madcat arabic: clean scripts, tuning, rescoring, text localization

egs/cifar/v1/image/ocr/make_features.py

@@ -45,10 +45,13 @@
                     'and right side of the image.')
 parser.add_argument('--num-channels', type=int, default=1,
                     help='Number of color channels')
+parser.add_argument('--vertical-shift', type=int, default=0,
+                    help='total number of padding pixel per column')
 parser.add_argument('--fliplr', type=lambda x: (str(x).lower()=='true'), default=False,
                    help="Flip the image left-right for right to left languages")
-parser.add_argument("--augment", type=lambda x: (str(x).lower()=='true'), default=False,
-                   help="performs image augmentation")
+parser.add_argument('--augment_type', type=str, default='no_aug',
+                    choices=['no_aug', 'random_scale','random_shift'],
+                    help='Subset of data to process.')
 args = parser.parse_args()
 
 
@@ -68,7 +71,6 @@ def write_kaldi_matrix(file_handle, matrix, key):
             file_handle.write("\n")
     file_handle.write(" ]\n")
 
-
 def horizontal_pad(im, allowed_lengths = None):
     if allowed_lengths is None:
         left_padding = right_padding = args.padding
@@ -112,6 +114,33 @@ def get_scaled_image_aug(im, mode='normal'):
         return im_scaled_up
     return im
 
+def vertical_shift(im, mode='normal'):
+    if args.vertical_shift == 0:
+        return im
+    total = args.vertical_shift
+    if mode == 'notmid':
+        val = random.randint(0, 1)
+        if val == 0:
+            mode = 'top'
+        else:
+            mode = 'bottom'
+    if mode == 'normal':
+        top = int(total / 2)
+        bottom = total - top
+    elif mode == 'top':  # more padding on top
+        top = random.randint(total / 2, total)
+        bottom = total - top
+    elif mode == 'bottom':  # more padding on bottom
+        top = random.randint(0, total / 2)
+        bottom = total - top
+    width = im.shape[1]
+    im_pad = np.concatenate(
+        (255 * np.ones((top, width), dtype=int) -
+         np.random.normal(2, 1, (top, width)).astype(int), im), axis=0)
+    im_pad = np.concatenate(
+        (im_pad, 255 * np.ones((bottom, width), dtype=int) -
+         np.random.normal(2, 1, (bottom, width)).astype(int)), axis=0)
+    return im_pad
 
 ### main ###
 random.seed(1)
@@ -134,7 +163,6 @@ def get_scaled_image_aug(im, mode='normal'):
 
 num_fail = 0
 num_ok = 0
-aug_setting = ['normal', 'scaled']
 with open(data_list_path) as f:
     for line in f:
         line = line.strip()
@@ -144,21 +172,25 @@ def get_scaled_image_aug(im, mode='normal'):
         im = misc.imread(image_path)
         if args.fliplr:
             im = np.fliplr(im)
-        if args.augment:
-            im_aug = get_scaled_image_aug(im, aug_setting[1])
-        else:
-            im_aug = get_scaled_image_aug(im, aug_setting[0])
-        im_horizontal_padded = horizontal_pad(im_aug, allowed_lengths)
-        if im_horizontal_padded is None:
+        if args.augment_type == 'no_aug' or 'random_shift':
+            im = get_scaled_image_aug(im, 'normal')
+        elif args.augment_type == 'random_scale':
+            im = get_scaled_image_aug(im, 'scaled')
+        im = horizontal_pad(im, allowed_lengths)
+        if im is None:
             num_fail += 1
             continue
+        if args.augment_type == 'no_aug' or 'random_scale':
+            im = vertical_shift(im, 'normal')
+        elif args.augment_type == 'random_shift':
+            im = vertical_shift(im, 'notmid')
         if args.num_channels == 1:
-            data = np.transpose(im_horizontal_padded, (1, 0))
+            data = np.transpose(im, (1, 0))
         elif args.num_channels == 3:
-            H = im_horizontal_padded.shape[0]
-            W = im_horizontal_padded.shape[1]
-            C = im_horizontal_padded.shape[2]
-            data = np.reshape(np.transpose(im_horizontal_padded, (1, 0, 2)), (W, H * C))
+            H = im.shape[0]
+            W = im.shape[1]
+            C = im.shape[2]
+            data = np.reshape(np.transpose(im, (1, 0, 2)), (W, H * C))
         data = np.divide(data, 255.0)
         num_ok += 1
         write_kaldi_matrix(out_fh, data, image_id)

egs/madcat_ar/v1/local/chain/compare_wer.sh

@@ -27,13 +27,27 @@ for x in $*; do
 done
 echo
 
+echo -n "# WER (rescored)             "
+for x in $*; do
+  wer=$(cat $x/decode_test_rescored/scoring_kaldi/best_wer | awk '{print $2}')
+  printf "% 10s" $wer
+done
+echo
+
 echo -n "# CER                        "
 for x in $*; do
   cer=$(cat $x/decode_test/scoring_kaldi/best_cer | awk '{print $2}')
   printf "% 10s" $cer
 done
 echo
 
+echo -n "# CER (rescored)             "
+for x in $*; do
+  cer=$(cat $x/decode_test_rescored/scoring_kaldi/best_cer | awk '{print $2}')
+  printf "% 10s" $cer
+done
+echo
+
 if $used_epochs; then
   exit 0;  # the diagnostics aren't comparable between regular and discriminatively trained systems.
 fi

egs/madcat_ar/v1/local/chain/tuning/run_cnn_1a.sh

@@ -21,18 +21,16 @@ reporting_email=
 # chain options
 train_stage=-10
 xent_regularize=0.1
-frame_subsampling_factor=4
 # training chunk-options
 chunk_width=340,300,200,100
 num_leaves=500
 # we don't need extra left/right context for TDNN systems.
-chunk_left_context=0
-chunk_right_context=0
 tdnn_dim=450
 # training options
 srand=0
 remove_egs=false
-lang_test=lang_test
+lang_decode=data/lang
+lang_rescore=data/lang_rescore_6g
 # End configuration section.
 echo "$0 $@"  # Print the command line for logging
 
@@ -168,13 +166,13 @@ if [ $stage -le 5 ]; then
     --chain.leaky-hmm-coefficient=0.1 \
     --chain.l2-regularize=0.00005 \
     --chain.apply-deriv-weights=false \
-    --chain.lm-opts="--num-extra-lm-states=500" \
-    --chain.frame-subsampling-factor=$frame_subsampling_factor \
-    --chain.alignment-subsampling-factor=$frame_subsampling_factor \
+    --chain.lm-opts="--ngram-order=2 --no-prune-ngram-order=1 --num-extra-lm-states=1000" \
+    --chain.frame-subsampling-factor=4 \
+    --chain.alignment-subsampling-factor=4 \
     --trainer.srand=$srand \
     --trainer.max-param-change=2.0 \
     --trainer.num-epochs=4 \
-    --trainer.frames-per-iter=1000000 \
+    --trainer.frames-per-iter=2000000 \
     --trainer.optimization.num-jobs-initial=3 \
     --trainer.optimization.num-jobs-final=16 \
     --trainer.optimization.initial-effective-lrate=0.001 \
@@ -183,10 +181,6 @@ if [ $stage -le 5 ]; then
     --trainer.num-chunk-per-minibatch=64,32 \
     --trainer.optimization.momentum=0.0 \
     --egs.chunk-width=$chunk_width \
-    --egs.chunk-left-context=$chunk_left_context \
-    --egs.chunk-right-context=$chunk_right_context \
-    --egs.chunk-left-context-initial=0 \
-    --egs.chunk-right-context-final=0 \
     --egs.dir="$common_egs_dir" \
     --egs.opts="--frames-overlap-per-eg 0" \
     --cleanup.remove-egs=$remove_egs \
@@ -207,18 +201,20 @@ if [ $stage -le 6 ]; then
   # as long as phones.txt was compatible.
 
   utils/mkgraph.sh \
-    --self-loop-scale 1.0 data/$lang_test \
+    --self-loop-scale 1.0 $lang_decode \
     $dir $dir/graph || exit 1;
 fi
 
 if [ $stage -le 7 ]; then
   frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
   steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-    --extra-left-context $chunk_left_context \
-    --extra-right-context $chunk_right_context \
-    --extra-left-context-initial 0 \
-    --extra-right-context-final 0 \
     --frames-per-chunk $frames_per_chunk \
     --nj $nj --cmd "$cmd" \
     $dir/graph data/test $dir/decode_test || exit 1;
+
+  steps/lmrescore_const_arpa.sh --cmd "$cmd" $lang_decode $lang_rescore \
+                                data/test $dir/decode_test{,_rescored} || exit 1
 fi
+
+echo "Done. Date: $(date). Results:"
+local/chain/compare_wer.sh $dir

egs/madcat_ar/v1/local/chain/tuning/run_cnn_chainali_1a.sh

@@ -18,18 +18,15 @@ lats_affix=
 # chain options
 train_stage=-10
 xent_regularize=0.1
-frame_subsampling_factor=4
 # training chunk-options
 chunk_width=340,300,200,100
 num_leaves=500
-# we don't need extra left/right context for TDNN systems.
-chunk_left_context=0
-chunk_right_context=0
 tdnn_dim=450
 # training options
 srand=0
 remove_egs=false
-lang_test=lang_test
+lang_decode=data/lang
+lang_rescore=data/lang_rescore_6g
 # End configuration section.
 echo "$0 $@"  # Print the command line for logging
 
@@ -170,13 +167,13 @@ if [ $stage -le 5 ]; then
     --chain.leaky-hmm-coefficient=0.1 \
     --chain.l2-regularize=0.00005 \
     --chain.apply-deriv-weights=false \
-    --chain.lm-opts="--num-extra-lm-states=500" \
-    --chain.frame-subsampling-factor=$frame_subsampling_factor \
+    --chain.lm-opts="--ngram-order=2 --no-prune-ngram-order=1 --num-extra-lm-states=1000" \
+    --chain.frame-subsampling-factor=4 \
     --chain.alignment-subsampling-factor=1 \
     --trainer.srand=$srand \
     --trainer.max-param-change=2.0 \
     --trainer.num-epochs=4 \
-    --trainer.frames-per-iter=1000000 \
+    --trainer.frames-per-iter=2000000 \
     --trainer.optimization.num-jobs-initial=3 \
     --trainer.optimization.num-jobs-final=16 \
     --trainer.optimization.initial-effective-lrate=0.001 \
@@ -185,10 +182,6 @@ if [ $stage -le 5 ]; then
     --trainer.num-chunk-per-minibatch=64,32 \
     --trainer.optimization.momentum=0.0 \
     --egs.chunk-width=$chunk_width \
-    --egs.chunk-left-context=$chunk_left_context \
-    --egs.chunk-right-context=$chunk_right_context \
-    --egs.chunk-left-context-initial=0 \
-    --egs.chunk-right-context-final=0 \
     --egs.dir="$common_egs_dir" \
     --egs.opts="--frames-overlap-per-eg 0" \
     --cleanup.remove-egs=$remove_egs \
@@ -209,18 +202,20 @@ if [ $stage -le 6 ]; then
   # as long as phones.txt was compatible.
 
   utils/mkgraph.sh \
-    --self-loop-scale 1.0 data/$lang_test \
+    --self-loop-scale 1.0 $lang_decode \
     $dir $dir/graph || exit 1;
 fi
 
 if [ $stage -le 7 ]; then
   frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
   steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-    --extra-left-context $chunk_left_context \
-    --extra-right-context $chunk_right_context \
-    --extra-left-context-initial 0 \
-    --extra-right-context-final 0 \
     --frames-per-chunk $frames_per_chunk \
     --nj $nj --cmd "$cmd" \
     $dir/graph data/test $dir/decode_test || exit 1;
+
+  steps/lmrescore_const_arpa.sh --cmd "$cmd" $lang_decode $lang_rescore \
+                                data/test $dir/decode_test{,_rescored} || exit 1
 fi
+
+echo "Done. Date: $(date). Results:"
+local/chain/compare_wer.sh $dir

egs/madcat_ar/v1/local/chain/tuning/run_cnn_e2eali_1a.sh

@@ -19,17 +19,14 @@ reporting_email=
 train_stage=-10
 xent_regularize=0.1
 frame_subsampling_factor=4
-# training chunk-options
 chunk_width=340,300,200,100
 num_leaves=500
-# we don't need extra left/right context for TDNN systems.
-chunk_left_context=0
-chunk_right_context=0
 tdnn_dim=450
 # training options
 srand=0
 remove_egs=true
-lang_test=lang_test
+lang_decode=data/lang
+lang_rescore=data/lang_rescore_6g
 # End configuration section.
 echo "$0 $@"  # Print the command line for logging
 
@@ -171,28 +168,24 @@ if [ $stage -le 5 ]; then
     --chain.leaky-hmm-coefficient=0.1 \
     --chain.l2-regularize=0.00005 \
     --chain.apply-deriv-weights=false \
-    --chain.lm-opts="--num-extra-lm-states=500" \
+    --chain.lm-opts="--ngram-order=2 --no-prune-ngram-order=1 --num-extra-lm-states=1000" \
     --chain.frame-subsampling-factor=$frame_subsampling_factor \
     --chain.alignment-subsampling-factor=1 \
     --chain.left-tolerance 3 \
     --chain.right-tolerance 3 \
     --trainer.srand=$srand \
     --trainer.max-param-change=2.0 \
-    --trainer.num-epochs=2 \
-    --trainer.frames-per-iter=1000000 \
+    --trainer.num-epochs=4 \
+    --trainer.frames-per-iter=2000000 \
     --trainer.optimization.num-jobs-initial=3 \
     --trainer.optimization.num-jobs-final=16 \
     --trainer.optimization.initial-effective-lrate=0.001 \
     --trainer.optimization.final-effective-lrate=0.0001 \
     --trainer.optimization.shrink-value=1.0 \
-    --trainer.num-chunk-per-minibatch=96,64 \
+    --trainer.num-chunk-per-minibatch=64,32 \
     --trainer.optimization.momentum=0.0 \
     --trainer.add-option="--optimization.memory-compression-level=2" \
     --egs.chunk-width=$chunk_width \
-    --egs.chunk-left-context=$chunk_left_context \
-    --egs.chunk-right-context=$chunk_right_context \
-    --egs.chunk-left-context-initial=0 \
-    --egs.chunk-right-context-final=0 \
     --egs.dir="$common_egs_dir" \
     --egs.opts="--frames-overlap-per-eg 0" \
     --cleanup.remove-egs=$remove_egs \
@@ -213,18 +206,20 @@ if [ $stage -le 6 ]; then
   # as long as phones.txt was compatible.
 
   utils/mkgraph.sh \
-    --self-loop-scale 1.0 data/$lang_test \
+    --self-loop-scale 1.0 $lang_decode \
     $dir $dir/graph || exit 1;
 fi
 
 if [ $stage -le 7 ]; then
   frames_per_chunk=$(echo $chunk_width | cut -d, -f1)
   steps/nnet3/decode.sh --acwt 1.0 --post-decode-acwt 10.0 \
-    --extra-left-context $chunk_left_context \
-    --extra-right-context $chunk_right_context \
-    --extra-left-context-initial 0 \
-    --extra-right-context-final 0 \
     --frames-per-chunk $frames_per_chunk \
     --nj $nj --cmd "$cmd" \
     $dir/graph data/test $dir/decode_test || exit 1;
+
+  steps/lmrescore_const_arpa.sh --cmd "$cmd" $lang_decode $lang_rescore \
+                                data/test $dir/decode_test{,_rescored} || exit 1
 fi
+
+echo "Done. Date: $(date). Results:"
+local/chain/compare_wer.sh $dir