Added the basic TDNN+LSTM scripts

egs/swbd/s5c/local/chain/run_tdnn_lstm.sh

@@ -0,0 +1 @@
+tuning/run_tdnn_lstm_1a.sh

egs/swbd/s5c/local/chain/tuning/run_blstm_6j.sh

@@ -6,6 +6,16 @@
 # and bias-stddev=0 initialization.
 # This run also accounts for changes in training due to the BackpropTruncationComponent
 
+#System                 blstm_6i  blstm_6j
+#WER on train_dev(tg)      14.11     13.80
+#WER on train_dev(fg)      13.04     12.64
+#WER on eval2000(tg)        16.2      15.6
+#WER on eval2000(fg)        14.6      14.2
+#Final train prob     -0.0615713-0.0552637
+#Final valid prob     -0.0829338-0.0765151
+#Final train prob (xent)      -1.16518 -0.777318
+#Final valid prob (xent)      -1.26028 -0.912595
+
 set -e
 
 # configs for 'chain'

egs/swbd/s5c/local/chain/tuning/run_lstm_6j.sh

@@ -3,23 +3,21 @@
 # 6j is same as 6i but using the xconfig format of network specification.
 # Also, the model is trained without layer-wise discriminative pretraining.
 # Another minor change is that the final-affine component has param-stddev-0
-# and bias-stddev=0 initialization.
+# and bias-stddev=0 initialization. The results also account for changes
+# due to BackpropTruncationComponent in place of ClipGradientComponent.
+# Note that removal of layerwise discriminative pretraining does not result
+# in a lot of improvement in LSTMs, compared to TDNNs (7f vs 7g).
 
-
-
-# This run is affected by the bug that per-element-scale components do not have
-# max-change. The updated results without the bug will be submitted soon.
 #System               lstm_6i_ld5  lstm_6j_ld5
-#WER on train_dev(tg)      14.65     14.43
-#WER on train_dev(fg)      13.38     13.17
-#WER on eval2000(tg)        16.9      16.9
-#WER on eval2000(fg)        15.4      15.3
-#Final train prob     -0.0751668-0.0795697
-#Final valid prob     -0.0928206-0.0926466
-#Final train prob (xent)      -1.34549  -1.16067
-#Final valid prob (xent)      -1.41301  -1.23679
+#WER on train_dev(tg)      14.65     14.66
+#WER on train_dev(fg)      13.38     13.42
+#WER on eval2000(tg)        16.9      16.8
+#WER on eval2000(fg)        15.4      15.4
+#Final train prob     -0.0751668-0.0824531
+#Final valid prob     -0.0928206-0.0989325
+#Final train prob (xent)      -1.34549  -1.15506
+#Final valid prob (xent)      -1.41301  -1.24364
 #
-
 set -e
 
 # configs for 'chain'

egs/swbd/s5c/local/chain/tuning/run_tdnn_7i.sh

@@ -0,0 +1,221 @@
+#!/bin/bash
+
+# Same as 7h but double the number of parameters (27983950 vs 15551509)
+
+set -e
+
+
+#System                  tdnn_7h   tdnn_7i
+#WER on train_dev(tg)      13.84     13.48
+#WER on train_dev(fg)      12.84     12.47
+#WER on eval2000(tg)        16.5      16.4
+#WER on eval2000(fg)        14.8      14.9
+#Final train prob     -0.0889771-0.0785415
+#Final valid prob      -0.113102 -0.105757
+#Final train prob (xent)       -1.2533  -1.15785
+#Final valid prob (xent)      -1.36743  -1.28397
+#
+# configs for 'chain'
+affix=
+stage=12
+train_stage=0
+get_egs_stage=-10
+speed_perturb=true
+dir=exp/chain/tdnn_7i  # Note: _sp will get added to this if $speed_perturb == true.
+decode_iter=
+
+# training options
+num_epochs=4
+initial_effective_lrate=0.001
+final_effective_lrate=0.0001
+leftmost_questions_truncate=-1
+max_param_change=2.0
+final_layer_normalize_target=0.5
+num_jobs_initial=3
+num_jobs_final=16
+minibatch_size=128
+frames_per_eg=150
+remove_egs=false
+common_egs_dir=exp/chain/tdnn_7g_sp/egs
+xent_regularize=0.1
+
+# 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:+_$affix}$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 [ $stage -le 9 ]; then
+  # Get the alignments as lattices (gives the LF-MMI training more freedom).
+  # use the same num-jobs as the alignments
+  nj=$(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. This is the critically different
+  # step compared with other recipes.
+  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 exp/chain/tri5_7d_tree_sp/tree |grep num-pdfs|awk '{print $2}')
+  learning_rate_factor=$(echo "print 0.5/$xent_regularize" | python)
+
+  mkdir -p $dir/configs
+  cat <<EOF > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda input=Append(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  # the first splicing is moved before the lda layer, so no splicing here
+  relu-renorm-layer name=tdnn1 dim=1024
+  relu-renorm-layer name=tdnn2 input=Append(-1,0,1) dim=1024
+  relu-renorm-layer name=tdnn3 input=Append(-1,0,1) dim=1024
+  relu-renorm-layer name=tdnn4 input=Append(-3,0,3) dim=1024
+  relu-renorm-layer name=tdnn5 input=Append(-3,0,3) dim=1024
+  relu-renorm-layer name=tdnn6 input=Append(-3,0,3) dim=1024
+  relu-renorm-layer name=tdnn7 input=Append(-3,0,3) dim=1024
+
+  ## adding the layers for chain branch
+  relu-renorm-layer name=prefinal-chain input=tdnn7 dim=625 target-rms=0.5
+  output-layer name=output include-log-softmax=false dim=$num_targets max-change=1.5
+
+  # adding the layers for xent branch
+  # This block prints the configs for a separate output that will be
+  # trained with a cross-entropy objective in the 'chain' models... this
+  # has the effect of regularizing the hidden parts of the model.  we use
+  # 0.5 / args.xent_regularize as the learning rate factor- the factor of
+  # 0.5 / args.xent_regularize is suitable as it means the xent
+  # final-layer learns at a rate independent of the regularization
+  # constant; and the 0.5 was tuned so as to make the relative progress
+  # similar in the xent and regular final layers.
+  relu-renorm-layer name=prefinal-xent input=tdnn7 dim=625 target-rms=0.5
+  output-layer name=output-xent dim=$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 $xent_regularize \
+    --chain.leaky-hmm-coefficient 0.1 \
+    --chain.l2-regularize 0.00005 \
+    --chain.apply-deriv-weights false \
+    --chain.lm-opts="--num-extra-lm-states=2000" \
+    --egs.dir "$common_egs_dir" \
+    --egs.stage $get_egs_stage \
+    --egs.opts "--frames-overlap-per-eg 0" \
+    --egs.chunk-width $frames_per_eg \
+    --trainer.num-chunk-per-minibatch $minibatch_size \
+    --trainer.frames-per-iter 1500000 \
+    --trainer.num-epochs $num_epochs \
+    --trainer.optimization.num-jobs-initial $num_jobs_initial \
+    --trainer.optimization.num-jobs-final $num_jobs_final \
+    --trainer.optimization.initial-effective-lrate $initial_effective_lrate \
+    --trainer.optimization.final-effective-lrate $final_effective_lrate \
+    --trainer.max-param-change $max_param_change \
+    --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 --left-biphone --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
+  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" $iter_opts \
+          --online-ivector-dir exp/nnet3/ivectors_${decode_set} \
+          $graph_dir data/${decode_set}_hires $dir/decode_${decode_set}${decode_iter:+_$decode_iter}_${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}${decode_iter:+_$decode_iter}_sw1_{tg,fsh_fg} || exit 1;
+      fi
+      ) &
+  done
+fi
+wait;
+exit 0;