Sprak nnet3

egs/sprakbanken/s5/RESULTS

@@ -1,25 +1,28 @@
-%WER 49.19 [ 5318 / 10811, 481 ins, 1511 del, 3326 sub ] exp/mono0a/decode_3g_test1k/wer_9
-%WER 47.28 [ 5111 / 10811, 443 ins, 1489 del, 3179 sub ] exp/mono0a/decode_b3g_test1k/wer_10
-%WER 16.19 [ 1750 / 10811, 397 ins, 323 del, 1030 sub ] exp/sgmm2_5a/decode_3g_test1k/wer_9
-%WER 15.10 [ 1632 / 10811, 404 ins, 305 del, 923 sub ] exp/sgmm2_5b/decode_3g_test1k/wer_9
-%WER 14.94 [ 1615 / 10811, 390 ins, 310 del, 915 sub ] exp/sgmm2_5b/decode_4g_test1k/wer_9
-%WER 14.36 [ 1553 / 10811, 376 ins, 264 del, 913 sub ] exp/sgmm2_5c/decode_3g_test1k/wer_9
-%WER 14.18 [ 1533 / 10811, 367 ins, 266 del, 900 sub ] exp/sgmm2_5c/decode_4g_test1k/wer_9
-%WER 25.61 [ 2769 / 10811, 511 ins, 539 del, 1719 sub ] exp/tri1/decode_3g_test1k/wer_10
-%WER 25.12 [ 2716 / 10811, 444 ins, 571 del, 1701 sub ] exp/tri1/decode_b3g_test1k/wer_11
-%WER 23.81 [ 2574 / 10811, 426 ins, 564 del, 1584 sub ] exp/tri2a/decode_3g_test1k/wer_12
-%WER 23.22 [ 2510 / 10811, 457 ins, 517 del, 1536 sub ] exp/tri2a/decode_3g_test1k_fromlats/wer_11
-%WER 22.18 [ 2398 / 10811, 436 ins, 495 del, 1467 sub ] exp/tri2b/decode_3g_test1k/wer_11
-%WER 21.87 [ 2364 / 10811, 380 ins, 553 del, 1431 sub ] exp/tri2b/decode_3g_test1k_mbr/wer_13
-%WER 18.98 [ 2052 / 10811, 451 ins, 372 del, 1229 sub ] exp/tri3b_20k/decode_3g_test1k/wer_11
-%WER 22.62 [ 2445 / 10811, 468 ins, 460 del, 1517 sub ] exp/tri3b_20k/decode_3g_test1k.si/wer_10
-%WER 19.31 [ 2088 / 10811, 440 ins, 388 del, 1260 sub ] exp/tri3b/decode_3g_test1k/wer_11
-%WER 23.19 [ 2507 / 10811, 435 ins, 520 del, 1552 sub ] exp/tri3b/decode_3g_test1k.si/wer_12
-%WER 19.06 [ 2061 / 10811, 427 ins, 384 del, 1250 sub ] exp/tri3b/decode_4g_test1k/wer_11
-%WER 23.20 [ 2508 / 10811, 447 ins, 520 del, 1541 sub ] exp/tri3b/decode_4g_test1k.si/wer_11
-%WER 17.42 [ 1883 / 10811, 416 ins, 359 del, 1108 sub ] exp/tri4a/decode_3g_test1k/wer_13
-%WER 20.86 [ 2255 / 10811, 403 ins, 473 del, 1379 sub ] exp/tri4a/decode_3g_test1k.si/wer_13
-%WER 17.52 [ 1894 / 10811, 396 ins, 372 del, 1126 sub ] exp/tri4b/decode_3g_test1k/wer_13
-%WER 20.82 [ 2251 / 10811, 399 ins, 471 del, 1381 sub ] exp/tri4b/decode_3g_test1k.si/wer_13
-%WER 17.53 [ 1895 / 10811, 403 ins, 375 del, 1117 sub ] exp/tri4b/decode_4g_test1k/wer_13
-%WER 20.99 [ 2269 / 10811, 438 ins, 436 del, 1395 sub ] exp/tri4b/decode_4g_test1k.si/wer_11
+GMM-based systems
+%WER 22.87 [ 24286 / 106172, 3577 ins, 5321 del, 15388 sub ] exp/tri1/decode_fg_dev/wer_12_0.5
+%WER 23.13 [ 24561 / 106172, 3602 ins, 5411 del, 15548 sub ] exp/tri1/decode_tg_dev/wer_12_0.5
+%WER 21.24 [ 22548 / 106172, 4028 ins, 4246 del, 14274 sub ] exp/tri2a/decode_tg_dev/wer_13_0.0
+%WER 19.46 [ 20664 / 106172, 3276 ins, 4332 del, 13056 sub ] exp/tri2b/decode_tg_dev/wer_15_0.5
+%WER 16.80 [ 17839 / 106172, 3238 ins, 3403 del, 11198 sub ] exp/tri3b/decode_fg_dev/wer_17_0.0
+%WER 19.45 [ 20651 / 106172, 3880 ins, 3671 del, 13100 sub ] exp/tri3b/decode_fg_dev.si/wer_15_0.0
+%WER 14.24 [ 9849 / 69165, 2046 ins, 1365 del, 6438 sub ] exp/tri3b/decode_fg_test/wer_16_0.5
+%WER 17.31 [ 11972 / 69165, 2330 ins, 1695 del, 7947 sub ] exp/tri3b/decode_fg_test.si/wer_15_0.5
+%WER 16.94 [ 17984 / 106172, 3361 ins, 3377 del, 11246 sub ] exp/tri3b/decode_tg_dev/wer_16_0.0
+%WER 19.52 [ 20720 / 106172, 3654 ins, 3846 del, 13220 sub ] exp/tri3b/decode_tg_dev.si/wer_17_0.0
+%WER 14.40 [ 9957 / 69165, 2291 ins, 1184 del, 6482 sub ] exp/tri3b/decode_tg_test/wer_16_0.0
+%WER 17.41 [ 12044 / 69165, 2291 ins, 1736 del, 8017 sub ] exp/tri3b/decode_tg_test.si/wer_15_0.5
+nnet3 xent systems
+%WER 11.57 [ 12279 / 106172, 2640 ins, 2442 del, 7197 sub ] exp/nnet3/tdnn0_sp/decode_dev/wer_10_0.0
+%WER 9.89 [ 6841 / 69165, 1542 ins, 917 del, 4382 sub ] exp/nnet3/tdnn0_sp/decode_test/wer_11_0.5
+%WER 10.45 [ 11098 / 106172, 2199 ins, 2272 del, 6627 sub ] exp/nnet3/lstm_0_ld5_sp/decode_dev/wer_9_0.0
+%WER 12.34 [ 8533 / 69165, 1740 ins, 1393 del, 5400 sub ] exp/nnet3/lstm_0_ld5_sp/decode_test/wer_11_1.0
+%WER 10.59 [ 11241 / 106172, 2208 ins, 2304 del, 6729 sub ] exp/nnet3/lstm_bidirectional_ld5_sp/decode_dev/wer_9_0.0
+%WER 12.43 [ 8596 / 69165, 1742 ins, 1426 del, 5428 sub ] exp/nnet3/lstm_bidirectional_ld5_sp/decode_test/wer_11_1.0
+%WER 9.18 [ 9747 / 106172, 1987 ins, 1913 del, 5847 sub ] exp/nnet3/lstm_bidirectional_sp/decode_dev/wer_8_0.0
+Nnet3 chain systems
+%WER 8.48 [ 9001 / 106172, 1559 ins, 1624 del, 5818 sub ] exp/chain/tdnn_lstm1a_sp_bi/decode_dev/wer_9_0.0
+%WER 7.20 [ 4981 / 69165, 915 ins, 402 del, 3664 sub ] exp/chain/tdnn_lstm1a_sp_bi/decode_test/wer_8_1.0
+%WER 10.00 [ 10619 / 106172, 1980 ins, 1896 del, 6743 sub ] exp/chain/tdnn_sp_bi/decode_dev/wer_9_0.0
+%WER 8.58 [ 5936 / 69165, 1059 ins, 667 del, 4210 sub ] exp/chain/tdnn_sp_bi/decode_test/wer_9_1.0
+%WER 9.39 [ 9969 / 106172, 1624 ins, 1912 del, 6433 sub ] exp/chain/lstm1e_sp_bi/decode_dev/wer_8_0.5
+%WER 7.72 [ 5341 / 69165, 1002 ins, 497 del, 3842 sub ] exp/chain/lstm1e_sp_bi/decode_test/wer_8_0.5

egs/sprakbanken/s5/conf/mfcc_hires.conf

@@ -0,0 +1,11 @@
+# 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.
+                  # Needs to be this low to be sensitive to creaky voice
+--high-freq=-400 # high cutoff frequently, relative to Nyquist of 8000 (=7600) 

egs/sprakbanken/s5/conf/online_cmvn.conf

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

egs/sprakbanken/s5/local/chain/compare_wer_general.sh

@@ -0,0 +1,50 @@
+#!/bin/bash
+
+# Prints a table makes it easy to compare WER and objective values across nnet3
+# and chain training runs
+
+echo -n "System               "
+for x in "$@"; do   printf "% 10s" $x;   done
+echo
+
+echo -n "WER on dev(tg) "
+for x in "$@"; do
+  wer=$(grep WER ${x}/decode_dev/wer_* | utils/best_wer.sh | awk '{print $2}')
+  printf "% 10s" $wer
+done
+echo
+
+echo -n "WER on test(tg)  "
+for x in "$@"; do
+  wer=$(grep WER ${x}/decode_test/wer_* | utils/best_wer.sh | awk '{print $2}')
+  printf "% 10s" $wer
+done
+echo
+
+echo -n "Final train prob     "
+for x in "$@"; 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

egs/sprakbanken/s5/local/chain/run_lstm.sh

@@ -0,0 +1 @@
+tuning/run_lstm_1e.sh

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

@@ -0,0 +1 @@
+tuning/run_tdnn_1b.sh

egs/sprakbanken/s5/local/chain/run_tdnn_lstm.sh

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

egs/sprakbanken/s5/local/chain/tuning/run_lstm_1a.sh

@@ -0,0 +1,260 @@
+#!/bin/bash
+
+# run_lstm_1a.sh is a first attempt at an LSTM system, based on xconfigs-- it's
+# probably not very well configured, e.g. the num-params might be too small.
+# recurrent-projection-dim is less than non-recurrent-projection-dim due to an
+# oversight.
+
+# comparison with TDNN system (WER is worse):
+# local/chain/compare_wer_general.sh exp/chain_cleaned/tdnn1b_sp_bi exp/chain_cleaned/lstm1a_sp_bi
+# System                tdnn1b_sp_bi lstm1a_sp_bi
+# WER on dev(orig)          10.2       10.8
+# WER on dev(rescored)       9.6       10.2
+# WER on test(orig)           9.7      10.0
+# WER on test(rescored)       9.2      9.6
+# Final train prob        -0.0928   -0.0848
+# Final valid prob        -0.1178   -0.1098
+# Final train prob (xent)   -1.4666   -1.1692
+# Final valid prob (xent)   -1.5473   -1.2520
+
+
+## how you run this (note: this assumes that the run_lstm.sh soft link points here;
+## otherwise call it directly in its location).
+# by default, with cleanup:
+# local/chain/run_lstm.sh
+
+# without cleanup:
+# local/chain/run_lstm.sh  --train-set train --gmm tri3 --nnet3-affix "" &
+
+# note, if you have already run one of the non-chain nnet3 systems
+# (e.g. local/nnet3/run_tdnn.sh), you may want to run with --stage 14.
+
+# This script (run_lstm_1a) is like run_tdnn_1b.sh except modified to use an LSTM
+# configuration (some aspects borrowed from egs/swbd/s5c/local/chain/tuning/run_lstm_6j.sh).
+
+
+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
+nj=30
+decode_nj=30
+min_seg_len=1.55
+chunk_left_context=40
+chunk_right_context=0
+label_delay=5
+xent_regularize=0.1
+train_set=train_cleaned
+gmm=tri3_cleaned  # the gmm for the target data
+num_threads_ubm=32
+nnet3_affix=_cleaned  # cleanup affix for nnet3 and chain dirs, e.g. _cleaned
+# decode options
+extra_left_context=50
+extra_right_context=0
+frames_per_chunk=150
+
+# 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.
+lstm_affix=1a  #affix for 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.
+
+# 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 \
+                                  --nj $nj \
+                                  --min-seg-len $min_seg_len \
+                                  --train-set $train_set \
+                                  --gmm $gmm \
+                                  --num-threads-ubm $num_threads_ubm \
+                                  --nnet3-affix "$nnet3_affix"
+
+
+gmm_dir=exp/$gmm
+ali_dir=exp/${gmm}_ali_${train_set}_sp_comb
+tree_dir=exp/chain${nnet3_affix}/tree_bi${tree_affix}
+lat_dir=exp/chain${nnet3_affix}/${gmm}_${train_set}_sp_comb_lats
+dir=exp/chain${nnet3_affix}/lstm${lstm_affix}_sp_bi
+train_data_dir=data/${train_set}_sp_hires_comb
+lores_train_data_dir=data/${train_set}_sp_comb
+train_ivector_dir=exp/nnet3${nnet3_affix}/ivectors_${train_set}_sp_hires_comb
+
+
+for f in $gmm_dir/final.mdl $train_data_dir/feats.scp $train_ivector_dir/ivector_online.scp \
+    $lores_train_data_dir/feats.scp $ali_dir/ali.1.gz $gmm_dir/final.mdl; do
+  [ ! -f $f ] && echo "$0: expected file $f to exist" && exit 1
+done
+
+if [ $stage -le 14 ]; 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 [ $stage -le 15 ]; then
+  # Get the alignments as lattices (gives the chain training more freedom).
+  # use the same num-jobs as the alignments
+  steps/align_fmllr_lats.sh --nj 100 --cmd "$train_cmd" ${lores_train_data_dir} \
+    data/lang $gmm_dir $lat_dir
+  rm $lat_dir/fsts.*.gz # save space
+fi
+
+if [ $stage -le 16 ]; 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" 4000 ${lores_train_data_dir} data/lang_chain $ali_dir $tree_dir
+fi
+
+
+if [ $stage -le 17 ]; then
+  mkdir -p $dir
+  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 > $dir/configs/network.xconfig
+  input dim=100 name=ivector
+  input dim=40 name=input
+
+  # please note that it is important to have input layer with the name=input
+  # as the layer immediately preceding the fixed-affine-layer to enable
+  # the use of short notation for the descriptor
+  fixed-affine-layer name=lda input=Append(-2,-1,0,1,2,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+  # check steps/libs/nnet3/xconfig/lstm.py for the other options and defaults
+  lstmp-layer name=lstm1 cell-dim=512 recurrent-projection-dim=128 non-recurrent-projection-dim=256 delay=-3
+  lstmp-layer name=lstm2 cell-dim=512 recurrent-projection-dim=128 non-recurrent-projection-dim=256 delay=-3
+  lstmp-layer name=lstm3 cell-dim=512 recurrent-projection-dim=128 non-recurrent-projection-dim=256 delay=-3
+
+  ## 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 18 ]; 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-$(date +'%m_%d_%H_%M')/s5/$dir/egs/storage $dir/egs/storage
+  fi
+
+ steps/nnet3/chain/train.py --stage $train_stage \
+    --cmd "$decode_cmd" \
+    --feat.online-ivector-dir $train_ivector_dir \
+    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
+    --chain.xent-regularize 0.1 \
+    --chain.leaky-hmm-coefficient 0.1 \
+    --chain.l2-regularize 0.00005 \
+    --chain.apply-deriv-weights false \
+    --chain.lm-opts="--num-extra-lm-states=2000" \
+    --egs.dir "$common_egs_dir" \
+    --egs.opts "--frames-overlap-per-eg 0" \
+    --egs.chunk-width "$frames_per_chunk" \
+    --egs.chunk-left-context "$chunk_left_context" \
+    --egs.chunk-right-context "$chunk_right_context" \
+    --trainer.num-chunk-per-minibatch 128 \
+    --trainer.frames-per-iter 1500000 \
+    --trainer.max-param-change 2.0 \
+    --trainer.num-epochs 4 \
+    --trainer.deriv-truncate-margin 10 \
+    --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.optimization.momentum 0.0 \
+    --cleanup.remove-egs true \
+    --feat-dir $train_data_dir \
+    --tree-dir $tree_dir \
+    --lat-dir $lat_dir \
+    --dir $dir
+fi
+
+
+
+if [ $stage -le 19 ]; 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 --self-loop-scale 1.0 data/lang $dir $dir/graph
+fi
+
+if [ $stage -le 20 ]; then
+  rm $dir/.error 2>/dev/null || true
+  for dset in dev test; do
+      (
+      steps/nnet3/decode.sh --num-threads 4 --nj $decode_nj --cmd "$decode_cmd" \
+          --acwt 1.0 --post-decode-acwt 10.0 \
+          --extra-left-context $extra_left_context  \
+          --extra-right-context $extra_right_context  \
+          --frames-per-chunk "$frames_per_chunk" \
+          --online-ivector-dir exp/nnet3${nnet3_affix}/ivectors_${dset}_hires \
+          --scoring-opts "--min-lmwt 5 " \
+         $dir/graph data/${dset}_hires $dir/decode_${dset} || exit 1;
+      steps/lmrescore_const_arpa.sh --cmd "$decode_cmd" data/lang data/lang_rescore \
+        data/${dset}_hires ${dir}/decode_${dset} ${dir}/decode_${dset}_rescore || 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