[egs] Added tri3b and chain training for Aurora4 (#3638)

egs/aurora4/s5/RESULTS

@@ -1,8 +1,19 @@
-for x in exp/{mono,tri,sgmm,nnet,dnn}*/decode*; do [ -d $x ] && grep WER $x/wer_* | utils/best_wer.sh; done
+for x in exp/{mono,tri,sgmm,nnet,dnn,chain/tdnn*}*/decode*; do [ -d $x ] && grep WER $x/wer_* | utils/best_wer.sh; done
 
-%WER 19.61 [ 14698 / 74942, 1233 ins, 3759 del, 9706 sub ] exp/tri2b_multi/decode_tgpr_5k_eval92/wer_13
-%WER 13.93 [ 10437 / 74942, 732 ins, 2695 del, 7010 sub ] exp/tri3a_dnn/decode_tgpr_5k_eval92/wer_10
-%WER 13.61 [ 10202 / 74942, 660 ins, 2987 del, 6555 sub ] exp/tri4a_dnn/decode_tgpr_5k_eval92/wer_11
+# mono
+%WER 37.42 [ 14223 / 38010, 1030 ins, 2613 del, 10580 sub ] exp/mono0a_multi/decode_tgpr_0166/wer_10
+%WER 38.18 [ 28612 / 74942, 1919 ins, 5319 del, 21374 sub ] exp/mono0a_multi/decode_tgpr_eval92/wer_10
 
+# tri2b
+%WER 20.42 [ 7763 / 38010, 827 ins, 1905 del, 5031 sub ] exp/tri2b_multi/decode_tgpr_5k_0166/wer_12
+%WER 19.61 [ 14728 / 74942, 1411 ins, 3548 del, 9769 sub ] exp/tri2b_multi/decode_tgpr_5k_eval92/wer_12
+
+# tri3b
+%WER 15.71 [ 5970 / 38010, 641 ins, 1403 del, 3926 sub ] exp/tri3b_multi/decode_tgpr_0166/wer_13
+%WER 15.28 [ 11454 / 74942, 1082 ins, 2633 del, 7739 sub ] exp/tri3b_multi/decode_tgpr_eval92/wer_13
+
+# chain
+%WER 7.88 [ 2994 / 38010, 216 ins, 1045 del, 1733 sub ] exp/chain/tdnn1a_sp/decode_tgpr_5k_0166/wer_15
+%WER 7.67 [ 5745 / 74942, 392 ins, 1758 del, 3595 sub ] exp/chain/tdnn1a_sp/decode_tgpr_5k_eval92/wer_13
 
 for x in /mnt/matylda3/qmallidi/Karels_New-Parametric-ReLU/kaldi/egs/aurora4/s5_PReLU/exp/{mono,tri,sgmm,nnet,dnn}*/decode*; do [ -d $x ] && grep WER $x/wer_* | /mnt/matylda5/iveselyk/DEVEL/kaldi-official/egs/aurora4/s5/utils/best_wer.sh; done

egs/aurora4/s5/conf/mfcc_hires.conf

@@ -0,0 +1,10 @@
+# config for high-resolution MFCC features, intended for neural network training
+# Note: we keep all cepstra, so it has the same info as filterbank features,
+# but MFCC is more easily compressible (because less correlated) which is why 
+# we prefer this method.
+--use-energy=false   # use average of log energy, not energy.
+--num-mel-bins=40     # similar to Google's setup.
+--num-ceps=40     # there is no dimensionality reduction.
+--low-freq=20     # low cutoff frequency for mel bins... this is high-bandwidth data, so
+                  # there might be some information at the low end.
+--high-freq=-400 # high cutoff frequently, relative to Nyquist of 8000 (=7600) 

egs/aurora4/s5/conf/online_cmvn.conf

@@ -0,0 +1 @@
+# configuration file for apply-cmvn-online, used in the script ../local/online/run_online_decoding_nnet2.sh

egs/aurora4/s5/local/chain/compare_wer.sh

@@ -0,0 +1,137 @@
+#!/bin/bash
+
+# this script is used for comparing decoding results between systems.
+# e.g. local/chain/compare_wer.sh exp/chain/tdnn_{c,d}_sp
+# For use with discriminatively trained systems you specify the epochs after a colon:
+# for instance,
+# local/chain/compare_wer.sh exp/chain/tdnn_c_sp exp/chain/tdnn_c_sp_smbr:{1,2,3}
+
+
+if [ $# == 0 ]; then
+  echo "Usage: $0: [--looped] [--online] <dir1> [<dir2> ... ]"
+  echo "e.g.: $0 exp/chain/tdnn_{b,c}_sp"
+  echo "or (with epoch numbers for discriminative training):"
+  echo "$0 exp/chain/tdnn_b_sp_disc:{1,2,3}"
+  exit 1
+fi
+
+echo "# $0 $*"
+
+include_looped=false
+if [ "$1" == "--looped" ]; then
+  include_looped=true
+  shift
+fi
+include_online=false
+if [ "$1" == "--online" ]; then
+  include_online=true
+  shift
+fi
+
+
+used_epochs=false
+
+# this function set_names is used to separate the epoch-related parts of the name
+# [for discriminative training] and the regular parts of the name.
+# If called with a colon-free directory name, like:
+#  set_names exp/chain/tdnn_lstm1e_sp_bi_smbr
+# it will set dir=exp/chain/tdnn_lstm1e_sp_bi_smbr and epoch_infix=""
+# If called with something like:
+#  set_names exp/chain/tdnn_d_sp_smbr:3
+# it will set dir=exp/chain/tdnn_d_sp_smbr and epoch_infix="_epoch3"
+
+
+set_names() {
+  if [ $# != 1 ]; then
+    echo "compare_wer_general.sh: internal error"
+    exit 1  # exit the program
+  fi
+  dirname=$(echo $1 | cut -d: -f1)
+  epoch=$(echo $1 | cut -s -d: -f2)
+  if [ -z $epoch ]; then
+    epoch_infix=""
+  else
+    used_epochs=true
+    epoch_infix=_epoch${epoch}
+  fi
+}
+
+
+
+echo -n "# System               "
+for x in $*; do   printf "% 10s" " $(basename $x)";   done
+echo
+
+strings=(
+"# WER eval92 (tgpr_5k)         "
+"# WER 0166 (tgpr_5k)           ")
+
+for n in 0 1; do
+   echo -n "${strings[$n]}"
+   for x in $*; do
+     set_names $x  # sets $dirname and $epoch_infix
+     decode_names=(tgpr_5k_eval92 tgpr_5k_0166)
+
+     wer=$(cat $dirname/decode_${decode_names[$n]}/wer* | utils/best_wer.sh | awk '{print $2}')
+     printf "% 10s" $wer
+   done
+   echo
+   if $include_looped; then
+     echo -n "#             [looped:]    "
+     for x in $*; do
+       set_names $x  # sets $dirname and $epoch_infix
+       wer=$(cat $dirname/decode_looped_${decode_names[$n]}/wer* | utils/best_wer.sh | awk '{print $2}')
+       printf "% 10s" $wer
+     done
+     echo
+   fi
+   if $include_online; then
+     echo -n "#             [online:]    "
+     for x in $*; do
+       set_names $x  # sets $dirname and $epoch_infix
+       wer=$(cat ${dirname}_online/decode_${decode_names[$n]}/wer* | utils/best_wer.sh | awk '{print $2}')
+       printf "% 10s" $wer
+     done
+     echo
+   fi
+done
+
+
+if $used_epochs; then
+  exit 0;  # the diagnostics aren't comparable between regular and discriminatively trained systems.
+fi
+
+
+echo -n "# Final train prob     "
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_train.final.log | grep -v xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final valid prob     "
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_valid.final.log | grep -v xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final train prob (xent)"
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_train.final.log | grep -w xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Final valid prob (xent)"
+for x in $*; do
+  prob=$(grep Overall $x/log/compute_prob_valid.final.log | grep -w xent | awk '{printf("%.4f", $8)}')
+  printf "% 10s" $prob
+done
+echo
+
+echo -n "# Num-params              "
+for x in $*; do
+  printf "% 10s" $(grep num-parameters $x/log/progress.1.log | awk '{print $2}')
+done
+echo

egs/aurora4/s5/local/chain/run_tdnn.sh

@@ -0,0 +1 @@
+tuning/run_tdnn_1a.sh