Added PerDimWeightedAverage component for chain models in SWBD, works…

egs/swbd/s5c/local/chain/run_tdnn_4q.sh

@@ -0,0 +1,182 @@
+#!/bin/bash
+
+# this is based on Dan's tdnn_2o script
+# it has a different splicing configuration
+# it uses the PerDimensionWeightedAverage pooling in place of the Jesus layer
+
+# it relies on new steps/nnet3/chain/train_tdnn_b.sh script which accepts more 
+# parameters are calls new config generator steps/nnet3/tdnn/make_configs.py
+# which is more in line with other config generators.
+
+set -e
+
+#%WER 11.1 | 1831 21395 | 90.2 6.3 3.5 1.3 11.1 46.6 | exp/chain/tdnn_v1_trial6_sp/decode_eval2000_sw1_fsh_fg/score_12_0.0/eval2000_hires.ctm.swbd.filt.sys
+#%WER 16.6 | 4459 42989 | 85.2 9.5 5.3 1.8 16.6 53.4 | exp/chain/tdnn_v1_trial6_sp/decode_eval2000_sw1_fsh_fg/score_11_0.0/eval2000_hires.ctm.filt.sys
+#%WER 15.59 [ 7671 / 49204, 883 ins, 2234 del, 4554 sub ] exp/chain/tdnn_v1_trial6_sp/decode_train_dev_sw1_fsh_fg/wer_12_0.0
+
+
+# configs for 'chain'
+affix=
+stage=10
+train_stage=-10
+get_egs_stage=-10
+speed_perturb=true
+dir=exp/chain/tdnn_4q  # Note: _sp will get added to this if $speed_perturb == true.
+decode_iter=
+
+# TDNN options
+# this script uses the new tdnn config generator so it needs a final 0 to reflect that the final layer input has no splicing
+splice_indexes="-2,-1,0,1,2 -1,0,2 -3,0,3 -6,-3,0,3 -6,-3,0,3"
+# smoothing options
+pool_window=7 
+pool_type='per-dim-weighted-average'
+pool_lpfilter_width=
+
+# training options
+num_epochs=4
+initial_effective_lrate=0.001
+final_effective_lrate=0.0001
+leftmost_questions_truncate=-1
+max_param_change=1.0
+final_layer_normalize_target=0.5
+num_jobs_initial=3
+num_jobs_final=16
+minibatch_size=64
+relu_dim=700
+frames_per_eg=150
+remove_egs=false
+common_egs_dir=
+
+
+
+# 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_2o_tree$suffix
+lang=data/lang_chain_2o
+
+
+# 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 [ $stage -le 9 ]; then
+  # Get the alignments as lattices (gives the CTC training more freedom).
+  # use the same num-jobs as the alignments
+  nj=$(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.
+  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
+      --leftmost-questions-truncate $leftmost_questions_truncate \
+      --cmd "$train_cmd" 9000 data/$train_set $lang $ali_dir $treedir
+fi
+
+if [ $stage -le 12 ]; 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
+
+ touch $dir/egs/.nodelete # keep egs around when that run dies.
+
+ steps/nnet3/chain/train_tdnn_b.sh --stage $train_stage \
+    --apply-deriv-weights false \
+    --lm-opts "--num-extra-lm-states=2000" \
+    --get-egs-stage $get_egs_stage \
+    --minibatch-size $minibatch_size \
+    --egs-opts "--frames-overlap-per-eg 0" \
+    --frames-per-eg $frames_per_eg \
+    --num-epochs $num_epochs --num-jobs-initial $num_jobs_initial --num-jobs-final $num_jobs_final \
+    --splice-indexes "$splice_indexes" \
+    --pool-type "$pool_type" \
+    --pool-window "$pool_window" \
+    --pool-lpfilter-width "$pool_lpfilter_width" \
+    --feat-type raw \
+    --online-ivector-dir exp/nnet3/ivectors_${train_set} \
+    --cmvn-opts "--norm-means=false --norm-vars=false" \
+    --initial-effective-lrate $initial_effective_lrate --final-effective-lrate $final_effective_lrate \
+    --max-param-change $max_param_change \
+    --final-layer-normalize-target $final_layer_normalize_target \
+    --relu-dim $relu_dim \
+    --cmd "$decode_cmd" \
+    --remove-egs $remove_egs \
+    --egs-dir "$common_egs_dir" \
+    data/${train_set}_hires $treedir exp/tri4_lats_nodup$suffix $dir  || exit 1;
+fi
+
+if [ $stage -le 13 ]; 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 --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 14 ]; 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_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}_sw1_{tg,fsh_fg} || exit 1;
+      fi
+      ) &
+  done
+fi
+wait;
+exit 0;