From 7109c43e9673c9c58b8ee2f29e3edd5d2ed52891 Mon Sep 17 00:00:00 2001
From: Gaofeng Cheng <770579626@qq.com>
Date: Mon, 12 Dec 2016 21:06:05 +0800
Subject: [PATCH 1/2] change dropout_parser strategy

---
 .../local/chain/tuning/run_tdnn_lstm_1i_dp.sh | 293 ++++++++++++++++++
 egs/wsj/s5/steps/libs/nnet3/train/common.py   |  16 +-
 egs/wsj/s5/steps/libs/nnet3/xconfig/lstm.py   |  25 +-
 3 files changed, 325 insertions(+), 9 deletions(-)
 create mode 100644 egs/ami/s5b/local/chain/tuning/run_tdnn_lstm_1i_dp.sh

diff --git a/egs/ami/s5b/local/chain/tuning/run_tdnn_lstm_1i_dp.sh b/egs/ami/s5b/local/chain/tuning/run_tdnn_lstm_1i_dp.sh
new file mode 100644
index 00000000000..ff083875835
--- /dev/null
+++ b/egs/ami/s5b/local/chain/tuning/run_tdnn_lstm_1i_dp.sh
@@ -0,0 +1,293 @@
+#!/bin/bash
+
+# same as 1g but with TDNN output dim 1024 instead of 512
+# (num-params 1g:21309812 1i: 43447156)
+# results on sdm1 using ihm ali
+#System               tdnn_lstm1g_sp_bi_ihmali_ld5 tdnn_lstm1i_sp_bi_ihmali_ld5
+#WER on dev        38.3      37.6
+#WER on eval        41.6      40.9
+#Final train prob      -0.138017 -0.114135
+#Final valid prob      -0.238659 -0.245208
+#Final train prob (xent)      -1.66834  -1.47648
+#Final valid prob (xent)      -2.17419  -2.16365
+
+
+set -e -o pipefail
+
+# First the options that are passed through to run_ivector_common.sh
+# (some of which are also used in this script directly).
+stage=0
+mic=ihm
+nj=30
+min_seg_len=1.55
+use_ihm_ali=false
+train_set=train_cleaned
+gmm=tri3_cleaned  # the gmm for the target data
+ihm_gmm=tri3  # the gmm for the IHM system (if --use-ihm-ali true).
+num_threads_ubm=32
+nnet3_affix=_cleaned  # cleanup affix for nnet3 and chain dirs, e.g. _cleaned
+dropout_schedule=
+chunk_width=150
+chunk_left_context=40
+chunk_right_context=0
+label_delay=5
+# The rest are configs specific to this script.  Most of the parameters
+# are just hardcoded at this level, in the commands below.
+train_stage=-10
+tree_affix=  # affix for tree directory, e.g. "a" or "b", in case we change the configuration.
+tlstm_affix=1i  #affix for TDNN-LSTM directory, e.g. "a" or "b", in case we change the configuration.
+common_egs_dir=  # you can set this to use previously dumped egs.
+
+
+# decode options
+extra_left_context=50
+frames_per_chunk=
+
+# 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
+
+
+local/nnet3/run_ivector_common.sh --stage $stage \
+                                  --mic $mic \
+                                  --nj $nj \
+                                  --min-seg-len $min_seg_len \
+                                  --train-set $train_set \
+                                  --gmm $gmm \
+                                  --num-threads-ubm $num_threads_ubm \
+                                  --nnet3-affix "$nnet3_affix"
+
+# Note: the first stage of the following script is stage 8.
+local/nnet3/prepare_lores_feats.sh --stage $stage \
+                                   --mic $mic \
+                                   --nj $nj \
+                                   --min-seg-len $min_seg_len \
+                                   --use-ihm-ali $use_ihm_ali \
+                                   --train-set $train_set
+
+if $use_ihm_ali; then
+  gmm_dir=exp/ihm/${ihm_gmm}
+  ali_dir=exp/${mic}/${ihm_gmm}_ali_${train_set}_sp_comb_ihmdata
+  lores_train_data_dir=data/$mic/${train_set}_ihmdata_sp_comb
+  tree_dir=exp/$mic/chain${nnet3_affix}/tree_bi${tree_affix}_ihmdata
+  lat_dir=exp/$mic/chain${nnet3_affix}/${gmm}_${train_set}_sp_comb_lats_ihmdata
+  dir=exp/$mic/chain${nnet3_affix}/tdnn_lstm${tlstm_affix}_sp_bi_ihmali
+  # note: the distinction between when we use the 'ihmdata' suffix versus
+  # 'ihmali' is pretty arbitrary.
+else
+  gmm_dir=exp/${mic}/$gmm
+  ali_dir=exp/${mic}/${gmm}_ali_${train_set}_sp_comb
+  lores_train_data_dir=data/$mic/${train_set}_sp_comb
+  tree_dir=exp/$mic/chain${nnet3_affix}/tree_bi${tree_affix}
+  lat_dir=exp/$mic/chain${nnet3_affix}/${gmm}_${train_set}_sp_comb_lats
+  dir=exp/$mic/chain${nnet3_affix}/tdnn_lstm${tlstm_affix}_sp_bi
+fi
+
+if [ $label_delay -gt 0 ]; then dir=${dir}_ld$label_delay; fi
+
+train_data_dir=data/$mic/${train_set}_sp_hires_comb
+train_ivector_dir=exp/$mic/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires_comb
+final_lm=`cat data/local/lm/final_lm`
+LM=$final_lm.pr1-7
+
+
+for f in $gmm_dir/final.mdl $lores_train_data_dir/feats.scp \
+   $train_data_dir/feats.scp $train_ivector_dir/ivector_online.scp; do
+  [ ! -f $f ] && echo "$0: expected file $f to exist" && exit 1
+done
+
+
+if [ $stage -le 11 ]; then
+  if [ -f $ali_dir/ali.1.gz ]; then
+    echo "$0: alignments in $ali_dir appear to already exist.  Please either remove them "
+    echo " ... or use a later --stage option."
+    exit 1
+  fi
+  echo "$0: aligning perturbed, short-segment-combined ${maybe_ihm}data"
+  steps/align_fmllr.sh --nj $nj --cmd "$train_cmd" \
+     ${lores_train_data_dir} data/lang $gmm_dir $ali_dir
+fi
+
+[ ! -f $ali_dir/ali.1.gz ] && echo  "$0: expected $ali_dir/ali.1.gz to exist" && exit 1
+
+if [ $stage -le 12 ]; then
+  echo "$0: creating lang directory with one state per phone."
+  # 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.]
+  if [ -d data/lang_chain ]; then
+    if [ data/lang_chain/L.fst -nt data/lang/L.fst ]; then
+      echo "$0: data/lang_chain already exists, not overwriting it; continuing"
+    else
+      echo "$0: data/lang_chain already exists and seems to be older than data/lang..."
+      echo " ... not sure what to do.  Exiting."
+      exit 1;
+    fi
+  else
+    cp -r data/lang data/lang_chain
+    silphonelist=$(cat data/lang_chain/phones/silence.csl) || exit 1;
+    nonsilphonelist=$(cat data/lang_chain/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 >data/lang_chain/topo
+  fi
+fi
+
+if [ $stage -le 13 ]; then
+  # Get the alignments as lattices (gives the chain training more freedom).
+  # use the same num-jobs as the alignments
+  steps/align_fmllr_lats.sh --nj 100 --cmd "$train_cmd" ${lores_train_data_dir} \
+    data/lang $gmm_dir $lat_dir
+  rm $lat_dir/fsts.*.gz # save space
+fi
+
+if [ $stage -le 14 ]; then
+  # Build a tree using our new topology.  We know we have alignments for the
+  # speed-perturbed data (local/nnet3/run_ivector_common.sh made them), so use
+  # those.
+  if [ -f $tree_dir/final.mdl ]; then
+    echo "$0: $tree_dir/final.mdl already exists, refusing to overwrite it."
+    exit 1;
+  fi
+  steps/nnet3/chain/build_tree.sh --frame-subsampling-factor 3 \
+      --context-opts "--context-width=2 --central-position=1" \
+      --leftmost-questions-truncate -1 \
+      --cmd "$train_cmd" 4200 ${lores_train_data_dir} data/lang_chain $ali_dir $tree_dir
+fi
+
+xent_regularize=0.1
+
+if [ $stage -le 15 ]; then
+  echo "$0: creating neural net configs using the xconfig parser";
+
+  num_targets=$(tree-info $tree_dir/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
+
+  # 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 dropout-proportion=0.0
+  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
+  lstmp-layer name=lstm2 cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 dropout-proportion=0.0
+  relu-renorm-layer name=tdnn7 input=Append(-3,0,3) dim=1024
+  relu-renorm-layer name=tdnn8 input=Append(-3,0,3) dim=1024
+  relu-renorm-layer name=tdnn9 input=Append(-3,0,3) dim=1024
+  lstmp-layer name=lstm3 cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 dropout-proportion=0.0
+
+  ## adding the layers for chain branch
+  output-layer name=output input=lstm3 output-delay=$label_delay include-log-softmax=false dim=$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.
+  output-layer name=output-xent input=lstm3 output-delay=$label_delay dim=$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 16 ]; 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/ami-$(date +'%m_%d_%H_%M')/s5b/$dir/egs/storage $dir/egs/storage
+  fi
+
+ steps/nnet3/chain/train.py --stage $train_stage \
+    --cmd "$decode_cmd" \
+    --feat.online-ivector-dir $train_ivector_dir \
+    --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.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 \
+    --trainer.dropout-schedule $dropout_schedule \
+    --trainer.num-chunk-per-minibatch 64 \
+    --trainer.frames-per-iter 1500000 \
+    --trainer.num-epochs 4 \
+    --trainer.optimization.shrink-value 0.99 \
+    --trainer.optimization.num-jobs-initial 2 \
+    --trainer.optimization.num-jobs-final 12 \
+    --trainer.optimization.initial-effective-lrate 0.001 \
+    --trainer.optimization.final-effective-lrate 0.0001 \
+    --trainer.max-param-change 2.0 \
+    --trainer.deriv-truncate-margin 8 \
+    --cleanup.remove-egs true \
+    --feat-dir $train_data_dir \
+    --tree-dir $tree_dir \
+    --lat-dir $lat_dir \
+    --dir $dir
+fi
+
+
+graph_dir=$dir/graph_${LM}
+if [ $stage -le 17 ]; then
+  # Note: it might appear that this data/lang_chain directory is mismatched, and it is as
+  # far as the 'topo' is concerned, but this script doesn't read the 'topo' from
+  # the lang directory.
+  utils/mkgraph.sh --left-biphone --self-loop-scale 1.0 data/lang_${LM} $dir $graph_dir
+fi
+
+if [ $stage -le 18 ]; then
+  rm $dir/.error 2>/dev/null || true
+
+  [ -z $extra_left_context ] && extra_left_context=$chunk_left_context;
+  [ -z $frames_per_chunk ] && frames_per_chunk=$chunk_width;
+
+  for decode_set in dev eval; do
+      (
+      steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
+          --nj $nj --cmd "$decode_cmd" \
+          --extra-left-context $extra_left_context  \
+          --frames-per-chunk "$frames_per_chunk" \
+          --online-ivector-dir exp/$mic/nnet3${nnet3_affix}/ivectors_${decode_set}_hires \
+          --scoring-opts "--min-lmwt 5 " \
+         $graph_dir data/$mic/${decode_set}_hires $dir/decode_${decode_set} || exit 1;
+      ) || touch $dir/.error &
+  done
+  wait
+  if [ -f $dir/.error ]; then
+    echo "$0: something went wrong in decoding"
+    exit 1
+  fi
+fi
+exit 0
diff --git a/egs/wsj/s5/steps/libs/nnet3/train/common.py b/egs/wsj/s5/steps/libs/nnet3/train/common.py
index 7009b1cba6e..6c73d417009 100644
--- a/egs/wsj/s5/steps/libs/nnet3/train/common.py
+++ b/egs/wsj/s5/steps/libs/nnet3/train/common.py
@@ -395,8 +395,8 @@ def _parse_dropout_string(num_archives_to_process, dropout_str):
                             "at least the start and end dropouts")
 
         # Starting dropout proportion
-        dropout_values.append((0, float(parts[0])))
-
+        dropout_values.append((0, float(parts[0]))) 
+        data_fraction_one_previous='' # used to control situations like: 0.2@0.75,0@0.75
         for i in range(1, len(parts) - 1):
             value_x_pair = parts[i].split('@')
             if len(value_x_pair) == 1:
@@ -407,9 +407,15 @@ def _parse_dropout_string(num_archives_to_process, dropout_str):
             else:
                 assert len(value_x_pair) == 2
                 dropout_proportion, data_fraction = value_x_pair
-                dropout_values.append(
-                    (float(data_fraction) * num_archives_to_process,
-                     float(dropout_proportion)))
+                if data_fraction == data_fraction_one_previous :
+                    dropout_values.append(
+                        (float(data_fraction) * num_archives_to_process + 1.0,
+                        float(dropout_proportion)))                    
+                else:
+                    dropout_values.append(
+                        (float(data_fraction) * num_archives_to_process,
+                        float(dropout_proportion)))
+            _, data_fraction_one_previous = value_x_pair
 
         dropout_values.append((num_archives_to_process, float(parts[-1])))
     except Exception:
diff --git a/egs/wsj/s5/steps/libs/nnet3/xconfig/lstm.py b/egs/wsj/s5/steps/libs/nnet3/xconfig/lstm.py
index 1ac860ffa9c..023eece93da 100644
--- a/egs/wsj/s5/steps/libs/nnet3/xconfig/lstm.py
+++ b/egs/wsj/s5/steps/libs/nnet3/xconfig/lstm.py
@@ -249,7 +249,8 @@ def set_default_configs(self):
                         'ng-affine-options' : ' max-change=0.75 ',
                         'self-repair-scale-nonlinearity' : 0.00001,
                         'zeroing-interval' : 20,
-                        'zeroing-threshold' : 15.0
+                        'zeroing-threshold' : 15.0,
+                        'dropout-proportion' : -1.0 # -1.0 stands for no dropout will be added
                        }
 
     def set_derived_configs(self):
@@ -280,6 +281,11 @@ def check_configs(self):
                 raise xparser_error("{0} has invalid value {2}.".format(self.layer_type,
                                                                                key,
                                                                                self.config[key]))
+        if ((self.config['dropout-proportion'] > 1.0 or
+             self.config['dropout-proportion'] < 0.0) and
+             self.config['dropout-proportion'] != -1.0 ):
+             raise xparser_error("dropout-proportion has invalid value {0}.".format(self.config['dropout-proportion']))
+
     def auxiliary_outputs(self):
         return ['c_t']
 
@@ -339,6 +345,8 @@ def generate_lstm_config(self):
                                 abs(delay)))
         affine_str = self.config['ng-affine-options']
         pes_str = self.config['ng-per-element-scale-options']
+        lstm_dropout_value = self.config['dropout-proportion']
+        lstm_dropout_str = 'dropout-proportion='+str(self.config['dropout-proportion'])
 
         # Natural gradient per element scale parameters
         # TODO: decide if we want to keep exposing these options
@@ -418,13 +426,21 @@ def generate_lstm_config(self):
 
         # add the recurrent connections
         configs.append("# projection matrices : Wrm and Wpm")
+        if lstm_dropout_value != -1.0:
+            configs.append("component name={0}.W_rp.m.dropout type=DropoutComponent dim={1} {2}".format(name, cell_dim, lstm_dropout_str))
         configs.append("component name={0}.W_rp.m type=NaturalGradientAffineComponent input-dim={1} output-dim={2} {3}".format(name, cell_dim, rec_proj_dim + nonrec_proj_dim, affine_str))
         configs.append("component name={0}.r type=BackpropTruncationComponent dim={1} {2}".format(name, rec_proj_dim, bptrunc_str))
 
         configs.append("# r_t and p_t : rp_t will be the output")
-        configs.append("component-node name={0}.rp_t component={0}.W_rp.m input={0}.m_t".format(name))
-        configs.append("dim-range-node name={0}.r_t_preclip input-node={0}.rp_t dim-offset=0 dim={1}".format(name, rec_proj_dim))
-        configs.append("component-node name={0}.r_t component={0}.r input={0}.r_t_preclip".format(name))
+        if lstm_dropout_value != -1.0:
+            configs.append("component-node name={0}.rp_t.dropout component={0}.W_rp.m.dropout input={0}.m_t".format(name))
+            configs.append("component-node name={0}.rp_t component={0}.W_rp.m input={0}.rp_t.dropout".format(name))
+            configs.append("dim-range-node name={0}.r_t_preclip input-node={0}.rp_t dim-offset=0 dim={1}".format(name, rec_proj_dim))
+            configs.append("component-node name={0}.r_t component={0}.r input={0}.r_t_preclip".format(name))
+        else:
+            configs.append("component-node name={0}.rp_t component={0}.W_rp.m input={0}.m_t".format(name))
+            configs.append("dim-range-node name={0}.r_t_preclip input-node={0}.rp_t dim-offset=0 dim={1}".format(name, rec_proj_dim))
+            configs.append("component-node name={0}.r_t component={0}.r input={0}.r_t_preclip".format(name))
 
         return configs
 
@@ -745,6 +761,7 @@ def set_default_configs(self):
                         'ng-affine-options' : ' max-change=1.5',
                         'zeroing-interval' : 20,
                         'zeroing-threshold' : 15.0
+
                         }
 
     def set_derived_configs(self):

From 5435f236edb09a864e8a4c9847bf63588f5f8693 Mon Sep 17 00:00:00 2001
From: Gaofeng Cheng <770579626@qq.com>
Date: Wed, 14 Dec 2016 22:00:00 +0800
Subject: [PATCH 2/2] adding frame level dropout to TDNN+LSTM on AMI SDM1 #1248

---
 .../local/chain/tuning/run_tdnn_lstm_1i_dp.sh | 20 ++++++++++---------
 1 file changed, 11 insertions(+), 9 deletions(-)

diff --git a/egs/ami/s5b/local/chain/tuning/run_tdnn_lstm_1i_dp.sh b/egs/ami/s5b/local/chain/tuning/run_tdnn_lstm_1i_dp.sh
index ff083875835..ea7c01b79ec 100644
--- a/egs/ami/s5b/local/chain/tuning/run_tdnn_lstm_1i_dp.sh
+++ b/egs/ami/s5b/local/chain/tuning/run_tdnn_lstm_1i_dp.sh
@@ -1,15 +1,17 @@
 #!/bin/bash
+# Copyright 2016  University of Chinese Academy of Sciences (Author: Cheng Gaofeng)
+# Apache 2.0
 
-# same as 1g but with TDNN output dim 1024 instead of 512
+# same as 1i but with frame level dropout
 # (num-params 1g:21309812 1i: 43447156)
 # results on sdm1 using ihm ali
-#System               tdnn_lstm1g_sp_bi_ihmali_ld5 tdnn_lstm1i_sp_bi_ihmali_ld5
-#WER on dev        38.3      37.6
-#WER on eval        41.6      40.9
-#Final train prob      -0.138017 -0.114135
-#Final valid prob      -0.238659 -0.245208
-#Final train prob (xent)      -1.66834  -1.47648
-#Final valid prob (xent)      -2.17419  -2.16365
+#System            tdnn_lstm1i_sp_bi_ihmali_ld5
+#WER on dev        37.6            36.7
+#WER on eval       40.9            39.9
+#Final train prob      -0.114135   -0.118
+#Final valid prob      -0.245208   -0.246
+#Final train prob (xent)      -1.47648  -1.54
+#Final valid prob (xent)      -2.16365  -2.10
 
 
 set -e -o pipefail
@@ -26,7 +28,7 @@ gmm=tri3_cleaned  # the gmm for the target data
 ihm_gmm=tri3  # the gmm for the IHM system (if --use-ihm-ali true).
 num_threads_ubm=32
 nnet3_affix=_cleaned  # cleanup affix for nnet3 and chain dirs, e.g. _cleaned
-dropout_schedule=
+dropout_schedule='0,0@0.20,0.5@0.50,0@0.50,0'
 chunk_width=150
 chunk_left_context=40
 chunk_right_context=0