clean up scripts in #2572

[xiaohui-zhang]

No description provided.

[yangxueruivs]

Hi, I noticed that you didn't add a recipe for librispeech, is it because it won't give any improvement in that dataset? I tried myself and found it's not as good as pure tdnn-f.

[danpovey]

We didn't run it on Librispeech. When you say you ran it on librispeech, can you be more specific about the configuration you ran? Because we'd tend to use slightly larger models for librispeech, and other settings like num-epochs may be different there too.

[yangxueruivs]

@danpovey Sorry, the configurations are:

input dim=100 name=ivector
input dim=40 name=input

# MFCC to filterbank
idct-layer name=idct input=input dim=40 cepstral-lifter=22 affine-transform-file=$dir/configs/idct.mat
linear-component name=ivector-linear $ivector_affine_opts dim=200 input=ReplaceIndex(ivector, t, 0)
batchnorm-component name=ivector-batchnorm target-rms=0.025
batchnorm-component name=idct-batchnorm input=idct

combine-feature-maps-layer name=combine_inputs input=Append(idct-batchnorm, ivector-batchnorm) num-filters1=1 num-filters2=5 height=40
conv-relu-batchnorm-layer name=cnn1 $cnn_opts height-in=40 height-out=40 time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=64
conv-relu-batchnorm-layer name=cnn2 $cnn_opts height-in=40 height-out=40 time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=64
conv-relu-batchnorm-layer name=cnn3 $cnn_opts height-in=40 height-out=20 height-subsample-out=2 time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=128
conv-relu-batchnorm-layer name=cnn4 $cnn_opts height-in=20 height-out=20 time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=128
conv-relu-batchnorm-layer name=cnn5 $cnn_opts height-in=20 height-out=10 height-subsample-out=2 time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=256
conv-relu-batchnorm-layer name=cnn6 $cnn_opts height-in=10 height-out=10 time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=256

# the first TDNN-F layer has no bypass
tdnnf-layer name=tdnnf7 $tdnnf_first_opts dim=1536 bottleneck-dim=256 time-stride=0
tdnnf-layer name=tdnnf8 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf9 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf10 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf11 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf12 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf13 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf14 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf15 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf16 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf17 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf18 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
tdnnf-layer name=tdnnf19 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3
linear-component name=prefinal-l dim=256 $linear_opts

prefinal-layer name=prefinal-chain input=prefinal-l $prefinal_opts big-dim=1536 small-dim=256
output-layer name=output include-log-softmax=false dim=$num_targets $output_opts

prefinal-layer name=prefinal-xent input=prefinal-l $prefinal_opts big-dim=1536 small-dim=256
output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor $output_opts

different tdnnf layers are tried(17/18/19), bottleneck dims(256/160). I set the epochs always equals to 4 so that I could compare the loss with pure tdnnf.

And the relative loss is like(first column is tdnnf, the others are cnn-tdnnf):

2.95 | 3.16 | 3.18
2.94 | 3.14 | 3.18
2.93 | 3.13 | 3.17
2.93 | 3.13 | 3.16
2.94 | 3.12 | 3.14
2.92 | 3.12 | 3.13
2.92 | 3.12 | 3.07
2.89 | 3.11 | 3.1
2.91 | 3.12 | 3.1
2.92 | 3.12 | 3.1
2.89 | 3.12 | 3.08
2.9 | 3.09 | 3.08
2.86 | 3.08 | 3.06
2.88 | 3.07 | 3.05
2.85 | 3.07 | 3.05
2.9 | 3.09 | 3.06
2.88 | 3.05 | 3.04
2.85 | 3.08 | 3.06
2.85 | 3.05 | 3.03
2.83 | 3.03 | 3.07
2.83 | 3.05 | 3.02

[danpovey]

Don't focus on that one test condition. The compare_wer.sh script compares about 16 different test conditions. Is there degradation consistently across those-- or at least, on average? The configuration seems reasonable.

…

On Tue, Sep 18, 2018 at 9:58 PM YangXuerui ***@***.***> wrote: @danpovey <https://github.com/danpovey> Sorry, the configurations are: input dim=100 name=ivector input dim=40 name=input # MFCC to filterbank idct-layer name=idct input=input dim=40 cepstral-lifter=22 affine-transform-file=$dir/configs/idct.mat linear-component name=ivector-linear $ivector_affine_opts dim=200 input=ReplaceIndex(ivector, t, 0) batchnorm-component name=ivector-batchnorm target-rms=0.025 batchnorm-component name=idct-batchnorm input=idct combine-feature-maps-layer name=combine_inputs input=Append(idct-batchnorm, ivector-batchnorm) num-filters1=1 num-filters2=5 height=40 conv-relu-batchnorm-layer name=cnn1 $cnn_opts height-in=40 height-out=40 time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=64 conv-relu-batchnorm-layer name=cnn2 $cnn_opts height-in=40 height-out=40 time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=64 conv-relu-batchnorm-layer name=cnn3 $cnn_opts height-in=40 height-out=20 height-subsample-out=2 time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=128 conv-relu-batchnorm-layer name=cnn4 $cnn_opts height-in=20 height-out=20 time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=128 conv-relu-batchnorm-layer name=cnn5 $cnn_opts height-in=20 height-out=10 height-subsample-out=2 time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=256 conv-relu-batchnorm-layer name=cnn6 $cnn_opts height-in=10 height-out=10 time-offsets=-1,0,1 height-offsets=-1,0,1 num-filters-out=256 # the first TDNN-F layer has no bypass tdnnf-layer name=tdnnf7 $tdnnf_first_opts dim=1536 bottleneck-dim=256 time-stride=0 tdnnf-layer name=tdnnf8 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3 tdnnf-layer name=tdnnf9 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3 tdnnf-layer name=tdnnf10 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3 tdnnf-layer name=tdnnf11 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3 tdnnf-layer name=tdnnf12 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3 tdnnf-layer name=tdnnf13 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3 tdnnf-layer name=tdnnf14 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3 tdnnf-layer name=tdnnf15 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3 tdnnf-layer name=tdnnf16 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3 tdnnf-layer name=tdnnf17 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3 tdnnf-layer name=tdnnf18 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3 tdnnf-layer name=tdnnf19 $tdnnf_opts dim=1536 bottleneck-dim=160 time-stride=3 linear-component name=prefinal-l dim=256 $linear_opts prefinal-layer name=prefinal-chain input=prefinal-l $prefinal_opts big-dim=1536 small-dim=256 output-layer name=output include-log-softmax=false dim=$num_targets $output_opts prefinal-layer name=prefinal-xent input=prefinal-l $prefinal_opts big-dim=1536 small-dim=256 output-layer name=output-xent dim=$num_targets learning-rate-factor=$learning_rate_factor $output_opts different tdnnf layers are tried(17/18/19), bottleneck dims(256/160). I set the epochs always equals to 4 so that I could compare the loss with pure tdnnf. And the relative loss is like(first column is tdnnf, the others are cnn-tdnnf): 2.95 | 3.16 | 3.18 2.94 | 3.14 | 3.18 2.93 | 3.13 | 3.17 2.93 | 3.13 | 3.16 2.94 | 3.12 | 3.14 2.92 | 3.12 | 3.13 2.92 | 3.12 | 3.07 2.89 | 3.11 | 3.1 2.91 | 3.12 | 3.1 2.92 | 3.12 | 3.1 2.89 | 3.12 | 3.08 2.9 | 3.09 | 3.08 2.86 | 3.08 | 3.06 2.88 | 3.07 | 3.05 2.85 | 3.07 | 3.05 2.9 | 3.09 | 3.06 2.88 | 3.05 | 3.04 2.85 | 3.08 | 3.06 2.85 | 3.05 | 3.03 2.83 | 3.03 | 3.07 2.83 | 3.05 | 3.02 — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub <#2643 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVuxodF692KB4T-e2T2Irpw5wasGuSks5ucaS8gaJpZM4WKr0E> .