changed the definition of deriv-truncate-margin option. Now the margi…

egs/ami/s5/local/chain/run_blstm_ami_5.sh

@@ -128,7 +128,7 @@ if [ $stage -le 17 ]; then
     --trainer.optimization.initial-effective-lrate 0.001 \
     --trainer.optimization.final-effective-lrate 0.0001 \
     --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 10 \
+    --trainer.deriv-truncate-margin 8 \
     --egs.stage $get_egs_stage \
     --egs.opts "--frames-overlap-per-eg 0" \
     --egs.chunk-width $chunk_width \

egs/aspire/s5/local/chain/tuning/run_blstm_7b.sh

@@ -192,7 +192,7 @@ if [ $stage -le 12 ]; then
     --trainer.optimization.final-effective-lrate 0.0001 \
     --trainer.optimization.shrink-value 0.99 \
     --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 10 \
+    --trainer.deriv-truncate-margin 8 \
     --cleanup.remove-egs $remove_egs \
     --feat-dir data/train_rvb_min${min_seg_len}_hires \
     --tree-dir $treedir \

egs/aspire/s5/local/chain/tuning/run_blstm_asp_1.sh

@@ -187,7 +187,7 @@ if [ $stage -le 12 ]; then
     --trainer.optimization.final-effective-lrate 0.0001 \
     --trainer.optimization.shrink-value 0.99 \
     --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 10 \
+    --trainer.deriv-truncate-margin 8 \
     --cleanup.remove-egs $remove_egs \
     --feat-dir data/train_rvb_min${min_seg_len}_hires \
     --tree-dir $treedir \

egs/fisher_swbd/s5/local/chain/run_blstm_6h.sh

@@ -127,7 +127,7 @@ if [ $stage -le 13 ]; then
     --trainer.optimization.initial-effective-lrate 0.001 \
     --trainer.optimization.final-effective-lrate 0.0001 \
     --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 10 \
+    --trainer.deriv-truncate-margin 8 \
     --egs.stage $get_egs_stage \
     --egs.opts "--frames-overlap-per-eg 0" \
     --egs.chunk-width $chunk_width \

egs/multi_en/s5/local/chain/run_blstm_6h.sh

@@ -134,7 +134,7 @@ if [ $stage -le 13 ]; then
     --trainer.optimization.initial-effective-lrate 0.001 \
     --trainer.optimization.final-effective-lrate 0.0001 \
     --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 10 \
+    --trainer.deriv-truncate-margin 8 \
     --egs.stage $get_egs_stage \
     --egs.opts "--frames-overlap-per-eg 0" \
     --egs.chunk-width $chunk_width \

egs/swbd/s5c/local/chain/tuning/run_blstm_6h.sh

@@ -154,7 +154,7 @@ if [ $stage -le 13 ]; then
     --trainer.optimization.initial-effective-lrate 0.001 \
     --trainer.optimization.final-effective-lrate 0.0001 \
     --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 10 \
+    --trainer.deriv-truncate-margin 8 \
     --egs.stage $get_egs_stage \
     --egs.opts "--frames-overlap-per-eg 0" \
     --egs.chunk-width $chunk_width \

egs/swbd/s5c/local/chain/tuning/run_blstm_6i.sh

@@ -160,7 +160,7 @@ if [ $stage -le 13 ]; then
     --trainer.optimization.initial-effective-lrate 0.001 \
     --trainer.optimization.final-effective-lrate 0.0001 \
     --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 10 \
+    --trainer.deriv-truncate-margin 8 \
     --egs.stage $get_egs_stage \
     --egs.opts "--frames-overlap-per-eg 0" \
     --egs.chunk-width $chunk_width \

egs/swbd/s5c/local/chain/tuning/run_blstm_6j.sh

@@ -185,7 +185,7 @@ if [ $stage -le 13 ]; then
     --trainer.optimization.initial-effective-lrate 0.001 \
     --trainer.optimization.final-effective-lrate 0.0001 \
     --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 10 \
+    --trainer.deriv-truncate-margin 8 \
     --egs.stage $get_egs_stage \
     --egs.opts "--frames-overlap-per-eg 0" \
     --egs.chunk-width $chunk_width \

egs/swbd/s5c/local/chain/tuning/run_lstm_6h.sh

@@ -159,7 +159,7 @@ if [ $stage -le 13 ]; then
     --trainer.optimization.initial-effective-lrate 0.001 \
     --trainer.optimization.final-effective-lrate 0.0001 \
     --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 10 \
+    --trainer.deriv-truncate-margin 8 \
     --egs.stage $get_egs_stage \
     --egs.opts "--frames-overlap-per-eg 0" \
     --egs.chunk-width $chunk_width \

egs/swbd/s5c/local/chain/tuning/run_lstm_6i.sh

@@ -159,7 +159,7 @@ if [ $stage -le 13 ]; then
     --trainer.optimization.initial-effective-lrate 0.001 \
     --trainer.optimization.final-effective-lrate 0.0001 \
     --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 10 \
+    --trainer.deriv-truncate-margin 8 \
     --egs.stage $get_egs_stage \
     --egs.opts "--frames-overlap-per-eg 0" \
     --egs.chunk-width $chunk_width \

egs/swbd/s5c/local/chain/tuning/run_lstm_6j.sh

@@ -181,7 +181,7 @@ if [ $stage -le 13 ]; then
     --trainer.optimization.initial-effective-lrate 0.001 \
     --trainer.optimization.final-effective-lrate 0.0001 \
     --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 10 \
+    --trainer.deriv-truncate-margin 8 \
     --egs.stage $get_egs_stage \
     --egs.opts "--frames-overlap-per-eg 0" \
     --egs.chunk-width $chunk_width \

egs/swbd/s5c/local/chain/tuning/run_lstm_d.sh

@@ -164,7 +164,7 @@ if [ $stage -le 13 ]; then
     --trainer.optimization.initial-effective-lrate 0.001 \
     --trainer.optimization.final-effective-lrate 0.0001 \
     --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 10 \
+    --trainer.deriv-truncate-margin 8 \
     --egs.stage $get_egs_stage \
     --egs.opts="--frames-overlap-per-eg 0" \
     --egs.chunk-width $chunk_width \

egs/swbd/s5c/local/chain/tuning/run_tdnn_lstm_1a.sh

@@ -188,7 +188,7 @@ if [ $stage -le 13 ]; then
     --trainer.optimization.initial-effective-lrate 0.001 \
     --trainer.optimization.final-effective-lrate 0.0001 \
     --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 10 \
+    --trainer.deriv-truncate-margin 8 \
     --egs.stage $get_egs_stage \
     --egs.opts "--frames-overlap-per-eg 0" \
     --egs.chunk-width $chunk_width \

egs/swbd/s5c/local/chain/tuning/run_tdnn_lstm_1b.sh

@@ -184,7 +184,7 @@ if [ $stage -le 13 ]; then
     --trainer.optimization.initial-effective-lrate 0.001 \
     --trainer.optimization.final-effective-lrate 0.0001 \
     --trainer.optimization.momentum 0.0 \
-    --trainer.deriv-truncate-margin 10 \
+    --trainer.deriv-truncate-margin 8 \
     --egs.stage $get_egs_stage \
     --egs.opts "--frames-overlap-per-eg 0" \
     --egs.chunk-width $chunk_width \

egs/wsj/s5/steps/nnet3/chain/train.py

@@ -222,10 +222,12 @@ def GetArgs():
                         help="Number of sequences to be processed in parallel every minibatch" )
     parser.add_argument("--trainer.deriv-truncate-margin", type=int, dest='deriv_truncate_margin',
                         default = None,
-                        help="If specified, it is the number of frames that the derivative will be backpropagated through the chunk boundaries, "
-                        "e.g., During BLSTM model training if the chunk-width=150 and deriv-truncate-margin=5, then the derivative will be "
-                        "backpropagated up to t=-5 and t=154 in the forward and backward LSTM sequence respectively; "
-                        "otherwise, the derivative will be backpropagated to the end of the sequence.")
+                        help="(Relevant only for recurrent models). If specified, gives the margin "
+                        "(in input frames) around the 'required' part of each chunk that the "
+                        "derivatives are backpropagated to. If unset, the derivatives are "
+                        "backpropagated all the way to the boundaries of the input data. E.g. 8 is "
+                        "a reasonable setting. Note: the 'required' part of the chunk is defined by "
+                        "the model's {left,right}-context.")
 
     # General options
     parser.add_argument("--stage", type=int, default=-4,
@@ -675,8 +677,8 @@ def Train(args, run_opts):
     min_deriv_time = None
     max_deriv_time = None
     if not args.deriv_truncate_margin is None:
-        min_deriv_time = -args.deriv_truncate_margin
-        max_deriv_time = args.chunk_width - 1 + args.deriv_truncate_margin
+        min_deriv_time = -args.deriv_truncate_margin - model_left_context
+        max_deriv_time = args.chunk_width - 1 + args.deriv_truncate_margin + model_right_context
 
     logger.info("Training will run for {0} epochs = {1} iterations".format(args.num_epochs, num_iters))
     for iter in range(num_iters):