madcat Arabic (optional extra corpus text for LM), IAM (making word-based and BPE based setup similar), madcat Chinese (clean scripts, tuning)

egs/iam/v1/local/augment_data.sh

@@ -0,0 +1,34 @@
+#!/bin/bash
+# Copyright   2018 Hossein Hadian
+#             2018 Ashish Arora
+
+# Apache 2.0
+# This script performs data augmentation.
+
+nj=4
+cmd=run.pl
+feat_dim=40
+echo "$0 $@"
+
+. ./cmd.sh
+. ./path.sh
+. ./utils/parse_options.sh || exit 1;
+
+srcdir=$1
+outdir=$2
+datadir=$3
+aug_set=aug1
+mkdir -p $datadir/augmentations
+echo "copying $srcdir to $datadir/augmentations/$aug_set, allowed length, creating feats.scp"
+
+for set in $aug_set; do
+  image/copy_data_dir.sh --spk-prefix $set- --utt-prefix $set- \
+    $srcdir $datadir/augmentations/$set
+  cat $srcdir/allowed_lengths.txt > $datadir/augmentations/$set/allowed_lengths.txt
+  local/extract_features.sh --nj $nj --cmd "$cmd" --feat-dim $feat_dim \
+    --fliplr false --augment true $datadir/augmentations/$set
+done
+
+echo " combine original data and data from different augmentations"
+utils/combine_data.sh --extra-files images.scp $outdir $srcdir $datadir/augmentations/$aug_set
+cat $srcdir/allowed_lengths.txt > $outdir/allowed_lengths.txt

egs/iam/v1/local/chain/compare_wer.sh

@@ -34,6 +34,20 @@ for x in $*; do
 done
 echo
 
+echo -n "# WER val                    "
+for x in $*; do
+  wer=$(cat $x/decode_val/scoring_kaldi/best_wer | awk '{print $2}')
+  printf "% 10s" $wer
+done
+echo
+
+echo -n "# CER val                    "
+for x in $*; do
+  cer=$(cat $x/decode_val/scoring_kaldi/best_cer | awk '{print $2}')
+  printf "% 10s" $cer
+done
+echo
+
 if $used_epochs; then
   exit 0;  # the diagnostics aren't comparable between regular and discriminatively trained systems.
 fi

egs/iam/v1/local/chain/run_cnn.sh

@@ -0,0 +1 @@
+tuning/run_cnn_1a.sh

egs/iam/v1/local/chain/run_cnn_chainali.sh

@@ -0,0 +1 @@
+tuning/run_cnn_chainali_1d.sh

egs/iam/v1/local/chain/run_cnn_e2eali.sh

@@ -0,0 +1 @@
+tuning/run_cnn_e2eali_1c.sh

egs/iam/v1/local/chain/run_e2e_cnn.sh

@@ -0,0 +1 @@
+tuning/run_e2e_cnn_1a.sh

egs/iam/v1/local/chain/run_cnn_chainali_1a.sh → egs/iam/v1/local/chain/tuning/run_cnn_chainali_1a.sh

@@ -8,6 +8,7 @@ stage=0
 
 nj=30
 train_set=train
+decode_val=true
 gmm=tri3        # this is the source gmm-dir that we'll use for alignments; it
                 # should have alignments for the specified training data.
 nnet3_affix=    # affix for exp dirs, e.g. it was _cleaned in tedlium.
@@ -32,7 +33,8 @@ tdnn_dim=450
 # training options
 srand=0
 remove_egs=false
-lang_test=lang_test
+lang_test=lang_unk
+if $decode_val; then maybe_val=val; else maybe_val= ; fi
 # End configuration section.
 echo "$0 $@"  # Print the command line for logging
 
@@ -219,12 +221,13 @@ fi
 
 if [ $stage -le 7 ]; then
   frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
-  steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-    --extra-left-context $chunk_left_context \
-    --extra-right-context $chunk_right_context \
-    --extra-left-context-initial 0 \
-    --extra-right-context-final 0 \
-    --frames-per-chunk $frames_per_chunk \
-    --nj $nj --cmd "$cmd" \
-    $dir/graph data/test $dir/decode_test || exit 1;
+  for decode_set in test $maybe_val; do
+    steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+      --frames-per-chunk $frames_per_chunk \
+      --nj $nj --cmd "$cmd" \
+      $dir/graph data/$decode_set $dir/decode_$decode_set || exit 1;
+  done
 fi
+
+echo "Done. Date: $(date). Results:"
+local/chain/compare_wer.sh $dir

egs/iam/v1/local/chain/run_cnn_chainali_1b.sh → egs/iam/v1/local/chain/tuning/run_cnn_chainali_1b.sh

@@ -22,6 +22,7 @@ stage=0
 
 nj=30
 train_set=train
+decode_val=true
 gmm=tri3        # this is the source gmm-dir that we'll use for alignments; it
                 # should have alignments for the specified training data.
 nnet3_affix=    # affix for exp dirs, e.g. it was _cleaned in tedlium.
@@ -47,6 +48,7 @@ tdnn_dim=450
 srand=0
 remove_egs=false
 lang_test=lang_unk
+if $decode_val; then maybe_val=val; else maybe_val= ; fi
 # End configuration section.
 echo "$0 $@"  # Print the command line for logging
 
@@ -235,12 +237,13 @@ fi
 
 if [ $stage -le 7 ]; then
   frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
-  steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-    --extra-left-context $chunk_left_context \
-    --extra-right-context $chunk_right_context \
-    --extra-left-context-initial 0 \
-    --extra-right-context-final 0 \
-    --frames-per-chunk $frames_per_chunk \
-    --nj $nj --cmd "$cmd" \
-    $dir/graph data/test $dir/decode_test || exit 1;
+  for decode_set in test $maybe_val; do
+    steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+      --frames-per-chunk $frames_per_chunk \
+      --nj $nj --cmd "$cmd" \
+      $dir/graph data/$decode_set $dir/decode_$decode_set || exit 1;
+  done
 fi
+
+echo "Done. Date: $(date). Results:"
+local/chain/compare_wer.sh $dir

egs/iam/v1/local/chain/run_cnn_chainali_1c.sh → egs/iam/v1/local/chain/tuning/run_cnn_chainali_1c.sh

@@ -20,6 +20,7 @@ stage=0
 
 nj=30
 train_set=train
+decode_val=true
 gmm=tri3        # this is the source gmm-dir that we'll use for alignments; it
                 # should have alignments for the specified training data.
 nnet3_affix=    # affix for exp dirs, e.g. it was _cleaned in tedlium.
@@ -44,6 +45,7 @@ tdnn_dim=450
 srand=0
 remove_egs=false
 lang_test=lang_unk
+if $decode_val; then maybe_val=val; else maybe_val= ; fi
 # End configuration section.
 echo "$0 $@"  # Print the command line for logging
 
@@ -236,12 +238,13 @@ fi
 
 if [ $stage -le 7 ]; then
   frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
-  steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-    --extra-left-context $chunk_left_context \
-    --extra-right-context $chunk_right_context \
-    --extra-left-context-initial 0 \
-    --extra-right-context-final 0 \
-    --frames-per-chunk $frames_per_chunk \
-    --nj $nj --cmd "$cmd" \
-    $dir/graph data/test $dir/decode_test || exit 1;
+  for decode_set in test $maybe_val; do
+    steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+      --frames-per-chunk $frames_per_chunk \
+      --nj $nj --cmd "$cmd" \
+      $dir/graph data/$decode_set $dir/decode_$decode_set || exit 1;
+  done
 fi
+
+echo "Done. Date: $(date). Results:"
+local/chain/compare_wer.sh $dir

egs/iam/v1/local/chain/run_cnn_e2eali_1b.sh → egs/iam/v1/local/chain/tuning/run_cnn_e2eali_1b.sh

@@ -1,22 +1,20 @@
 #!/bin/bash
 
 # e2eali_1b is the same as e2eali_1a but uses unconstrained egs
-
-# local/chain/compare_wer.sh /home/hhadian/kaldi-rnnlm/egs/iam/v1/exp/chain/cnn_e2eali_1a exp/chain/cnn_e2eali_1b
-# System                      cnn_e2eali_1a cnn_e2eali_1b
-# WER                             12.79     12.23
-# CER                              5.73      5.48
-# Final train prob              -0.0556   -0.0367
-# Final valid prob              -0.0795   -0.0592
-# Final train prob (xent)       -0.9178   -0.8382
-# Final valid prob (xent)       -1.0604   -0.9853
-# Parameters                      3.95M     3.95M
+# local/chain/compare_wer.sh exp/chain/cnn_e2eali_1b
+# System                      cnn_e2eali_1b (dict_50k) cnn_e2eali_1b (dict_500k)
+# WER                             11.41                   10.25
+# CER                              4.87                    4.60
+# Final train prob              -0.0384                 -0.0384
+# Final valid prob              -0.0444                 -0.0444
+# Final train prob (xent)       -0.8084                 -0.8084
+# Final valid prob (xent)       -0.8470                 -0.8470
+# Parameters                      3.97M                   3.97M
 
 # steps/info/chain_dir_info.pl exp/chain/cnn_e2eali_1b
-# exp/chain/cnn_e2eali_1b: num-iters=21 nj=2..4 num-params=4.0M dim=40->360 combine=-0.038->-0.038 (over 1) xent:train/valid[13,20,final]=(-1.34,-0.967,-0.838/-1.40,-1.07,-0.985) logprob:train/valid[13,20,final]=(-0.075,-0.054,-0.037/-0.083,-0.072,-0.059)
+# exp/chain/cnn_e2eali_1b: num-iters=42 nj=2..4 num-params=4.0M dim=40->376 combine=-0.039->-0.039 (over 1) xent:train/valid[27,41,final]=(-1.28,-0.846,-0.808/-1.27,-0.871,-0.847) logprob:train/valid[27,41,final]=(-0.064,-0.043,-0.038/-0.065,-0.051,-0.044)
 
 set -e -o pipefail
-
 stage=0
 
 nj=30
@@ -141,7 +139,6 @@ if [ $stage -le 4 ]; then
   mkdir -p $dir/configs
   cat <<EOF > $dir/configs/network.xconfig
   input dim=40 name=input
-
   conv-relu-batchnorm-layer name=cnn1 height-in=40 height-out=40 time-offsets=-3,-2,-1,0,1,2,3 $common1
   conv-relu-batchnorm-layer name=cnn2 height-in=40 height-out=20 time-offsets=-2,-1,0,1,2 $common1 height-subsample-out=2
   conv-relu-batchnorm-layer name=cnn3 height-in=20 height-out=20 time-offsets=-4,-2,0,2,4 $common2
@@ -152,11 +149,9 @@ if [ $stage -le 4 ]; then
   relu-batchnorm-layer name=tdnn1 input=Append(-4,-2,0,2,4) dim=$tdnn_dim $tdnn_opts
   relu-batchnorm-layer name=tdnn2 input=Append(-4,0,4) dim=$tdnn_dim $tdnn_opts
   relu-batchnorm-layer name=tdnn3 input=Append(-4,0,4) dim=$tdnn_dim $tdnn_opts
-
   ## adding the layers for chain branch
   relu-batchnorm-layer name=prefinal-chain dim=$tdnn_dim target-rms=0.5 $tdnn_opts
   output-layer name=output include-log-softmax=false dim=$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' mod?els... this