add reverberated TDNN+LSTM recipe on AMI; a fix to aspire recipe#1314
Merged
vijayaditya merged 3 commits intokaldi-asr:masterfrom Jan 7, 2017
Merged
add reverberated TDNN+LSTM recipe on AMI; a fix to aspire recipe#1314vijayaditya merged 3 commits intokaldi-asr:masterfrom
vijayaditya merged 3 commits intokaldi-asr:masterfrom
Conversation
Contributor
Contributor
|
Will do.
Vijay
…On Jan 3, 2017 6:58 PM, "Daniel Povey" ***@***.***> wrote:
@vijayaditya, do you have time to review this?
On Tue, Jan 3, 2017 at 6:36 PM, Tom Ko ***@***.***> wrote:
> ------------------------------
> You can view, comment on, or merge this pull request online at:
>
> #1314
> Commit Summary
>
> - add reverberated TDNN+LSTM recipe on AMI; a fix to aspire recipe
>
> File Changes
>
> - *M* egs/ami/s5b/RESULTS_ihm
> <https://github.com/kaldi-asr/kaldi/pull/1314/files#diff-0> (9)
> - *M* egs/ami/s5b/RESULTS_mdm
> <https://github.com/kaldi-asr/kaldi/pull/1314/files#diff-1> (9)
> - *M* egs/ami/s5b/RESULTS_sdm
> <https://github.com/kaldi-asr/kaldi/pull/1314/files#diff-2> (14)
> - *M* egs/ami/s5b/local/chain/multi_condition/run_tdnn.sh
> <https://github.com/kaldi-asr/kaldi/pull/1314/files#diff-3> (6)
> - *A* egs/ami/s5b/local/chain/multi_condition/run_tdnn_lstm.sh
> <https://github.com/kaldi-asr/kaldi/pull/1314/files#diff-4> (334)
> - *M* egs/aspire/s5/local/chain/tuning/run_tdnn_7b.sh
> <https://github.com/kaldi-asr/kaldi/pull/1314/files#diff-5> (5)
>
> Patch Links:
>
> - https://github.com/kaldi-asr/kaldi/pull/1314.patch
> - https://github.com/kaldi-asr/kaldi/pull/1314.diff
>
> —
> You are receiving this because you are subscribed to this thread.
> Reply to this email directly, view it on GitHub
> <#1314>, or mute the thread
> <https://github.com/notifications/unsubscribe-auth/
ADJVu9wd3IHuy0D32zMOQ8f6xADi7fu3ks5rOwWvgaJpZM4LaPIs>
> .
>
—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
<#1314 (comment)>, or mute
the thread
<https://github.com/notifications/unsubscribe-auth/ADtwoPWKpp5Ea-6fbHyxtbu6XcxZ7UWzks5rOwrQgaJpZM4LaPIs>
.
|
vijayaditya
requested changes
Jan 4, 2017
| # local/chain/multi_condition/run_tdnn_lstm.sh --mic mdm8 --use-ihm-ali true --train-set train_cleaned --gmm tri3_cleaned | ||
| # cleanup + chain TDNN+LSTM model, MDM original + IHM reverberated data, alignments from IHM data | ||
| # *** best system *** | ||
| %WER 31.8 | 14488 94497 | 71.8 15.4 12.8 3.5 31.8 62.7 | 0.698 | exp/mdm8/chain_cleaned_rvb/tdnn_lstm1i_sp_rvb_bi_ihmali/decode_dev/ascore_10/dev_hires_o4.ctm.filt.sys |
Contributor
There was a problem hiding this comment.
@tomkocse Very impressive results. Please add the version number of your script which generates these results (eg. run_tdnn_lstm_1a.sh ) as we might replace run_tdnn_lstm.sh with a better tdnn_lstm architecture in the future.
Move this script to tuning directory as run_tdnn_lstm_1b.sh and softlink it from run_tdnn_lstm.sh
egs/ami/s5b/RESULTS_sdm
Outdated
| %WER 40.9 | 13807 89961 | 62.4 20.0 17.6 3.3 40.9 65.7 | 0.612 | exp/sdm1/chain_cleaned/tdnn_lstm1i_sp_bi_ihmali_ld5/decode_eval/ascore_10/eval_hires_o4.ctm.filt.sys | ||
|
|
||
|
|
||
| # local/chain/multi_condition/run_tdnn_lstm.sh --mic sdm1 --use-ihm-ali true --train-set train_cleaned --gmm tri3_cleaned |
Contributor
There was a problem hiding this comment.
same comment as above; start versioning your run*.sh scripts.
| @@ -0,0 +1,334 @@ | |||
| #!/bin/bash | |||
|
|
|||
| # This is a chain-training script with TDNN+LSTM neural networks. | |||
Contributor
There was a problem hiding this comment.
Not checking this in detail as I assume it is similar to multi_condition/run_tdnn.sh except for the xconfig and some lstm specific parameters.
|
|
||
| # check steps/libs/nnet3/xconfig/lstm.py for the other options and defaults | ||
| lstmp-layer name=lstm1 cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3 | ||
| relu-renorm-layer name=tdnn4 input=Append(-3,0,3) dim=1024 |
Contributor
There was a problem hiding this comment.
Try increasing the parameters further as you have more data than the SDM/MDM only case. Though we might not ultimately recommend a large parameter model it is better to know the least possible WER with this architecture and simulated multi_condition data. You can try increasing the relu-norm-layer dim in increments of 256 or 128.
| --egs.chunk-right-context $chunk_right_context \ | ||
| --trainer.num-chunk-per-minibatch 64 \ | ||
| --trainer.frames-per-iter 1500000 \ | ||
| --trainer.num-epochs 4 \ |
Contributor
There was a problem hiding this comment.
Check if you can increase the number of epochs.
| data/train data/train_temp_for_lats | ||
| utils/data/combine_short_segments.sh \ | ||
| data/train_temp_for_lats $min_seg_len data/train_min${min_seg_len} | ||
| steps/compute_cmvn_stats.sh data/train_min${min_seg_len} exp/make_mfcc/train_min${min_seg_len} mfcc || exit 1; |
Contributor
There was a problem hiding this comment.
I recommend using the new style of not specifying logdirs.
Contributor
|
Vijay, wouldn't the training be rather slow with more epochs? IIRC he's
already doubled the amount of data versus the regular 3x-perturbed data, so
that affects the meaning of epochs.
…On Tue, Jan 3, 2017 at 7:39 PM, Vijayaditya Peddinti < ***@***.***> wrote:
***@***.**** requested changes on this pull request.
Could you also move softlink run_tdnn_lstm.sh in the chain/ directory to
tdnn_lstm_1i.sh script.
------------------------------
In egs/ami/s5b/RESULTS_mdm
<#1314 (review)>:
> %WER 34.6 | 15116 94508 | 69.6 17.6 12.9 4.1 34.6 62.3 | 0.687 | exp/mdm8/chain_cleaned/tdnn_lstm1i_sp_bi_ihmali_ld5/decode_dev/ascore_9/dev_hires_o4.ctm.filt.sys
%WER 37.1 | 14343 90002 | 66.3 18.8 14.9 3.4 37.1 62.3 | 0.659 | exp/mdm8/chain_cleaned/tdnn_lstm1i_sp_bi_ihmali_ld5/decode_eval/ascore_9/eval_hires_o4.ctm.filt.sys
+
+
+# local/chain/multi_condition/run_tdnn_lstm.sh --mic mdm8 --use-ihm-ali true --train-set train_cleaned --gmm tri3_cleaned
+# cleanup + chain TDNN+LSTM model, MDM original + IHM reverberated data, alignments from IHM data
+# *** best system ***
+%WER 31.8 | 14488 94497 | 71.8 15.4 12.8 3.5 31.8 62.7 | 0.698 | exp/mdm8/chain_cleaned_rvb/tdnn_lstm1i_sp_rvb_bi_ihmali/decode_dev/ascore_10/dev_hires_o4.ctm.filt.sys
@tomkocse <https://github.com/tomkocse> Very impressive results. Please
add the version number of your script which generates these results (eg.
run_tdnn_lstm_1a.sh ) as we might replace run_tdnn_lstm.sh with a better
tdnn_lstm architecture in the future.
Move this script to tuning directory as run_tdnn_lstm_1b.sh and softlink
it from run_tdnn_lstm.sh
------------------------------
In egs/ami/s5b/RESULTS_sdm
<#1314 (review)>:
> %WER 37.6 | 15122 94495 | 66.1 18.7 15.1 3.7 37.6 63.2 | 0.646 | exp/sdm1/chain_cleaned/tdnn_lstm1i_sp_bi_ihmali_ld5/decode_dev/ascore_10/dev_hires_o4.ctm.filt.sys
%WER 40.9 | 13807 89961 | 62.4 20.0 17.6 3.3 40.9 65.7 | 0.612 | exp/sdm1/chain_cleaned/tdnn_lstm1i_sp_bi_ihmali_ld5/decode_eval/ascore_10/eval_hires_o4.ctm.filt.sys
+
+
+# local/chain/multi_condition/run_tdnn_lstm.sh --mic sdm1 --use-ihm-ali true --train-set train_cleaned --gmm tri3_cleaned
same comment as above; start versioning your run*.sh scripts.
------------------------------
In egs/ami/s5b/local/chain/multi_condition/run_tdnn_lstm.sh
<#1314 (review)>:
> @@ -0,0 +1,334 @@
+#!/bin/bash
+
+# This is a chain-training script with TDNN+LSTM neural networks.
Not checking this in detail as I assume it is similar to
multi_condition/run_tdnn.sh except for the xconfig and some lstm specific
parameters.
------------------------------
In egs/ami/s5b/local/chain/multi_condition/run_tdnn_lstm.sh
<#1314 (review)>:
> + 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(-1,0,1,ReplaceIndex(ivector, t, 0)) affine-transform-file=$dir/configs/lda.mat
+
+ # the first splicing is moved before the lda layer, so no splicing here
+ relu-renorm-layer name=tdnn1 dim=1024
+ relu-renorm-layer name=tdnn2 input=Append(-1,0,1) dim=1024
+ relu-renorm-layer name=tdnn3 input=Append(-1,0,1) dim=1024
+
+ # check steps/libs/nnet3/xconfig/lstm.py for the other options and defaults
+ lstmp-layer name=lstm1 cell-dim=1024 recurrent-projection-dim=256 non-recurrent-projection-dim=256 delay=-3
+ relu-renorm-layer name=tdnn4 input=Append(-3,0,3) dim=1024
Try increasing the parameters further as you have more data than the
SDM/MDM only case. Though we might not ultimately recommend a large
parameter model it is better to know the least possible WER with this
architecture and simulated multi_condition data. You can try increasing the
relu-norm-layer dim in increments of 256 or 128.
------------------------------
In egs/ami/s5b/local/chain/multi_condition/run_tdnn_lstm.sh
<#1314 (review)>:
> + --cmd "$decode_cmd" \
+ --feat.online-ivector-dir $train_ivector_dir \
+ --feat.cmvn-opts "--norm-means=false --norm-vars=false" \
+ --chain.xent-regularize $xent_regularize \
+ --chain.leaky-hmm-coefficient 0.1 \
+ --chain.l2-regularize 0.00005 \
+ --chain.apply-deriv-weights false \
+ --chain.lm-opts="--num-extra-lm-states=2000" \
+ --egs.dir "$common_egs_dir" \
+ --egs.opts "--frames-overlap-per-eg 0" \
+ --egs.chunk-width $chunk_width \
+ --egs.chunk-left-context $chunk_left_context \
+ --egs.chunk-right-context $chunk_right_context \
+ --trainer.num-chunk-per-minibatch 64 \
+ --trainer.frames-per-iter 1500000 \
+ --trainer.num-epochs 4 \
Check if you can increase the number of epochs.
------------------------------
In egs/aspire/s5/local/chain/tuning/run_tdnn_7b.sh
<#1314 (review)>:
> @@ -100,6 +100,7 @@ if [ $stage -le 9 ]; then
data/train data/train_temp_for_lats
utils/data/combine_short_segments.sh \
data/train_temp_for_lats $min_seg_len data/train_min${min_seg_len}
+ steps/compute_cmvn_stats.sh data/train_min${min_seg_len} exp/make_mfcc/train_min${min_seg_len} mfcc || exit 1;
I recommend using the new style of not specifying logdirs.
—
You are receiving this because you commented.
Reply to this email directly, view it on GitHub
<#1314 (review)>,
or mute the thread
<https://github.com/notifications/unsubscribe-auth/ADJVuy5A7sJGhLt89e5TciPGf8LcJAG1ks5rOxRXgaJpZM4LaPIs>
.
|
Contributor
vijayaditya
approved these changes
Jan 7, 2017
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.