Added LSTM(P,C) layers. Moved the code to conform to the latest package

structure and many more changes. Successfully generates an LSTM model.

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

@@ -0,0 +1,235 @@
+#!/bin/bash
+
+# 6i is based on run_lstm_6h.sh, but changing the HMM context from triphone to left biphone.
+
+# System                       6h         6i
+# WER on train_dev(tg)      14.73      15.02
+# WER on train_dev(fg)      14.05      14.17
+# WER on eval2000(tg)        17.0       16.8
+# WER on eval2000(fg)        15.8       15.6
+# Final train prob     -0.0955829 -0.0842581
+# Final valid prob       -0.11419  -0.101598
+# Final train prob (xent)      -2.28923  -1.78616
+# Final valid prob (xent)      -2.25641  -1.79062
+# Real-time factor       3.338339   3.131303
+
+set -ex
+
+# configs for 'chain'
+stage=12
+train_stage=-10
+get_egs_stage=-10
+speed_perturb=true
+dir=exp/chain/lstm_6i_xconf  # Note: _sp will get added to this if $speed_perturb == true.
+decode_iter=
+decode_dir_affix=
+
+# training options
+leftmost_questions_truncate=-1
+chunk_width=150
+chunk_left_context=40
+chunk_right_context=0
+xent_regularize=0.025
+self_repair_scale=0.00001
+label_delay=5
+# decode options
+extra_left_context=50
+extra_right_context=0
+frames_per_chunk=
+
+remove_egs=false
+common_egs_dir=
+
+affix=
+# 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}
+if [ $label_delay -gt 0 ]; then dir=${dir}_ld$label_delay; fi
+dir=${dir}$suffix
+train_set=train_nodup$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 [ $stage -le 9 ]; then
+  # Get the alignments as lattices (gives the CTC training more freedom).
+  # use the same num-jobs as the alignments
+  nj=$(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 $leftmost_questions_truncate \
+      --context-opts "--context-width=2 --central-position=1" \
+      --cmd "$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 > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat';
+
+  # check steps/libs/nnet3/xconfig/lstm.py for the other options and defaults
+  lstmp-layer name=lstm1 cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3
+  lstmp-layer name=lstm2 cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3
+  lstmp-layer name=lstm3 cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3
+
+  ## adding the layers for chain branch
+  relu-renorm-layer name=chain-prefinal input=lstm3 target-rms=0.5
+  affine-layer name=chain-final dim=$num_targets bias-stddev=0.0 param-stddev=0.0
+  output-layer name=output input=-$label_delay include-log-softmax=false
+
+  # 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
+  # 1.0 / 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=xent-prefinal input=lstm3 target-rms=0.5
+  affine-layer name=xent-final dim=$num_targets bias-stddev=0.0 param-stddev=0.0 learning-rate-factor=$learning_rate_factor
+  output-layer name=output-xent input=-$label_delay
+
+EOF
+
+  steps/nnet3/xconfig_to_configs.py --xconfig-file $dir/configs/network.xconfig --config-dir $dir/configs/
+  nnet3-init $dir/configs/ref.config  $dir/ref.raw
+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" \
+    --chain.left-deriv-truncate 0 \
+    --trainer.num-chunk-per-minibatch 64 \
+    --trainer.frames-per-iter 1200000 \
+    --trainer.max-param-change 2.0 \
+    --trainer.num-epochs 4 \
+    --trainer.optimization.shrink-value 0.99 \
+    --trainer.optimization.num-jobs-initial 3 \
+    --trainer.optimization.num-jobs-final 16 \
+    --trainer.optimization.initial-effective-lrate 0.001 \
+    --trainer.optimization.final-effective-lrate 0.0001 \
+    --trainer.optimization.momentum 0.0 \
+    --egs.stage $get_egs_stage \
+    --egs.opts "--frames-overlap-per-eg 0" \
+    --egs.chunk-width $chunk_width \
+    --egs.chunk-left-context $chunk_left_context \
+    --egs.chunk-right-context $chunk_right_context \
+    --egs.dir "$common_egs_dir" \
+    --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
+  [ -z $extra_left_context ] && extra_left_context=$chunk_left_context;
+  [ -z $extra_right_context ] && extra_right_context=$chunk_right_context;
+  [ -z $frames_per_chunk ] && frames_per_chunk=$chunk_width;
+  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 \
+          --extra-left-context $extra_left_context  \
+          --extra-right-context $extra_right_context  \
+          --frames-per-chunk "$frames_per_chunk" \
+          --online-ivector-dir exp/nnet3/ivectors_${decode_set} \
+         $graph_dir data/${decode_set}_hires \
+         $dir/decode_${decode_set}${decode_dir_affix:+_$decode_dir_affix}_${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}${decode_dir_affix:+_$decode_dir_affix}_sw1_{tg,fsh_fg} || exit 1;
+      fi
+      ) &
+  done
+fi
+wait;
+exit 0;

egs/wsj/s5/steps/libs/__init__.py

@@ -0,0 +1,9 @@
+
+
+# Copyright 2016    Vimal Manohar
+# Apache 2.0.
+
+""" This package contains modules and subpackages used in kaldi scripts.
+"""
+
+__all__ = ["common"]

egs/wsj/s5/steps/libs/nnet3/__init__.py

@@ -0,0 +1,12 @@
+
+# Copyright 2016    Johns Hopkins University (Dan Povey)
+#           2016    Vimal Manohar
+#           2016    Vijayaditya Peddinti
+#           2016    Yiming Wang
+# Apache 2.0.
+
+
+# This module has the python functions which facilitate the use of nnet3 toolkit
+# It has two sub-modules
+# xconfig : Library for parsing high level description of neural networks
+# train : Library for training scripts

egs/wsj/s5/steps/libs/nnet3/xconfig/__init__.py

@@ -0,0 +1,30 @@
+# This library has classes and methods to form neural network computation graphs,
+# in the nnet3 framework, using higher level abstractions called 'layers'
+# (e.g. sub-graphs like LSTMS )
+# Note : We use the term 'layer' though the computation graph can have a
+# highly non-linear structure as, other terms such as nodes/components have
+# already been used in C++ codebase of nnet3.
+
+# This is basically a config parser module, where the configs have very concise
+# descriptions of a neural network.
+
+# This module has methods to convert the xconfigs into a configs interpretable
+# by nnet3 C++ library. It generates three different configs:
+#
+# init.config : which is the config with the info necessary for computing
+#               the preconditioning matrix i.e., LDA transform
+#               e.g.
+#                 input-node name=input dim=40
+#                 input-node name=ivector dim=100
+#                 output-node name=output input=Append(Offset(input, -2), Offset(input, -1), input, Offset(input, 1), Offset(input, 2), ReplaceIndex(ivector, t, 0)) objective=linear
+
+
+# 'ref.config' : which is a version of the config file used to generate
+#                a model for getting left and right context it doesn't read anything for the
+#                LDA-like transform and/or presoftmax-prior-scale components)
+# 'final.config' : which has the actual config used to initialize the model used
+#                 in training i.e, it has file paths for LDA transform and
+#                 other initialization files
+
+
+__all__ = ["utils", "layers", "parser"]