modifications for adding resnet

egs/iam/s5/local/chain/run_cnn_1a.sh

@@ -29,8 +29,8 @@ alignment_subsampling_factor=1
 chunk_width=340,300,200,100
 num_leaves=500
 # we don't need extra left/right context for TDNN systems.
-chunk_left_context=32
-chunk_right_context=32
+chunk_left_context=0
+chunk_right_context=0
 tdnn_dim=450
 # training options
 srand=0
@@ -127,8 +127,8 @@ if [ $stage -le 4 ]; then
 
   num_targets=$(tree-info $tree_dir/tree | grep num-pdfs | awk '{print $2}')
   learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-  common1="required-time-offsets=0 height-offsets=-2,-1,0,1,2 num-filters-out=36"
-  common2="required-time-offsets=0 height-offsets=-2,-1,0,1,2 num-filters-out=70"
+  common1="height-offsets=-2,-1,0,1,2 num-filters-out=36"
+  common2="height-offsets=-2,-1,0,1,2 num-filters-out=70"
   mkdir -p $dir/configs
   cat <<EOF > $dir/configs/network.xconfig
   input dim=40 name=input
@@ -223,8 +223,8 @@ if [ $stage -le 7 ]; then
   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 32 \
-    --extra-right-context-final 32 \
+    --extra-left-context-initial 0 \
+    --extra-right-context-final 0 \
     --frames-per-chunk $frames_per_chunk \
     --nj $nj --cmd "$decode_cmd" \
     $dir/graph data/test $dir/decode_test || exit 1;

egs/iam/s5/local/chain/run_cnn_chainali_1a.sh

@@ -33,8 +33,8 @@ alignment_subsampling_factor=1
 chunk_width=340,300,200,100
 num_leaves=500
 # we don't need extra left/right context for TDNN systems.
-chunk_left_context=32
-chunk_right_context=32
+chunk_left_context=0
+chunk_right_context=0
 tdnn_dim=450
 # training options
 srand=0
@@ -131,9 +131,9 @@ if [ $stage -le 4 ]; then
 
   num_targets=$(tree-info $tree_dir/tree | grep num-pdfs | awk '{print $2}')
   learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
-  common1="required-time-offsets=0 height-offsets=-2,-1,0,1,2 num-filters-out=36"
-  common2="required-time-offsets=0 height-offsets=-2,-1,0,1,2 num-filters-out=70"
-  common3="required-time-offsets=0 height-offsets=-1,0,1 num-filters-out=70"
+  common1="height-offsets=-2,-1,0,1,2 num-filters-out=36"
+  common2="height-offsets=-2,-1,0,1,2 num-filters-out=70"
+  common3="height-offsets=-1,0,1 num-filters-out=70"
   mkdir -p $dir/configs
   cat <<EOF > $dir/configs/network.xconfig
   input dim=40 name=input
@@ -228,8 +228,8 @@ if [ $stage -le 7 ]; then
   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 32 \
-    --extra-right-context-final 32 \
+    --extra-left-context-initial 0 \
+    --extra-right-context-final 0 \
     --frames-per-chunk $frames_per_chunk \
     --nj $nj --cmd "$decode_cmd" \
     $dir/graph data/test $dir/decode_test || exit 1;

egs/iam/s5/local/chain/run_cnn_chainali_1b.sh

@@ -1,13 +1,10 @@
 #!/bin/bash
 
-# chainali_1a uses chain model for lattice instead of gmm-hmm model. It has more cnn layers as compared to 1a
-# (18.34% -> 13.68%)
+# chainali_1b uses chain model for lattice instead of gmm-hmm model. It has more cnn layers as compared to 1a
+# (15.17% -> 14.54%)
 
-# steps/info/chain_dir_info.pl exp/chain/cnn1a_chainali/
-# exp/chain/cnn_chainali_1a/: num-iters=21 nj=2..4 num-params=3.8M dim=40->380 combine=-0.009->-0.006 xent:train/valid[13,20,final]=(-0.870,-0.593,-0.568/-1.08,-0.889,-0.874) logprob:train/valid[13,20,final]=(-0.035,-0.003,-0.001/-0.077,-0.055,-0.054)
-
-# head exp/chain/cnn_chainali_1a/decode_test/scoring_kaldi/best_wer
-# %WER 13.68 [ 2410 / 17616, 243 ins, 633 del, 1534 sub ] exp/chain/cnn_chainali_1a/decode_test/wer_8_1.0
+# steps/info/chain_dir_info.pl exp/chain/chainali_cnn_1b/
+# exp/chain/chainali_cnn_1b/: num-iters=21 nj=2..4 num-params=4.0M dim=40->364 combine=-0.009->-0.006 xent:train/valid[13,20,final]=(-0.870,-0.593,-0.568/-1.08,-0.889,-0.874) logprob:train/valid[13,20,final]=(-0.035,-0.003,-0.001/-0.077,-0.055,-0.054)
 
 set -e -o pipefail
 
@@ -18,7 +15,7 @@ train_set=train
 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.
-affix=_1a  #affix for TDNN+LSTM directory e.g. "1a" or "1b", in case we change the configuration.
+affix=_1b  #affix for TDNN+LSTM directory e.g. "1a" or "1b", in case we change the configuration.
 ali=tri3_ali
 chain_model_dir=exp/chain${nnet3_affix}/cnn${affix}
 common_egs_dir=

egs/iam/s5/local/chain/run_cnn_chainali_1c.sh

@@ -1,13 +1,10 @@
 #!/bin/bash
 
 # chainali_1a uses chain model for lattice instead of gmm-hmm model. It has more cnn layers as compared to 1a
-# (18.34% -> 13.68%)
+# (15.17% -> 14.51%)
 
-# steps/info/chain_dir_info.pl exp/chain/cnn1a_chainali/
-# exp/chain/cnn_chainali_1a/: num-iters=21 nj=2..4 num-params=3.8M dim=40->380 combine=-0.009->-0.006 xent:train/valid[13,20,final]=(-0.870,-0.593,-0.568/-1.08,-0.889,-0.874) logprob:train/valid[13,20,final]=(-0.035,-0.003,-0.001/-0.077,-0.055,-0.054)
-
-# head exp/chain/cnn_chainali_1a/decode_test/scoring_kaldi/best_wer
-# %WER 13.68 [ 2410 / 17616, 243 ins, 633 del, 1534 sub ] exp/chain/cnn_chainali_1a/decode_test/wer_8_1.0
+# steps/info/chain_dir_info.pl exp/chain/chainali_cnn_1c/
+# exp/chain/chainali_cnn_1c/: num-iters=21 nj=2..4 num-params=3.9M dim=40->364 combine=-0.009->-0.006 xent:train/valid[13,20,final]=(-0.870,-0.593,-0.568/-1.08,-0.889,-0.874) logprob:train/valid[13,20,final]=(-0.035,-0.003,-0.001/-0.077,-0.055,-0.054)
 
 set -e -o pipefail
 
@@ -18,7 +15,7 @@ train_set=train
 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.
-affix=_1a  #affix for TDNN+LSTM directory e.g. "1a" or "1b", in case we change the configuration.
+affix=_1c  #affix for TDNN+LSTM directory e.g. "1a" or "1b", in case we change the configuration.
 ali=tri3_ali
 chain_model_dir=exp/chain${nnet3_affix}/cnn${affix}
 common_egs_dir=

egs/iam/s5/local/chain/run_cnn_chainali_1d.sh

@@ -1,14 +1,10 @@
 #!/bin/bash
 
 # chainali_1a uses chain model for lattice instead of gmm-hmm model. It has more cnn layers as compared to 1a
-# (18.34% -> 13.68%)
-
-# steps/info/chain_dir_info.pl exp/chain/cnn1a_chainali/
-# exp/chain/cnn_chainali_1a/: num-iters=21 nj=2..4 num-params=3.8M dim=40->380 combine=-0.009->-0.006 xent:train/valid[13,20,final]=(-0.870,-0.593,-0.568/-1.08,-0.889,-0.874) logprob:train/valid[13,20,final]=(-0.035,-0.003,-0.001/-0.077,-0.055,-0.054)
-
-# head exp/chain/cnn_chainali_1a/decode_test/scoring_kaldi/best_wer
-# %WER 13.68 [ 2410 / 17616, 243 ins, 633 del, 1534 sub ] exp/chain/cnn_chainali_1a/decode_test/wer_8_1.0
+# (19.10% -> 14.51%)
 
+# steps/info/chain_dir_info.pl exp/chain/chainali_cnn_1d/
+# exp/chain/chainali_cnn_1d/: num-iters=21 nj=2..4 num-params=3.6M dim=40->364 combine=-0.009->-0.006 xent:train/valid[13,20,final]=(-0.870,-0.593,-0.568/-1.08,-0.889,-0.874) logprob:train/valid[13,20,final]=(-0.035,-0.003,-0.001/-0.077,-0.055,-0.054)
 set -e -o pipefail
 
 stage=0
@@ -18,7 +14,7 @@ train_set=train
 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.
-affix=_1a  #affix for TDNN+LSTM directory e.g. "1a" or "1b", in case we change the configuration.
+affix=_1d  #affix for TDNN+LSTM directory e.g. "1a" or "1b", in case we change the configuration.
 ali=tri3_ali
 chain_model_dir=exp/chain${nnet3_affix}/cnn${affix}
 common_egs_dir=