Bug fixes and cosmetic changes in Iban recipe. (#746)

egs/iban/s5/cmd.sh

@@ -1,2 +1,5 @@
 export train_cmd="run.pl --max-jobs-run 32"
 export decode_cmd="run.pl --max-jobs-run 32"
+
+#export train_cmd="queue.pl"
+#export decode_cmd="queue.pl --mem 4G"

egs/iban/s5/local/nnet3/run_tdnn.sh

@@ -1,6 +1,6 @@
 #!/bin/bash
 
-# Combined from from WSJ + RM 
+# Combined from from WSJ + RM
 
 # this is the standard "tdnn" system, built in nnet3; it's what we use to
 # call multi-splice.
@@ -12,10 +12,11 @@
 # If you want to run without GPU you'd have to call train_tdnn.sh with --gpu false,
 # --num-threads 16 and --minibatch-size 128.
 
-stage=9
+stage=1
 train_stage=-10
-dir=exp/nnet3/nnet_tdnn_h_sp_4_850_170
-. cmd.sh
+dir=exp/nnet3/tdnn_1
+
+. ./cmd.sh
 . ./path.sh
 . ./utils/parse_options.sh
 
@@ -31,38 +32,55 @@ fi
 local/nnet3/run_ivector_common.sh --stage $stage || exit 1;
 
 if [ $stage -le 9 ]; then
-  #  --splice-indexes "-2,-1,0,1,2 -3,1 -5,3 0 0 " \
-  #  --splice-indexes "-4,-3,-2,-1,0,1,2,3,4  0  -2,2  0  -4,4 0" \
-
-  steps/nnet3/train_tdnn.sh --stage $train_stage \
-    --num-jobs-initial 2 --num-jobs-final 4 \
-    --splice-indexes "-4,-3,-2,-1,0,1,2,3,4  0  -2,2  0  -4,4 0 0" \
-    --num-epochs 4 \
-    --add-layers-period 1 \
-    --feat-type raw \
-    --online-ivector-dir exp/nnet3/ivectors_train\
-    --cmvn-opts "--norm-means=false --norm-vars=false" \
-    --initial-effective-lrate 0.005 --final-effective-lrate 0.0005 \
-    --cmd "$decode_cmd" \
-    --pnorm-input-dim 850 \
-    --pnorm-output-dim 170 \
-    --num-jobs-compute-prior 4\
-    data/train_hires data/lang exp/nnet3/tri3b_ali_sp $dir  || exit 1;
+  echo "$0: creating neural net configs";
+
+  # create the config files for nnet initialization
+  python steps/nnet3/tdnn/make_configs.py  \
+    --feat-dir data/train_hires \
+    --ivector-dir exp/nnet3/ivectors_train \
+    --ali-dir exp/nnet3/tri3b_ali_sp \
+    --relu-dim 256 \
+    --splice-indexes=" -2,-1,0,1,2  -1,0,1  -1,0,1  -1,0,1  -1,0,1 -1,0,1 0 " \
+    --use-presoftmax-prior-scale true \
+   $dir/configs || exit 1;
 fi
 
 
+
 if [ $stage -le 10 ]; then
+
+  steps/nnet3/train_dnn.py --stage $train_stage \
+    --cmd="$decode_cmd" \
+    --trainer.optimization.num-jobs-initial 2 \
+    --trainer.optimization.num-jobs-final 4 \
+    --trainer.num-epochs 4 \
+    --trainer.add-layers-period 1 \
+    --feat.online-ivector-dir exp/nnet3/ivectors_train\
+    --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
+    --trainer.num-epochs 2 \
+    --trainer.optimization.initial-effective-lrate 0.005 \
+    --trainer.optimization.final-effective-lrate 0.0005 \
+    --trainer.samples-per-iter 120000 \
+    --cleanup.preserve-model-interval 10 \
+    --feat-dir data/train_hires \
+    --ali-dir exp/nnet3/tri3b_ali_sp \
+    --lang data/lang \
+    --dir=$dir  || exit 1;
+fi
+
+
+if [ $stage -le 11 ]; then
   # this does offline decoding that should give the same results as the real
   # online decoding.
   graph_dir=exp/tri3b/graph
   # use already-built graphs.
     steps/nnet3/decode.sh --nj 6 --cmd "$decode_cmd" \
         --online-ivector-dir exp/nnet3/ivectors_dev --iter final\
        $graph_dir data/dev_hires $dir/decode_dev || exit 1;
-  
+
 fi
 
-if [ $stage -le 11 ]; then
+if [ $stage -le 12 ]; then
    steps/lmrescore_const_arpa.sh  --cmd "$decode_cmd" \
      data/lang_test/ data/lang_big/ data/dev \
     ${dir}/decode_dev ${dir}/decode_dev.rescored

egs/iban/s5/local/prepare_lm.sh

@@ -10,7 +10,7 @@ set -e -o pipefail
 
 local/train_lms_srilm.sh --train-text data/train/text data/ data/srilm
 
-nl -nrz -w10  iban/LM/iban-bp-2012.txt | sort -R > data/local/external_text
+nl -nrz -w10  corpus/LM/iban-bp-2012.txt | sort -R > data/local/external_text
 local/train_lms_srilm.sh --train-text data/local/external_text data/ data/srilm_external
 
 # let's do ngram interpolation of the previous two LMs