring-opts "--min-lmwt 5 " \
-         $graph_dir data/$mic/${decode_set}_hires $dir/decode_${decode_set} || exit 1;
+         $graph_dir data/$mic/${decode_set}_hires \
+         $dir/decode_${decode_set}${decode_iter:+_$decode_iter} || exit 1;
       ) || touch $dir/.error &
   done
   wait
@@ -273,4 +282,5 @@ if [ $stage -le 18 ]; then
     exit 1
   fi
 fi
+
 exit 0