How to use fbank feature as input

[yangxueruivs]

In YAML file, I set fbank-config="path-to-fbank.conf", but it doesnt work. Is it because gstreamer doesnt support extraction of fbank?

[yangxueruivs]

Now I set feature-type=fbank in YAML, and got this:

ERROR ([5.5.0-17cd]:OnlineTransform():online-feature.cc:436) Dimension mismatch: source features have dimension 539 and LDA #cols is 281

[ Stack-Trace: ] kaldi::MessageLogger::HandleMessage(kaldi::LogMessageEnvelope const&, char const) kaldi::FatalMessageLogger::~FatalMessageLogger() kaldi::OnlineTransform::OnlineTransform(kaldi::MatrixBase const&, kaldi::OnlineFeatureInterface) kaldi::OnlineIvectorFeature::OnlineIvectorFeature(kaldi::OnlineIvectorExtractionInfo const&, kaldi::OnlineFeatureInterface*) kaldi::OnlineNnet2FeaturePipeline::OnlineNnet2FeaturePipeline(kaldi::OnlineNnet2FeaturePipelineInfo const&)

clone

terminate called after throwing an instance of 'std::runtime_error' what():

[alumae]

Probably the fbank-config that you give to kaldi-gstreamer-server is not compatible with the model.

[yangxueruivs]

Actually I trained this model with 80-dim fbank and 100-dim ivector. And when I set the fbank.conf to be same as what I trained. It still returns this error. I also tried various settings, none of them match.

[yangxueruivs]

When I set "--num-mel-bins=80", the error will be Dimension mismatch: source features have dimension 560 and LDA #cols is 281. And whatever I set in num-mei-bins, it will time it by 7, e.g. ,if --num-mel-bins=60, the error would be "source features have dimension 420"

[alumae]

Are you using pitch features?

[yangxueruivs]

Yes, but when I use mfcc feature, the pitch can be generated correctly.

[alumae]

Please do:

nnet3-am-info $model_dir/final.mdl

and paste the output.

[yangxueruivs]

WARNING (nnet3-am-info[5.5]:Check():nnet-nnet.cc:789) Node relu11.batchnorm is never used to compute any output. WARNING (nnet3-am-info[5.5]:Check():nnet-nnet.cc:789) Node relu11.dropout is never used to compute any output. input-dim: 80 ivector-dim: 100 num-pdfs: 6184 prior-dimension: 0

Nnet info follows.

left-context: 138 right-context: 138 num-parameters: 22726608 modulus: 1 input-node name=ivector dim=100 input-node name=input dim=80 component-node name=ivector-linear component=ivector-linear input=ReplaceIndex(ivector, t, 0) input-dim=100 output-dim=160 component-node name=ivector-batchnorm component=ivector-batchnorm input=ivector-linear input-dim=160 output-dim=160 component-node name=combine_inputs component=combine_inputs input=Append(input, ivector-batchnorm) input-dim=240 output-dim=240 component-node name=cnn1.conv component=cnn1.conv input=combine_inputs input-dim=240 output-dim=2560 component-node name=cnn1.relu component=cnn1.relu input=cnn1.conv input-dim=2560 output-dim=2560 component-node name=cnn1.batchnorm component=cnn1.batchnorm input=cnn1.relu input-dim=2560 output-dim=2560 component-node name=cnn2.conv component=cnn2.conv input=cnn1.batchnorm input-dim=2560 output-dim=2560 component-node name=cnn2.relu component=cnn2.relu input=cnn2.conv input-dim=2560 output-dim=2560 component-node name=cnn2.batchnorm component=cnn2.batchnorm input=cnn2.relu input-dim=2560 output-dim=2560 component-node name=no-op1 component=no-op1 input=Sum(cnn2.batchnorm, cnn1.batchnorm) input-dim=2560 output-dim=2560 component-node name=cnn3.conv component=cnn3.conv input=no-op1 input-dim=2560 output-dim=2560 component-node name=cnn3.relu component=cnn3.relu input=cnn3.conv input-dim=2560 output-dim=2560 component-node name=cnn3.batchnorm component=cnn3.batchnorm input=cnn3.relu input-dim=2560 output-dim=2560 component-node name=cnn4.conv component=cnn4.conv input=cnn3.batchnorm input-dim=2560 output-dim=2560 component-node name=cnn4.relu component=cnn4.relu input=cnn4.conv input-dim=2560 output-dim=2560 component-node name=cnn4.batchnorm component=cnn4.batchnorm input=cnn4.relu input-dim=2560 output-dim=2560 component-node name=no-op2 component=no-op2 input=Sum(cnn3.batchnorm, cnn4.batchnorm) input-dim=2560 output-dim=2560 component-node name=cnn5.conv component=cnn5.conv input=no-op2 input-dim=2560 output-dim=2560 component-node name=cnn5.relu component=cnn5.relu input=cnn5.conv input-dim=2560 output-dim=2560 component-node name=cnn5.batchnorm component=cnn5.batchnorm input=cnn5.relu input-dim=2560 output-dim=2560 component-node name=cnn6.conv component=cnn6.conv input=cnn5.batchnorm input-dim=2560 output-dim=2560 component-node name=cnn6.relu component=cnn6.relu input=cnn6.conv input-dim=2560 output-dim=2560 component-node name=cnn6.batchnorm component=cnn6.batchnorm input=cnn6.relu input-dim=2560 output-dim=2560 component-node name=relu1.affine component=relu1.affine input=cnn6.batchnorm input-dim=2560 output-dim=1536 component-node name=relu1.relu component=relu1.relu input=relu1.affine input-dim=1536 output-dim=1536 component-node name=relu1.batchnorm component=relu1.batchnorm input=relu1.relu input-dim=1536 output-dim=1536 component-node name=relu1.dropout component=relu1.dropout input=relu1.batchnorm input-dim=1536 output-dim=1536 component-node name=linear1 component=linear1 input=relu1.dropout input-dim=1536 output-dim=256 component-node name=dfsmn1.element_wise_scale component=dfsmn1.element_wise_scale input=Append(Offset(linear1, -4), Offset(linear1, -2), linear1, Offset(linear1, 2), Offset( linear1, 4)) input-dim=1280 output-dim=1280 component-node name=dfsmn1 component=dfsmn1 input=dfsmn1.element_wise_scale input-dim=1280 output-dim=256 component-node name=relu2.affine component=relu2.affine input=dfsmn1 input-dim=256 output-dim=1536 component-node name=relu2.relu component=relu2.relu input=relu2.affine input-dim=1536 output-dim=1536 component-node name=relu2.batchnorm component=relu2.batchnorm input=relu2.relu input-dim=1536 output-dim=1536 component-node name=relu2.dropout component=relu2.dropout input=relu2.batchnorm input-dim=1536 output-dim=1536 component-node name=linear2 component=linear2 input=relu2.dropout input-dim=1536 output-dim=256 component-node name=dfsmn2.element_wise_scale component=dfsmn2.element_wise_scale input=Append(Offset(linear2, -4), Offset(linear2, -2), linear2, Offset(linear2, 2), Offset( linear2, 4)) input-dim=1280 output-dim=1280 component-node name=dfsmn2 component=dfsmn2 input=dfsmn2.element_wise_scale input-dim=1280 output-dim=256 component-node name=relu3.affine component=relu3.affine input=dfsmn2 input-dim=256 output-dim=1536 component-node name=relu3.relu component=relu3.relu input=relu3.affine input-dim=1536 output-dim=1536 component-node name=relu3.batchnorm component=relu3.batchnorm input=relu3.relu input-dim=1536 output-dim=1536 component-node name=relu3.dropout component=relu3.dropout input=relu3.batchnorm input-dim=1536 output-dim=1536 component-node name=linear3 component=linear3 input=relu3.dropout input-dim=1536 output-dim=256 component-node name=no-op3 component=no-op3 input=Sum(linear3, Scale(0.66, linear1)) input-dim=256 output-dim=256 component-node name=dfsmn3.element_wise_scale component=dfsmn3.element_wise_scale input=Append(Offset(no-op3, -8), Offset(no-op3, -6), Offset(no-op3, -4), Offset(no-op3, -2) , no-op3, Offset(no-op3, 2), Offset(no-op3, 4), Offset(no-op3, 6), Offset(no-op3, 8)) input-dim=2304 output-dim=2304 component-node name=dfsmn3 component=dfsmn3 input=dfsmn3.element_wise_scale input-dim=2304 output-dim=256 component-node name=relu4.affine component=relu4.affine input=dfsmn3 input-dim=256 output-dim=1536 component-node name=relu4.relu component=relu4.relu input=relu4.affine input-dim=1536 output-dim=1536 component-node name=relu4.batchnorm component=relu4.batchnorm input=relu4.relu input-dim=1536 output-dim=1536 component-node name=relu4.dropout component=relu4.dropout input=relu4.batchnorm input-dim=1536 output-dim=1536 component-node name=linear4 component=linear4 input=relu4.dropout input-dim=1536 output-dim=256 component-node name=dfsmn4.element_wise_scale component=dfsmn4.element_wise_scale input=Append(Offset(linear4, -8), Offset(linear4, -6), Offset(linear4, -4), Offset(linear4, -2), linear4, Offset(linear4, 2), Offset(linear4, 4), Offset(linear4, 6), Offset(linear4, 8)) input-dim=2304 output-dim=2304 component-node name=dfsmn4 component=dfsmn4 input=dfsmn4.element_wise_scale input-dim=2304 output-dim=256 component-node name=relu5.affine component=relu5.affine input=dfsmn4 input-dim=256 output-dim=1536 component-node name=relu5.relu component=relu5.relu input=relu5.affine input-dim=1536 output-dim=1536 component-node name=relu5.batchnorm component=relu5.batchnorm input=relu5.relu input-dim=1536 output-dim=1536 component-node name=relu5.dropout component=relu5.dropout input=relu5.batchnorm input-dim=1536 output-dim=1536 component-node name=linear5 component=linear5 input=relu5.dropout input-dim=1536 output-dim=256 component-node name=no-op5 component=no-op5 input=Sum(linear5, Scale(0.66, no-op3)) input-dim=256 output-dim=256 component-node name=dfsmn5.element_wise_scale component=dfsmn5.element_wise_scale input=Append(Offset(no-op5, -12), Offset(no-op5, -10), Offset(no-op5, -8), Offset(no-op5, - 6), Offset(no-op5, -4), Offset(no-op5, -2), no-op5, Offset(no-op5, 2), Offset(no-op5, 4), Offset(no-op5, 6), Offset(no-op5, 8), Offset(no-op5, 10), Offset(no-op5, 12)) input -dim=3328 output-dim=3328 component-node name=dfsmn5 component=dfsmn5 input=dfsmn5.element_wise_scale input-dim=3328 output-dim=256 component-node name=relu6.affine component=relu6.affine input=dfsmn5 input-dim=256 output-dim=1536 component-node name=relu6.relu component=relu6.relu input=relu6.affine input-dim=1536 output-dim=1536 component-node name=relu6.batchnorm component=relu6.batchnorm input=relu6.relu input-dim=1536 output-dim=1536 component-node name=relu6.dropout component=relu6.dropout input=relu6.batchnorm input-dim=1536 output-dim=1536 component-node name=linear6 component=linear6 input=relu6.dropout input-dim=1536 output-dim=256 component-node name=dfsmn6.element_wise_scale component=dfsmn6.element_wise_scale input=Append(Offset(linear6, -12), Offset(linear6, -10), Offset(linear6, -8), Offset(linear 6, -6), Offset(linear6, -4), Offset(linear6, -2), linear6, Offset(linear6, 2), Offset(linear6, 4), Offset(linear6, 6), Offset(linear6, 8), Offset(linear6, 10), Offset(linear 6, 12)) input-dim=3328 output-dim=3328 component-node name=dfsmn6 component=dfsmn6 input=dfsmn6.element_wise_scale input-dim=3328 output-dim=256 component-node name=relu7.affine component=relu7.affine input=dfsmn6 input-dim=256 output-dim=1536 component-node name=relu7.relu component=relu7.relu input=relu7.affine input-dim=1536 output-dim=1536 component-node name=relu7.batchnorm component=relu7.batchnorm input=relu7.relu input-dim=1536 output-dim=1536 component-node name=relu7.dropout component=relu7.dropout input=relu7.batchnorm input-dim=1536 output-dim=1536 component-node name=linear7 component=linear7 input=relu7.dropout input-dim=1536 output-dim=256 component-node name=no-op7 component=no-op7 input=Sum(linear7, Scale(0.66, no-op5)) input-dim=256 output-dim=256 component-node name=dfsmn7.element_wise_scale component=dfsmn7.element_wise_scale input=Append(Offset(no-op7, -16), Offset(no-op7, -14), Offset(no-op7, -12), Offset(no-op7, -10), Offset(no-op7, -8), Offset(no-op7, -6), Offset(no-op7, -4), Offset(no-op7, -2), no-op7, Offset(no-op7, 2), Offset(no-op7, 4), Offset(no-op7, 6), Offset(no-op7, 8), Off set(no-op7, 10), Offset(no-op7, 12), Offset(no-op7, 14), Offset(no-op7, 16)) input-dim=4352 output-dim=4352 component-node name=dfsmn7 component=dfsmn7 input=dfsmn7.element_wise_scale input-dim=4352 output-dim=256 component-node name=relu8.affine component=relu8.affine input=dfsmn7 input-dim=256 output-dim=1536 component-node name=relu8.relu component=relu8.relu input=relu8.affine input-dim=1536 output-dim=1536 component-node name=relu8.batchnorm component=relu8.batchnorm input=relu8.relu input-dim=1536 output-dim=1536 component-node name=relu8.dropout component=relu8.dropout input=relu8.batchnorm input-dim=1536 output-dim=1536 component-node name=linear8 component=linear8 input=relu8.dropout input-dim=1536 output-dim=256 component-node name=dfsmn8.element_wise_scale component=dfsmn8.element_wise_scale input=Append(Offset(linear8, -16), Offset(linear8, -14), Offset(linear8, -12), Offset(linea r8, -10), Offset(linear8, -8), Offset(linear8, -6), Offset(linear8, -4), Offset(linear8, -2), linear8, Offset(linear8, 2), Offset(linear8, 4), Offset(linear8, 6), Offset(lin ear8, 8), Offset(linear8, 10), Offset(linear8, 12), Offset(linear8, 14), Offset(linear8, 16)) input-dim=4352 output-dim=4352 component-node name=dfsmn8 component=dfsmn8 input=dfsmn8.element_wise_scale input-dim=4352 output-dim=256 component-node name=relu9.affine component=relu9.affine input=dfsmn8 input-dim=256 output-dim=1536 component-node name=relu9.relu component=relu9.relu input=relu9.affine input-dim=1536 output-dim=1536 component-node name=relu9.batchnorm component=relu9.batchnorm input=relu9.relu input-dim=1536 output-dim=1536 component-node name=relu9.dropout component=relu9.dropout input=relu9.batchnorm input-dim=1536 output-dim=1536 component-node name=linear9 component=linear9 input=relu9.dropout input-dim=1536 output-dim=256 component-node name=no-op9 component=no-op9 input=Sum(linear9, Scale(0.66, no-op7)) input-dim=256 output-dim=256 component-node name=dfsmn9.element_wise_scale component=dfsmn9.element_wise_scale input=Append(Offset(no-op9, -20), Offset(no-op9, -18), Offset(no-op9, -16), Offset(no-op9, -14), Offset(no-op9, -12), Offset(no-op9, -10), Offset(no-op9, -8), Offset(no-op9, -6), Offset(no-op9, -4), Offset(no-op9, -2), no-op9, Offset(no-op9, 2), Offset(no-op9, 4), Offset(no-op9, 6), Offset(no-op9, 8), Offset(no-op9, 10), Offset(no-op9, 12), Offset(no-op9, 14), Offset(no-op9, 16), Offset(no-op9, 18), Offset(no-op9, 20)) input-dim=5376 output-dim=5376 component-node name=dfsmn9 component=dfsmn9 input=dfsmn9.element_wise_scale input-dim=5376 output-dim=256 component-node name=relu10.affine component=relu10.affine input=dfsmn9 input-dim=256 output-dim=1536 component-node name=relu10.relu component=relu10.relu input=relu10.affine input-dim=1536 output-dim=1536 component-node name=relu10.batchnorm component=relu10.batchnorm input=relu10.relu input-dim=1536 output-dim=1536 component-node name=relu10.dropout component=relu10.dropout input=relu10.batchnorm input-dim=1536 output-dim=1536 component-node name=linear10 component=linear10 input=relu10.dropout input-dim=1536 output-dim=256 component-node name=dfsmn10.element_wise_scale component=dfsmn10.element_wise_scale input=Append(Offset(linear10, -20), Offset(linear10, -18), Offset(linear10, -16), Offset( linear10, -14), Offset(linear10, -12), Offset(linear10, -10), Offset(linear10, -8), Offset(linear10, -6), Offset(linear10, -4), Offset(linear10, -2), linear10, Offset(linear 10, 2), Offset(linear10, 4), Offset(linear10, 6), Offset(linear10, 8), Offset(linear10, 10), Offset(linear10, 12), Offset(linear10, 14), Offset(linear10, 16), Offset(linear1 0, 18), Offset(linear10, 20)) input-dim=5376 output-dim=5376 component-node name=dfsmn10 component=dfsmn10 input=dfsmn10.element_wise_scale input-dim=5376 output-dim=256 component-node name=relu11.batchnorm component=relu11.batchnorm input=relu11.relu input-dim=1536 output-dim=1536 component-node name=relu11.dropout component=relu11.dropout input=relu11.batchnorm input-dim=1536 output-dim=1536 component-node name=relu11.affine component=relu11.affine input=dfsmn10 input-dim=256 output-dim=1536 component-node name=relu11.relu component=relu11.relu input=relu11.affine input-dim=1536 output-dim=1536 component-node name=relu11.renorm component=relu11.renorm input=relu11.relu input-dim=1536 output-dim=1536 component-node name=prefinal-l component=prefinal-l input=relu11.renorm input-dim=1536 output-dim=256 component-node name=prefinal-chain.affine component=prefinal-chain.affine input=prefinal-l input-dim=256 output-dim=1024 component-node name=prefinal-chain.relu component=prefinal-chain.relu input=prefinal-chain.affine input-dim=1024 output-dim=1024 component-node name=prefinal-chain.batchnorm1 component=prefinal-chain.batchnorm1 input=prefinal-chain.relu input-dim=1024 output-dim=1024 component-node name=prefinal-chain.linear component=prefinal-chain.linear input=prefinal-chain.batchnorm1 input-dim=1024 output-dim=256 component-node name=prefinal-chain.batchnorm2 component=prefinal-chain.batchnorm2 input=prefinal-chain.linear input-dim=256 output-dim=256 component-node name=output.affine component=output.affine input=prefinal-chain.batchnorm2 input-dim=256 output-dim=6184 output-node name=output input=output.affine dim=6184 objective=linear component-node name=prefinal-xent.affine component=prefinal-xent.affine input=prefinal-l input-dim=256 output-dim=1024 component-node name=prefinal-xent.relu component=prefinal-xent.relu input=prefinal-xent.affine input-dim=1024 output-dim=1024 component-node name=prefinal-xent.batchnorm1 component=prefinal-xent.batchnorm1 input=prefinal-xent.relu input-dim=1024 output-dim=1024 component-node name=prefinal-xent.linear component=prefinal-xent.linear input=prefinal-xent.batchnorm1 input-dim=1024 output-dim=256 component-node name=prefinal-xent.batchnorm2 component=prefinal-xent.batchnorm2 input=prefinal-xent.linear input-dim=256 output-dim=256 component-node name=output-xent.affine component=output-xent.affine input=prefinal-xent.batchnorm2 input-dim=256 output-dim=6184 component-node name=output-xent.log-softmax component=output-xent.log-softmax input=output-xent.affine input-dim=6184 output-dim=6184 output-node name=output-xent input=output-xent.log-softmax dim=6184 objective=linear component name=ivector-linear type=LinearComponent, input-dim=100, output-dim=160, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-change=0.75, para ms-rms=0.0159, params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.09,0.09,0.09,0.12 0.12,0.13,0.15,0.18,0.19 0.21,0.22,0.22,0.23), mean=0.156, stddev=0.02 86], params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.06,0.06,0.06,0.09 0.10,0.12,0.17,0.26,0.30 0.33,0.39,0.39,0.40), mean=0.185, stddev=0.0785], use-n atural-gradient=true, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=ivector-batchnorm type=BatchNormComponent, dim=160, block-dim=160, epsilon=0.001, target-rms=0.025, count=342592, test-mode=false, data-mean=[percentiles(0,1, 2,5 10,20,50,80,90 95,98,99,100)=(-0.03,-0.03,-0.03,-0.03 -0.02,-0.01,-0.002,0.01,0.02 0.02,0.02,0.03,0.03), mean=-0.0011, stddev=0.0134], data-stddev=[percentiles(0,1,2,5 1 0,20,50,80,90 95,98,99,100)=(0.07,0.07,0.08,0.09 0.10,0.11,0.12,0.15,0.16 0.16,0.17,0.18,0.18), mean=0.127, stddev=0.0221] component name=combine_inputs type=PermuteComponent, dim=240 , column-map=[ 0 1 80 81 82 ... ] component name=cnn1.conv type=TimeHeightConvolutionComponent, input-dim=240, output-dim=2560, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-change =0.75 num-filters-in=6, num-filters-out=64, height-in=40, height-out=40, height-subsample-out=1, {time,height}-offsets=[-3,-3 -3,-2 -3,-1 -3,0 -3,1 -3,2 -3,3 -2,-3 -2,-2 -2, -1 -2,0 -2,1 -2,2 -2,3 -1,-3 -1,-2 -1,-1 -1,0 -1,1 -1,2 -1,3 0,-3 0,-2 0,-1 0,0 0,1 0,2 0,3 1,-3 1,-2 1,-1 1,0 1,1 1,2 1,3 2,-3 2,-2 2,-1 2,0 2,1 2,2 2,3 3,-3 3,-2 3,-1 3,0 3,1 3,2 3,3], required-time-offsets=[-3,-2,-1,0,1,2,3], input-dim=240, output-dim=2560, filter-params-rms=0.08826, bias-params-{mean,stddev}=-0.005034,0.05817, num-params=18 880, max-memory-mb=200, use-natural-gradient=1, num-minibatches-history=4, rank-in=80, rank-out=32, alpha=4 component name=cnn1.relu type=RectifiedLinearComponent, dim=2560, block-dim=64, self-repair-lower-threshold=0.05, self-repair-scale=2e-05, count=1.53e+05, self-repaired-prop ortion=0, value-avg=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0,0,0,0 0,0.0003,0.08,4.4,11 15,21,22,34), mean=2.78, stddev=5.33], deriv-avg=[percentiles(0,1,2,5 10, 20,50,80,90 95,98,99,100)=(0,0,0,0 0,0.0005,0.05,0.82,0.95 1.0,1.0,1.0,1.0), mean=0.296, stddev=0.381] component name=cnn1.batchnorm type=BatchNormComponent, dim=2560, block-dim=64, epsilon=0.001, target-rms=1, count=1.3504e+07, test-mode=false, data-mean=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.28,0.28,0.36,0.44 0.47,0.59,0.99,3.2,8.6 12,14,16,18), mean=2.8, stddev=4.1], data-stddev=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)= (1.4,1.4,1.6,1.9 2.1,2.3,3.1,5.6,7.6 8.7,10,11,12), mean=4.17, stddev=2.44] component name=cnn2.conv type=TimeHeightConvolutionComponent, input-dim=2560, output-dim=2560, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-chang e=0.75 num-filters-in=64, num-filters-out=64, height-in=40, height-out=40, height-subsample-out=1, {time,height}-offsets=[-3,-3 -3,-2 -3,-1 -3,0 -3,1 -3,2 -3,3 -2,-3 -2,-2 - 2,-1 -2,0 -2,1 -2,2 -2,3 -1,-3 -1,-2 -1,-1 -1,0 -1,1 -1,2 -1,3 0,-3 0,-2 0,-1 0,0 0,1 0,2 0,3 1,-3 1,-2 1,-1 1,0 1,1 1,2 1,3 2,-3 2,-2 2,-1 2,0 2,1 2,2 2,3 3,-3 3,-2 3,-1 3, 0 3,1 3,2 3,3], required-time-offsets=[-3,-2,-1,0,1,2,3], input-dim=2560, output-dim=2560, filter-params-rms=0.02099, bias-params-{mean,stddev}=-0.06376,0.05827, num-params= 200768, max-memory-mb=200, use-natural-gradient=1, num-minibatches-history=4, rank-in=80, rank-out=32, alpha=4 component name=cnn2.relu type=RectifiedLinearComponent, dim=2560, block-dim=64, self-repair-lower-threshold=0.05, self-repair-scale=2e-05, count=1.28e+05, self-repaired-prop ortion=0, value-avg=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0,4e-08,0.0001,0.003 0.009,0.05,0.30,0.75,1.1 1.7,2.8,5.0,11), mean=0.521, stddev=0.864], deriv-avg=[p ercentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0,8e-06,0.0007,0.01 0.04,0.09,0.27,0.46,0.63 0.80,0.95,1.0,1.0), mean=0.299, stddev=0.234] component name=cnn2.batchnorm type=BatchNormComponent, dim=2560, block-dim=64, epsilon=0.001, target-rms=1, count=1.33043e+07, test-mode=false, data-mean=[percentiles(0,1,2, 5 10,20,50,80,90 95,98,99,100)=(0.08,0.08,0.09,0.17 0.18,0.23,0.47,0.71,0.81 1.1,1.2,1.3,1.5), mean=0.52, stddev=0.308], data-stddev=[percentiles(0,1,2,5 10,20,50,80,90 95,9 8,99,100)=(0.38,0.38,0.43,0.59 0.65,0.80,1.2,1.4,1.7 1.7,1.8,1.8,1.9), mean=1.15, stddev=0.367] component name=no-op1 type=NoOpComponent, dim=2560 component name=cnn3.conv type=TimeHeightConvolutionComponent, input-dim=2560, output-dim=2560, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-chang e=0.75 num-filters-in=64, num-filters-out=128, height-in=40, height-out=20, height-subsample-out=2, {time,height}-offsets=[-3,-3 -3,-2 -3,-1 -3,0 -3,1 -3,2 -3,3 -2,-3 -2,-2 -2,-1 -2,0 -2,1 -2,2 -2,3 -1,-3 -1,-2 -1,-1 -1,0 -1,1 -1,2 -1,3 0,-3 0,-2 0,-1 0,0 0,1 0,2 0,3 1,-3 1,-2 1,-1 1,0 1,1 1,2 1,3 2,-3 2,-2 2,-1 2,0 2,1 2,2 2,3 3,-3 3,-2 3,-1 3 ,0 3,1 3,2 3,3], required-time-offsets=[-3,-2,-1,0,1,2,3], input-dim=2560, output-dim=2560, filter-params-rms=0.01567, bias-params-{mean,stddev}=-0.01042,0.02196, num-params =401536, max-memory-mb=200, use-natural-gradient=1, num-minibatches-history=4, rank-in=80, rank-out=64, alpha=4 component name=cnn3.relu type=RectifiedLinearComponent, dim=2560, block-dim=128, self-repair-lower-threshold=0.05, self-repair-scale=2e-05, count=1.56e+05, self-repaired-pro portion=0, value-avg=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0,1e-05,0.0003,0.005 0.03,0.13,0.51,1.2,1.8 3.2,5.5,6.4,8.7), mean=0.862, stddev=1.19], deriv-avg=[pe rcentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0,6e-05,0.001,0.01 0.05,0.14,0.37,0.62,0.76 0.97,1.0,1.0,1.0), mean=0.395, stddev=0.265] component name=cnn3.batchnorm type=BatchNormComponent, dim=2560, block-dim=128, epsilon=0.001, target-rms=1, count=6.55232e+06, test-mode=false, data-mean=[percentiles(0,1,2 ,5 10,20,50,80,90 95,98,99,100)=(0.07,0.09,0.10,0.19 0.39,0.53,0.82,1.1,1.3 1.5,1.8,1.9,2.4), mean=0.857, stddev=0.404], data-stddev=[percentiles(0,1,2,5 10,20,50,80,90 95,9 8,99,100)=(0.45,0.52,0.59,0.81 1.1,1.3,1.6,1.8,2.0 2.1,2.1,2.2,2.4), mean=1.55, stddev=0.361] component name=cnn4.conv type=TimeHeightConvolutionComponent, input-dim=2560, output-dim=2560, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-chang e=0.75 num-filters-in=128, num-filters-out=128, height-in=20, height-out=20, height-subsample-out=1, {time,height}-offsets=[-3,-3 -3,-2 -3,-1 -3,0 -3,1 -3,2 -3,3 -2,-3 -2,-2 -2,-1 -2,0 -2,1 -2,2 -2,3 -1,-3 -1,-2 -1,-1 -1,0 -1,1 -1,2 -1,3 0,-3 0,-2 0,-1 0,0 0,1 0,2 0,3 1,-3 1,-2 1,-1 1,0 1,1 1,2 1,3 2,-3 2,-2 2,-1 2,0 2,1 2,2 2,3 3,-3 3,-2 3,-1 3,0 3,1 3,2 3,3], required-time-offsets=[-3,-2,-1,0,1,2,3], input-dim=2560, output-dim=2560, filter-params-rms=0.01056, bias-params-{mean,stddev}=-0.03392,0.03036, num-param s=802944, max-memory-mb=200, use-natural-gradient=1, num-minibatches-history=4, rank-in=80, rank-out=64, alpha=4 component name=cnn4.relu type=RectifiedLinearComponent, dim=2560, block-dim=128, self-repair-lower-threshold=0.05, self-repair-scale=2e-05, count=1.24e+05, self-repaired-pro portion=0, value-avg=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(3e-06,0.0009,0.003,0.01 0.02,0.07,0.21,0.52,0.94 1.7,2.5,2.8,3.8), mean=0.4, stddev=0.566], deriv-avg =[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(6e-05,0.004,0.01,0.03 0.05,0.10,0.21,0.45,0.63 0.92,1.0,1.0,1.0), mean=0.291, stddev=0.246] component name=cnn4.batchnorm type=BatchNormComponent, dim=2560, block-dim=128, epsilon=0.001, target-rms=1, count=6.45248e+06, test-mode=false, data-mean=[percentiles(0,1,2 ,5 10,20,50,80,90 95,98,99,100)=(0.05,0.06,0.07,0.09 0.12,0.16,0.31,0.65,0.82 0.94,1.1,1.1,1.3), mean=0.4, stddev=0.281], data-stddev=[percentiles(0,1,2,5 10,20,50,80,90 95, 98,99,100)=(0.32,0.35,0.37,0.42 0.50,0.59,0.80,1.0,1.1 1.2,1.3,1.3,1.3), mean=0.81, stddev=0.24] component name=no-op2 type=NoOpComponent, dim=2560 component name=cnn5.conv type=TimeHeightConvolutionComponent, input-dim=2560, output-dim=2560, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-chang e=0.75 num-filters-in=128, num-filters-out=256, height-in=20, height-out=10, height-subsample-out=2, {time,height}-offsets=[-3,-3 -3,-2 -3,-1 -3,0 -3,1 -3,2 -3,3 -2,-3 -2,-2 -2,-1 -2,0 -2,1 -2,2 -2,3 -1,-3 -1,-2 -1,-1 -1,0 -1,1 -1,2 -1,3 0,-3 0,-2 0,-1 0,0 0,1 0,2 0,3 1,-3 1,-2 1,-1 1,0 1,1 1,2 1,3 2,-3 2,-2 2,-1 2,0 2,1 2,2 2,3 3,-3 3,-2 3,-1 3,0 3,1 3,2 3,3], required-time-offsets=[-3,-2,-1,0,1,2,3], input-dim=2560, output-dim=2560, filter-params-rms=0.009226, bias-params-{mean,stddev}=-0.007338,0.004492, num-pa rams=1605888, max-memory-mb=200, use-natural-gradient=1, num-minibatches-history=4, rank-in=80, rank-out=80, alpha=4 component name=cnn5.relu type=RectifiedLinearComponent, dim=2560, block-dim=256, self-repair-lower-threshold=0.05, self-repair-scale=2e-05, count=2.21e+05, self-repaired-pro portion=0, value-avg=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.0006,0.007,0.01,0.02 0.04,0.08,0.22,0.53,0.89 1.9,2.9,3.2,4.4), mean=0.423, stddev=0.622], deriv-av g=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.002,0.02,0.03,0.05 0.07,0.10,0.18,0.34,0.52 0.79,0.99,1.0,1.0), mean=0.246, stddev=0.215] component name=cnn5.batchnorm type=BatchNormComponent, dim=2560, block-dim=256, epsilon=0.001, target-rms=1, count=3.17632e+06, test-mode=false, data-mean=[percentiles(0,1,2 ,5 10,20,50,80,90 95,98,99,100)=(0.05,0.06,0.07,0.09 0.12,0.16,0.36,0.68,0.85 0.97,1.0,1.1,1.1), mean=0.423, stddev=0.277], data-stddev=[percentiles(0,1,2,5 10,20,50,80,90 9 5,98,99,100)=(0.32,0.35,0.37,0.47 0.55,0.67,0.99,1.3,1.4 1.5,1.5,1.6,1.7), mean=0.986, stddev=0.319] component name=cnn6.conv type=TimeHeightConvolutionComponent, input-dim=2560, output-dim=2560, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-chang e=0.75 num-filters-in=256, num-filters-out=256, height-in=10, height-out=10, height-subsample-out=1, {time,height}-offsets=[-3,-3 -3,-2 -3,-1 -3,0 -3,1 -3,2 -3,3 -2,-3 -2,-2 -2,-1 -2,0 -2,1 -2,2 -2,3 -1,-3 -1,-2 -1,-1 -1,0 -1,1 -1,2 -1,3 0,-3 0,-2 0,-1 0,0 0,1 0,2 0,3 1,-3 1,-2 1,-1 1,0 1,1 1,2 1,3 2,-3 2,-2 2,-1 2,0 2,1 2,2 2,3 3,-3 3,-2 3,-1 3,0 3,1 3,2 3,3], required-time-offsets=[-3,-2,-1,0,1,2,3], input-dim=2560, output-dim=2560, filter-params-rms=0.006937, bias-params-{mean,stddev}=-0.01979,0.01155, num-para ms=3211520, max-memory-mb=200, use-natural-gradient=1, num-minibatches-history=4, rank-in=80, rank-out=80, alpha=4 component name=cnn6.relu type=RectifiedLinearComponent, dim=2560, block-dim=256, self-repair-lower-threshold=0.05, self-repair-scale=2e-05, count=1.2e+05, self-repaired-prop ortion=0, value-avg=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.004,0.02,0.03,0.05 0.08,0.11,0.20,0.39,0.69 1.3,2.1,2.7,3.5), mean=0.343, stddev=0.466], deriv-avg=[ percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.02,0.05,0.06,0.09 0.11,0.14,0.20,0.31,0.44 0.68,0.94,0.99,1.0), mean=0.25, stddev=0.181] component name=cnn6.batchnorm type=BatchNormComponent, dim=2560, block-dim=256, epsilon=0.001, target-rms=1, count=3.10976e+06, test-mode=false, data-mean=[percentiles(0,1,2 ,5 10,20,50,80,90 95,98,99,100)=(0.11,0.11,0.13,0.15 0.17,0.21,0.28,0.48,0.59 0.67,0.78,0.80,1.4), mean=0.342, stddev=0.178], data-stddev=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.43,0.48,0.52,0.56 0.61,0.68,0.81,1.0,1.1 1.2,1.2,1.3,1.6), mean=0.837, stddev=0.197] component name=relu1.affine type=NaturalGradientAffineComponent, input-dim=2560, output-dim=1536, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-ch ange=0.75, linear-params-rms=0.007281, linear-params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.13,0.16,0.18,0.22 0.29,0.33,0.37,0.41,0.43 0.44,0.46,0.47 ,0.55), mean=0.363, stddev=0.0614], linear-params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.04,0.06,0.08,0.11 0.15,0.19,0.27,0.35,0.38 0.42,0.47,0.50,0. 67), mean=0.27, stddev=0.0933], bias-{mean,stddev}=-0.001609,0.001725, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=relu1.relu type=RectifiedLinearComponent, dim=1536, self-repair-scale=1e-05, count=1.2e+05, self-repaired-proportion=0, value-avg=[percentiles(0,1,2,5 10,20,5 0,80,90 95,98,99,100)=(0.01,0.03,0.04,0.05 0.06,0.07,0.11,0.15,0.17 0.19,0.22,0.24,0.31), mean=0.113, stddev=0.0453], deriv-avg=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99, 100)=(0.06,0.09,0.11,0.14 0.16,0.19,0.26,0.35,0.40 0.43,0.47,0.50,0.60), mean=0.272, stddev=0.0914] component name=relu1.batchnorm type=BatchNormComponent, dim=1536, block-dim=1536, epsilon=0.001, target-rms=1, count=310976, test-mode=false, data-mean=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.01,0.03,0.04,0.05 0.06,0.07,0.11,0.15,0.17 0.20,0.22,0.23,0.31), mean=0.113, stddev=0.0454], data-stddev=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.08,0.14,0.15,0.17 0.19,0.22,0.28,0.33,0.36 0.38,0.41,0.43,0.49), mean=0.277, stddev=0.065] component name=relu1.dropout type=GeneralDropoutComponent, dim=1536, block-dim=1536, dropout-proportion=0, continuous=true component name=linear1 type=LinearComponent, input-dim=1536, output-dim=256, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-change=0.75, params-rms =0.01693, params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.35,0.35,0.45,0.50 0.52,0.54,0.61,0.77,0.86 0.93,0.99,1.0,1.1), mean=0.649, stddev=0.137], par ams-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.09,0.11,0.13,0.16 0.21,0.24,0.27,0.30,0.32 0.33,0.34,0.36,0.49), mean=0.267, stddev=0.048], use-natural-gr adient=true, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=dfsmn1.element_wise_scale type=NaturalGradientPerElementScaleComponent, input-dim=1280, output-dim=1280, learning-rate=0.0003, learning-rate-factor=0.25, max- change=0.75, scales-min=-4.58443, scales-max=4.51872, scales-{mean,stddev}=-0.007778,0.9049, rank=8, update-period=10, num-samples-history=2000, alpha=4 component name=dfsmn1 type=SumBlockComponent, input-dim=1280, output-dim=256, scale=0.2 component name=relu2.affine type=NaturalGradientAffineComponent, input-dim=256, output-dim=1536, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-cha nge=0.75, linear-params-rms=0.00958, linear-params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.03,0.03,0.04,0.04 0.05,0.06,0.15,0.20,0.22 0.24,0.25,0.26,0 .32), mean=0.138, stddev=0.0659], linear-params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.03,0.03,0.08,0.11 0.13,0.17,0.29,0.53,0.60 0.64,0.67,0.70,0.74 ), mean=0.332, stddev=0.175], bias-{mean,stddev}=-0.03457,0.02436, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=relu2.relu type=RectifiedLinearComponent, dim=1536, self-repair-scale=1e-05, count=9.29e+04, self-repaired-proportion=0, value-avg=[percentiles(0,1,2,5 10,20, 50,80,90 95,98,99,100)=(0.002,0.01,0.01,0.01 0.01,0.02,0.03,0.05,0.06 0.07,0.07,0.08,0.12), mean=0.0348, stddev=0.0174], deriv-avg=[percentiles(0,1,2,5 10,20,50,80,90 95,98, 99,100)=(0.10,0.18,0.19,0.22 0.23,0.25,0.31,0.39,0.44 0.46,0.51,0.53,0.60), mean=0.323, stddev=0.0802] component name=relu2.batchnorm type=BatchNormComponent, dim=1536, block-dim=1536, epsilon=0.001, target-rms=1, count=304320, test-mode=false, data-mean=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.002,0.009,0.01,0.01 0.01,0.02,0.03,0.05,0.06 0.07,0.07,0.08,0.12), mean=0.0347, stddev=0.0174], data-stddev=[percentiles(0,1,2,5 10,20,50,80, 90 95,98,99,100)=(0.008,0.03,0.03,0.03 0.04,0.04,0.07,0.10,0.11 0.12,0.12,0.13,0.15), mean=0.0733, stddev=0.0267] component name=relu2.dropout type=GeneralDropoutComponent, dim=1536, block-dim=1536, dropout-proportion=0, continuous=true component name=linear2 type=LinearComponent, input-dim=1536, output-dim=256, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-change=0.75, params-rms =0.01094, params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(1e-06,2e-05,2e-05,2e-05 0.002,0.24,0.46,0.52,0.54 0.57,0.58,0.59,0.62), mean=0.387, stddev=0.18 4], params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.03,0.03,0.04,0.04 0.05,0.06,0.17,0.23,0.25 0.27,0.28,0.30,0.36), mean=0.156, stddev=0.0786], use-na tural-gradient=true, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=dfsmn2.element_wise_scale type=NaturalGradientPerElementScaleComponent, input-dim=1280, output-dim=1280, learning-rate=0.0003, learning-rate-factor=0.25, max- change=0.75, scales-min=-3.62041, scales-max=3.48824, scales-{mean,stddev}=-0.0107,0.9752, rank=8, update-period=10, num-samples-history=2000, alpha=4 component name=dfsmn2 type=SumBlockComponent, input-dim=1280, output-dim=256, scale=0.2 component name=relu3.affine type=NaturalGradientAffineComponent, input-dim=256, output-dim=1536, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-cha nge=0.75, linear-params-rms=0.01004, linear-params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.02,0.03,0.03,0.05 0.06,0.08,0.14,0.21,0.24 0.27,0.30,0.31,0 .38), mean=0.145, stddev=0.0682], linear-params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(8e-07,4e-06,7e-06,1e-05 0.002,0.21,0.43,0.48,0.50 0.52,0.55,0.56 ,0.60), mean=0.353, stddev=0.173], bias-{mean,stddev}=-0.02898,0.01996, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=relu3.relu type=RectifiedLinearComponent, dim=1536, self-repair-scale=1e-05, count=2.3e+05, self-repaired-proportion=0, value-avg=[percentiles(0,1,2,5 10,20,5 0,80,90 95,98,99,100)=(0.0004,0.0008,0.001,0.002 0.002,0.003,0.007,0.01,0.02 0.02,0.03,0.04,0.06), mean=0.00852, stddev=0.00718], deriv-avg=[percentiles(0,1,2,5 10,20,50,80, 90 95,98,99,100)=(0.06,0.07,0.08,0.09 0.11,0.14,0.24,0.32,0.36 0.40,0.48,0.58,0.74), mean=0.24, stddev=0.105] component name=relu3.batchnorm type=BatchNormComponent, dim=1536, block-dim=1536, epsilon=0.001, target-rms=1, count=297664, test-mode=false, data-mean=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.0004,0.0008,0.001,0.002 0.002,0.003,0.007,0.01,0.02 0.02,0.03,0.04,0.06), mean=0.00847, stddev=0.00716], data-stddev=[percentiles(0,1,2,5 10, 20,50,80,90 95,98,99,100)=(0.002,0.003,0.004,0.005 0.007,0.01,0.02,0.03,0.03 0.04,0.05,0.05,0.09), mean=0.0203, stddev=0.0108] component name=relu3.dropout type=GeneralDropoutComponent, dim=1536, block-dim=1536, dropout-proportion=0, continuous=true component name=linear3 type=LinearComponent, input-dim=1536, output-dim=256, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-change=0.75, params-rms =0.01132, params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.20,0.21,0.25,0.27 0.30,0.34,0.42,0.51,0.58 0.65,0.68,0.71,0.94), mean=0.429, stddev=0.112], p arams-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.02,0.03,0.04,0.05 0.06,0.09,0.15,0.23,0.27 0.31,0.37,0.40,0.56), mean=0.161, stddev=0.0826], use-natural -gradient=true, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=no-op3 type=NoOpComponent, dim=256 component name=dfsmn3.element_wise_scale type=NaturalGradientPerElementScaleComponent, input-dim=2304, output-dim=2304, learning-rate=0.0003, learning-rate-factor=0.25, max- change=0.75, scales-min=-3.69449, scales-max=4.32537, scales-{mean,stddev}=0.00425,0.9956, rank=8, update-period=10, num-samples-history=2000, alpha=4 component name=dfsmn3 type=SumBlockComponent, input-dim=2304, output-dim=256, scale=0.111111 component name=relu4.affine type=NaturalGradientAffineComponent, input-dim=256, output-dim=1536, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-cha nge=0.75, linear-params-rms=0.01106, linear-params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.03,0.04,0.04,0.05 0.06,0.10,0.17,0.22,0.25 0.27,0.30,0.32,0 .40), mean=0.163, stddev=0.0685], linear-params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.10,0.13,0.14,0.18 0.23,0.28,0.40,0.54,0.60 0.63,0.69,0.70,0.75 ), mean=0.41, stddev=0.142], bias-{mean,stddev}=-0.03602,0.02756, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=relu4.relu type=RectifiedLinearComponent, dim=1536, self-repair-scale=1e-05, count=1.36e+05, self-repaired-proportion=0, value-avg=[percentiles(0,1,2,5 10,20, 50,80,90 95,98,99,100)=(0.0005,0.001,0.002,0.003 0.005,0.01,0.02,0.03,0.04 0.04,0.05,0.06,0.09), mean=0.0197, stddev=0.0125], deriv-avg=[percentiles(0,1,2,5 10,20,50,80,90 9 5,98,99,100)=(0.05,0.09,0.10,0.12 0.14,0.18,0.25,0.34,0.39 0.44,0.49,0.53,0.67), mean=0.261, stddev=0.097] component name=relu4.batchnorm type=BatchNormComponent, dim=1536, block-dim=1536, epsilon=0.001, target-rms=1, count=284352, test-mode=false, data-mean=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.0005,0.001,0.002,0.004 0.005,0.01,0.02,0.03,0.04 0.04,0.05,0.06,0.09), mean=0.0192, stddev=0.0122], data-stddev=[percentiles(0,1,2,5 10,20,50 ,80,90 95,98,99,100)=(0.002,0.005,0.007,0.01 0.02,0.03,0.04,0.06,0.07 0.07,0.09,0.09,0.12), mean=0.0433, stddev=0.0184] component name=relu4.dropout type=GeneralDropoutComponent, dim=1536, block-dim=1536, dropout-proportion=0, continuous=true component name=linear4 type=LinearComponent, input-dim=1536, output-dim=256, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-change=0.75, params-rms =0.01139, params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(2e-05,2e-05,3e-05,6e-05 0.21,0.39,0.46,0.51,0.54 0.56,0.59,0.61,0.64), mean=0.421, stddev=0.149 ], params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.03,0.04,0.05,0.05 0.06,0.10,0.17,0.23,0.26 0.28,0.30,0.31,0.47), mean=0.167, stddev=0.072], use-natu ral-gradient=true, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=dfsmn4.element_wise_scale type=NaturalGradientPerElementScaleComponent, input-dim=2304, output-dim=2304, learning-rate=0.0003, learning-rate-factor=0.25, max- change=0.75, scales-min=-3.50985, scales-max=3.81808, scales-{mean,stddev}=-0.01971,1.039, rank=8, update-period=10, num-samples-history=2000, alpha=4 component name=dfsmn4 type=SumBlockComponent, input-dim=2304, output-dim=256, scale=0.111111 component name=relu5.affine type=NaturalGradientAffineComponent, input-dim=256, output-dim=1536, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-cha nge=0.75, linear-params-rms=0.01194, linear-params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.02,0.02,0.02,0.02 0.02,0.02,0.06,0.30,0.36 0.40,0.44,0.46,0 .50), mean=0.132, stddev=0.138], linear-params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(8e-06,1e-05,1e-05,5e-05 0.25,0.38,0.48,0.54,0.57 0.60,0.62,0.63,0 .74), mean=0.44, stddev=0.159], bias-{mean,stddev}=-0.02296,0.03774, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=relu5.relu type=RectifiedLinearComponent, dim=1536, self-repair-scale=1e-05, count=1.26e+05, self-repaired-proportion=0, value-avg=[percentiles(0,1,2,5 10,20, 50,80,90 95,98,99,100)=(0.0002,0.0005,0.0007,0.002 0.003,0.006,0.01,0.01,0.02 0.03,0.03,0.04,0.47), mean=0.0117, stddev=0.014], deriv-avg=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.05,0.06,0.06,0.07 0.08,0.11,0.12,0.19,0.22 0.25,0.31,0.39,0.93), mean=0.146, stddev=0.0672] component name=relu5.batchnorm type=BatchNormComponent, dim=1536, block-dim=1536, epsilon=0.001, target-rms=1, count=271040, test-mode=false, data-mean=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.0002,0.0005,0.0008,0.002 0.003,0.006,0.01,0.01,0.02 0.02,0.03,0.04,0.47), mean=0.0115, stddev=0.014], data-stddev=[percentiles(0,1,2,5 10,20, 50,80,90 95,98,99,100)=(0.001,0.002,0.004,0.009 0.02,0.03,0.04,0.04,0.05 0.07,0.09,0.10,0.17), mean=0.0367, stddev=0.0175] component name=relu5.dropout type=GeneralDropoutComponent, dim=1536, block-dim=1536, dropout-proportion=0, continuous=true component name=linear5 type=LinearComponent, input-dim=1536, output-dim=256, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-change=0.75, params-rms =0.01457, params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.23,0.33,0.37,0.41 0.44,0.46,0.56,0.64,0.69 0.74,0.80,0.84,0.92), mean=0.561, stddev=0.106], p arams-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.02,0.02,0.02,0.02 0.02,0.03,0.06,0.34,0.42 0.49,0.58,0.66,0.88), mean=0.157, stddev=0.172], use-natural- gradient=true, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=no-op5 type=NoOpComponent, dim=256 component name=dfsmn5.element_wise_scale type=NaturalGradientPerElementScaleComponent, input-dim=3328, output-dim=3328, learning-rate=0.0003, learning-rate-factor=0.25, max- change=0.75, scales-min=-3.59081, scales-max=3.37754, scales-{mean,stddev}=0.01226,0.9403, rank=8, update-period=10, num-samples-history=2000, alpha=4 component name=dfsmn5 type=SumBlockComponent, input-dim=3328, output-dim=256, scale=0.0769231 component name=relu6.affine type=NaturalGradientAffineComponent, input-dim=256, output-dim=1536, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-cha nge=0.75, linear-params-rms=0.01012, linear-params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.001,0.02,0.03,0.05 0.07,0.09,0.15,0.20,0.23 0.25,0.28,0.30, 0.38), mean=0.149, stddev=0.0626], linear-params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.12,0.14,0.19,0.22 0.24,0.27,0.36,0.49,0.55 0.61,0.64,0.67,0.8 0), mean=0.377, stddev=0.123], bias-{mean,stddev}=-0.02671,0.01558, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=relu6.relu type=RectifiedLinearComponent, dim=1536, self-repair-scale=1e-05, count=1.71e+05, self-repaired-proportion=0, value-avg=[percentiles(0,1,2,5 10,20, 50,80,90 95,98,99,100)=(4e-05,0.0004,0.0007,0.001 0.002,0.003,0.006,0.009,0.01 0.02,0.02,0.04,0.07), mean=0.00696, stddev=0.00649], deriv-avg=[percentiles(0,1,2,5 10,20,50,8 0,90 95,98,99,100)=(0.08,0.09,0.10,0.11 0.13,0.14,0.19,0.26,0.28 0.32,0.39,0.48,0.66), mean=0.205, stddev=0.0733] component name=relu6.batchnorm type=BatchNormComponent, dim=1536, block-dim=1536, epsilon=0.001, target-rms=1, count=251072, test-mode=false, data-mean=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(4e-05,0.0004,0.0007,0.001 0.002,0.003,0.006,0.009,0.01 0.02,0.02,0.04,0.07), mean=0.00699, stddev=0.00651], data-stddev=[percentiles(0,1,2,5 10 ,20,50,80,90 95,98,99,100)=(0.0001,0.002,0.002,0.005 0.007,0.01,0.02,0.03,0.03 0.04,0.05,0.06,0.09), mean=0.019, stddev=0.0109] component name=relu6.dropout type=GeneralDropoutComponent, dim=1536, block-dim=1536, dropout-proportion=0, continuous=true component name=linear6 type=LinearComponent, input-dim=1536, output-dim=256, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-change=0.75, params-rms =0.009597, params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(2e-05,3e-05,3e-05,4e-05 0.02,0.29,0.39,0.45,0.47 0.49,0.52,0.56,0.59), mean=0.347, stddev=0.14 5], params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.001,0.02,0.03,0.04 0.06,0.09,0.14,0.19,0.22 0.25,0.28,0.30,0.43), mean=0.141, stddev=0.0619], use-n atural-gradient=true, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=dfsmn6.element_wise_scale type=NaturalGradientPerElementScaleComponent, input-dim=3328, output-dim=3328, learning-rate=0.0003, learning-rate-factor=0.25, max- change=0.75, scales-min=-3.78687, scales-max=5.40403, scales-{mean,stddev}=-0.005258,1.025, rank=8, update-period=10, num-samples-history=2000, alpha=4 component name=dfsmn6 type=SumBlockComponent, input-dim=3328, output-dim=256, scale=0.0769231 component name=relu7.affine type=NaturalGradientAffineComponent, input-dim=256, output-dim=1536, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-cha nge=0.75, linear-params-rms=0.01044, linear-params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.02,0.02,0.02,0.02 0.02,0.02,0.04,0.26,0.34 0.37,0.39,0.41,0 .47), mean=0.112, stddev=0.124], linear-params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(1e-05,1e-05,1e-05,3e-05 0.02,0.31,0.42,0.48,0.51 0.55,0.59,0.66,0 .78), mean=0.376, stddev=0.161], bias-{mean,stddev}=-0.009848,0.01768, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=relu7.relu type=RectifiedLinearComponent, dim=1536, self-repair-scale=1e-05, count=7.42e+04, self-repaired-proportion=0, value-avg=[percentiles(0,1,2,5 10,20, 50,80,90 95,98,99,100)=(0.0002,0.0008,0.001,0.002 0.003,0.004,0.006,0.006,0.007 0.009,0.01,0.02,0.05), mean=0.00557, stddev=0.00317], deriv-avg=[percentiles(0,1,2,5 10,20,50 ,80,90 95,98,99,100)=(0.05,0.06,0.07,0.07 0.07,0.07,0.10,0.15,0.18 0.20,0.27,0.31,0.73), mean=0.115, stddev=0.0574] component name=relu7.batchnorm type=BatchNormComponent, dim=1536, block-dim=1536, epsilon=0.001, target-rms=1, count=231104, test-mode=false, data-mean=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.0002,0.0008,0.001,0.002 0.003,0.004,0.005,0.006,0.007 0.009,0.01,0.02,0.05), mean=0.00542, stddev=0.00313], data-stddev=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.0009,0.004,0.005,0.009 0.01,0.02,0.02,0.02,0.03 0.03,0.04,0.05,0.11), mean=0.0214, stddev=0.00816] component name=relu7.dropout type=GeneralDropoutComponent, dim=1536, block-dim=1536, dropout-proportion=0, continuous=true component name=linear7 type=LinearComponent, input-dim=1536, output-dim=256, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-change=0.75, params-rms =0.0138, params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.15,0.26,0.30,0.34 0.36,0.40,0.49,0.61,0.71 0.83,0.96,0.98,1.2), mean=0.518, stddev=0.156], par ams-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.02,0.02,0.02,0.02 0.02,0.02,0.05,0.31,0.43 0.50,0.58,0.63,0.78), mean=0.14, stddev=0.171], use-natural-gra dient=true, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=no-op7 type=NoOpComponent, dim=256 component name=dfsmn7.element_wise_scale type=NaturalGradientPerElementScaleComponent, input-dim=4352, output-dim=4352, learning-rate=0.0003, learning-rate-factor=0.25, max- change=0.75, scales-min=-4.83806, scales-max=4.12363, scales-{mean,stddev}=0.003027,1.002, rank=8, update-period=10, num-samples-history=2000, alpha=4 component name=dfsmn7 type=SumBlockComponent, input-dim=4352, output-dim=256, scale=0.0588235 component name=relu8.affine type=NaturalGradientAffineComponent, input-dim=256, output-dim=1536, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-cha nge=0.75, linear-params-rms=0.01349, linear-params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.02,0.03,0.03,0.06 0.09,0.13,0.20,0.26,0.30 0.34,0.38,0.41,0 .52), mean=0.2, stddev=0.0817], linear-params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.18,0.24,0.25,0.28 0.32,0.40,0.50,0.63,0.70 0.73,0.74,0.76,0.80), mean=0.511, stddev=0.135], bias-{mean,stddev}=-0.02923,0.01643, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=relu8.relu type=RectifiedLinearComponent, dim=1536, self-repair-scale=1e-05, count=1.38e+05, self-repaired-proportion=0, value-avg=[percentiles(0,1,2,5 10,20, 50,80,90 95,98,99,100)=(0.0001,0.0003,0.0004,0.001 0.002,0.002,0.004,0.007,0.01 0.01,0.02,0.03,0.11), mean=0.00526, stddev=0.00655], deriv-avg=[percentiles(0,1,2,5 10,20,50, 80,90 95,98,99,100)=(0.06,0.08,0.09,0.10 0.11,0.13,0.17,0.23,0.30 0.38,0.45,0.50,0.87), mean=0.189, stddev=0.0905] component name=relu8.batchnorm type=BatchNormComponent, dim=1536, block-dim=1536, epsilon=0.001, target-rms=1, count=204480, test-mode=false, data-mean=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.0001,0.0003,0.0004,0.001 0.002,0.002,0.004,0.007,0.01 0.01,0.02,0.03,0.11), mean=0.00527, stddev=0.00656], data-stddev=[percentiles(0,1,2,5 1 0,20,50,80,90 95,98,99,100)=(0.0007,0.001,0.002,0.004 0.006,0.008,0.01,0.02,0.02 0.03,0.04,0.05,0.07), mean=0.0138, stddev=0.00906] component name=relu8.dropout type=GeneralDropoutComponent, dim=1536, block-dim=1536, dropout-proportion=0, continuous=true component name=linear8 type=LinearComponent, input-dim=1536, output-dim=256, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-change=0.75, params-rms =0.01101, params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(1e-05,1e-05,2e-05,0.007 0.17,0.32,0.44,0.51,0.55 0.57,0.63,0.65,0.74), mean=0.404, stddev=0.153 ], params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.01,0.02,0.03,0.05 0.08,0.11,0.16,0.21,0.24 0.27,0.32,0.35,0.48), mean=0.164, stddev=0.0655], use-nat ural-gradient=true, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=dfsmn8.element_wise_scale type=NaturalGradientPerElementScaleComponent, input-dim=4352, output-dim=4352, learning-rate=0.0003, learning-rate-factor=0.25, max- change=0.75, scales-min=-5.84124, scales-max=3.73046, scales-{mean,stddev}=0.007816,1.016, rank=8, update-period=10, num-samples-history=2000, alpha=4 component name=dfsmn8 type=SumBlockComponent, input-dim=4352, output-dim=256, scale=0.0588235 component name=relu9.affine type=NaturalGradientAffineComponent, input-dim=256, output-dim=1536, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-cha nge=0.75, linear-params-rms=0.01066, linear-params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.001,0.002,0.002,0.003 0.004,0.008,0.08,0.24,0.30 0.34,0.38, 0.39,0.66), mean=0.121, stddev=0.12], linear-params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(4e-06,5e-06,8e-06,0.009 0.19,0.31,0.43,0.49,0.53 0.56,0.58,0 .60,0.73), mean=0.392, stddev=0.144], bias-{mean,stddev}=-0.01017,0.01133, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=relu9.relu type=RectifiedLinearComponent, dim=1536, self-repair-scale=1e-05, count=1.09e+05, self-repaired-proportion=0, value-avg=[percentiles(0,1,2,5 10,20, 50,80,90 95,98,99,100)=(7e-06,2e-05,3e-05,4e-05 5e-05,0.0001,0.001,0.003,0.005 0.007,0.01,0.01,0.03), mean=0.002, stddev=0.00273], deriv-avg=[percentiles(0,1,2,5 10,20,50,80 ,90 95,98,99,100)=(0.04,0.09,0.10,0.12 0.13,0.15,0.19,0.24,0.27 0.29,0.34,0.39,0.83), mean=0.199, stddev=0.0619] component name=relu9.batchnorm type=BatchNormComponent, dim=1536, block-dim=1536, epsilon=0.001, target-rms=1, count=177856, test-mode=false, data-mean=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(7e-06,2e-05,3e-05,4e-05 5e-05,0.0001,0.001,0.003,0.005 0.007,0.01,0.01,0.03), mean=0.00199, stddev=0.00273], data-stddev=[percentiles(0,1,2,5 1 0,20,50,80,90 95,98,99,100)=(3e-05,7e-05,8e-05,0.0001 0.0002,0.0003,0.003,0.01,0.02 0.02,0.02,0.03,0.06), mean=0.00592, stddev=0.00715] component name=relu9.dropout type=GeneralDropoutComponent, dim=1536, block-dim=1536, dropout-proportion=0, continuous=true component name=linear9 type=LinearComponent, input-dim=1536, output-dim=256, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-change=0.75, params-rms =0.01046, params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.08,0.09,0.12,0.22 0.27,0.31,0.37,0.47,0.54 0.61,0.67,0.73,0.89), mean=0.392, stddev=0.121], p arams-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.0009,0.002,0.002,0.003 0.004,0.007,0.07,0.23,0.29 0.34,0.40,0.43,0.57), mean=0.117, stddev=0.119], use-n atural-gradient=true, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=no-op9 type=NoOpComponent, dim=256 component name=dfsmn9.element_wise_scale type=NaturalGradientPerElementScaleComponent, input-dim=5376, output-dim=5376, learning-rate=0.0003, learning-rate-factor=0.25, max- change=0.75, scales-min=-4.32281, scales-max=4.27952, scales-{mean,stddev}=-0.01476,0.9989, rank=8, update-period=10, num-samples-history=2000, alpha=4 component name=dfsmn9 type=SumBlockComponent, input-dim=5376, output-dim=256, scale=0.047619 component name=relu10.affine type=NaturalGradientAffineComponent, input-dim=256, output-dim=1536, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-ch ange=0.75, linear-params-rms=0.0237, linear-params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.06,0.11,0.13,0.20 0.24,0.29,0.36,0.43,0.50 0.56,0.62,0.67,1 .1), mean=0.363, stddev=0.108], linear-params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.30,0.35,0.47,0.56 0.66,0.78,0.93,1.0,1.1 1.2,1.2,1.2,1.3), mean= 0.91, stddev=0.184], bias-{mean,stddev}=-0.02952,0.02796, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=relu10.relu type=RectifiedLinearComponent, dim=1536, self-repair-scale=1e-05, count=5.32e+04, self-repaired-proportion=0, value-avg=[percentiles(0,1,2,5 10,20 ,50,80,90 95,98,99,100)=(0.0005,0.001,0.001,0.002 0.002,0.002,0.003,0.007,0.01 0.03,0.05,0.06,0.21), mean=0.00719, stddev=0.0146], deriv-avg=[percentiles(0,1,2,5 10,20,50,80 ,90 95,98,99,100)=(0.03,0.05,0.05,0.05 0.06,0.07,0.09,0.15,0.23 0.36,0.65,0.75,0.95), mean=0.13, stddev=0.125] component name=relu10.batchnorm type=BatchNormComponent, dim=1536, block-dim=1536, epsilon=0.001, target-rms=1, count=144576, test-mode=false, data-mean=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.0005,0.001,0.001,0.002 0.002,0.002,0.003,0.007,0.01 0.03,0.05,0.06,0.21), mean=0.00717, stddev=0.0146], data-stddev=[percentiles(0,1,2,5 10, 20,50,80,90 95,98,99,100)=(0.002,0.006,0.007,0.009 0.01,0.01,0.01,0.02,0.04 0.06,0.08,0.09,0.26), mean=0.0212, stddev=0.0192] component name=relu10.dropout type=GeneralDropoutComponent, dim=1536, block-dim=1536, dropout-proportion=0, continuous=true component name=linear10 type=LinearComponent, input-dim=1536, output-dim=256, learning-rate=0.0003, l2-regularize=0.01, learning-rate-factor=0.25, max-change=0.75, params-rm s=0.0182, params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.13,0.19,0.41,0.46 0.50,0.54,0.68,0.83,0.94 1.0,1.1,1.1,1.1), mean=0.692, stddev=0.174], param s-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.05,0.09,0.11,0.16 0.19,0.22,0.27,0.34,0.38 0.42,0.47,0.51,0.71), mean=0.28, stddev=0.0813], use-natural-grad ient=true, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=dfsmn10.element_wise_scale type=NaturalGradientPerElementScaleComponent, input-dim=5376, output-dim=5376, learning-rate=0.0003, learning-rate-factor=0.25, max -change=0.75, scales-min=-5.68368, scales-max=7.01841, scales-{mean,stddev}=-0.007358,0.9309, rank=8, update-period=10, num-samples-history=2000, alpha=4 component name=dfsmn10 type=SumBlockComponent, input-dim=5376, output-dim=256, scale=0.047619 component name=relu11.affine type=NaturalGradientAffineComponent, input-dim=256, output-dim=1536, learning-rate=0.0012, l2-regularize=0.01, max-change=0.75, linear-params-rm s=0.02592, linear-params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.11,0.11,0.11,0.12 0.12,0.12,0.45,0.54,0.61 0.65,0.72,0.75,0.88), mean=0.359, stddev=0 .207], linear-params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.33,0.38,0.60,0.67 0.73,0.81,0.99,1.2,1.3 1.3,1.4,1.4,1.5), mean=0.994, stddev=0.209], bia s-{mean,stddev}=0.01739,0.1122, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=relu11.relu type=RectifiedLinearComponent, dim=1536, self-repair-scale=1e-05, count=2.03e+04, self-repaired-proportion=0, value-avg=[percentiles(0,1,2,5 10,20 ,50,80,90 95,98,99,100)=(0,0,0,0 0,1e-07,0.03,0.11,0.16 0.22,0.28,0.31,0.54), mean=0.0575, stddev=0.0749], deriv-avg=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0,0,0 ,0 0,5e-05,0.44,0.94,0.99 1.0,1.0,1.0,1.0), mean=0.466, stddev=0.403] component name=relu11.batchnorm type=BatchNormComponent, dim=1536, block-dim=1536, epsilon=0.001, target-rms=1, count=0, test-mode=false component name=relu11.dropout type=GeneralDropoutComponent, dim=1536, block-dim=1536, dropout-proportion=0, continuous=true component name=prefinal-l type=LinearComponent, input-dim=1536, output-dim=256, learning-rate=0.0012, max-change=0.75, params-rms=0.05396, params-row-norms=[percentiles(0,1, 2,5 10,20,50,80,90 95,98,99,100)=(1.8,1.8,1.9,1.9 2.0,2.0,2.1,2.2,2.3 2.3,2.4,2.4,2.7), mean=2.11, stddev=0.127], params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98, 99,100)=(0.36,0.37,0.38,0.39 0.40,0.41,0.94,1.1,1.2 1.3,1.3,1.3,1.6), mean=0.792, stddev=0.343], use-natural-gradient=true, rank-in=20, rank-out=80, num-samples-history=2000 , update-period=4, alpha=4 component name=prefinal-chain.affine type=NaturalGradientAffineComponent, input-dim=256, output-dim=1024, learning-rate=0.0012, max-change=0.75, linear-params-rms=0.07357, l inear-params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.81,0.92,0.94,0.97 1.0,1.0,1.1,1.3,1.4 1.4,1.5,1.5,1.7), mean=1.17, stddev=0.153], linear-params-c ol-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(2.1,2.2,2.2,2.2 2.2,2.3,2.4,2.4,2.5 2.5,2.5,2.6,2.6), mean=2.35, stddev=0.0847], bias-{mean,stddev}=0.005361,0.97 7, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=prefinal-chain.relu type=RectifiedLinearComponent, dim=1024, self-repair-scale=1e-05, count=2.11e+04, self-repaired-proportion=0, value-avg=[percentiles(0,1,2 ,5 10,20,50,80,90 95,98,99,100)=(0.17,0.25,0.29,0.37 0.47,0.76,1.8,3.5,4.7 5.8,6.6,7.1,9.4), mean=2.25, stddev=1.69], deriv-avg=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99, 100)=(0.07,0.09,0.10,0.12 0.14,0.18,0.32,0.50,0.62 0.68,0.73,0.76,0.83), mean=0.345, stddev=0.174] component name=prefinal-chain.batchnorm1 type=BatchNormComponent, dim=1024, block-dim=1024, epsilon=0.001, target-rms=1, count=38208, test-mode=false, data-mean=[percentiles (0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.17,0.26,0.29,0.37 0.47,0.75,1.8,3.5,4.7 5.7,6.6,7.1,9.3), mean=2.25, stddev=1.69], data-stddev=[percentiles(0,1,2,5 10,20,50,80,90 9 5,98,99,100)=(0.73,1.0,1.1,1.3 1.6,2.2,4.0,5.6,6.2 6.5,7.0,7.3,7.9), mean=3.96, stddev=1.68] component name=prefinal-chain.linear type=LinearComponent, input-dim=1024, output-dim=256, learning-rate=0.0012, max-change=0.75, params-rms=0.06253, params-row-norms=[perce ntiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(2.0,2.0,2.0,2.0 2.0,2.0,2.0,2.0,2.0 2.0,2.0,2.0,2.0), mean=2, stddev=0.0012], params-col-norms=[percentiles(0,1,2,5 10,20,50,80, 90 95,98,99,100)=(0.60,0.74,0.77,0.81 0.84,0.87,0.96,1.1,1.2 1.2,1.3,1.3,1.4), mean=0.991, stddev=0.136], orthonormal-constraint=-1, use-natural-gradient=true, rank-in=20, r ank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=prefinal-chain.batchnorm2 type=BatchNormComponent, dim=256, block-dim=256, epsilon=0.001, target-rms=1, count=38208, test-mode=false, data-mean=[percentiles(0 ,1,2,5 10,20,50,80,90 95,98,99,100)=(-2e-07,-2e-07,-1e-07,-1e-07 -9e-08,-5e-08,2e-09,6e-08,9e-08 1e-07,1e-07,2e-07,2e-07), mean=3.05e-09, stddev=6.85e-08], data-stddev=[perc entiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(2.1,2.2,2.3,2.3 2.4,2.5,2.6,2.8,2.9 3.0,3.2,3.3,3.5), mean=2.66, stddev=0.232] component name=output.affine type=NaturalGradientAffineComponent, input-dim=256, output-dim=6184, learning-rate=0.0012, max-change=1.5, linear-params-rms=0.03309, linear-par ams-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(3e-10,0.07,0.11,0.17 0.24,0.32,0.48,0.66,0.76 0.84,0.93,1.0,1.3), mean=0.49, stddev=0.201], linear-params-co l-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(2.3,2.3,2.3,2.4 2.4,2.5,2.6,2.7,2.8 2.8,2.9,2.9,2.9), mean=2.6, stddev=0.13], bias-{mean,stddev}=0.0001979,0.05861 , rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=prefinal-xent.affine type=NaturalGradientAffineComponent, input-dim=256, output-dim=1024, learning-rate=0.0012, max-change=0.75, linear-params-rms=0.06393, li near-params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.66,0.71,0.73,0.75 0.77,0.81,1.1,1.2,1.2 1.2,1.2,1.3,1.3), mean=1.01, stddev=0.171], linear-params- col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(1.9,1.9,2.0,2.0 2.0,2.0,2.0,2.1,2.1 2.1,2.1,2.2,2.2), mean=2.05, stddev=0.0465], bias-{mean,stddev}=-0.006102,0. 9595, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=prefinal-xent.relu type=RectifiedLinearComponent, dim=1024, self-repair-scale=1e-05, count=6.4e+03, self-repaired-proportion=0, value-avg=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.04,0.08,0.11,0.15 0.18,0.24,1.2,2.5,3.0 3.3,3.8,4.3,7.1), mean=1.39, stddev=1.16], deriv-avg=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,10 0)=(0.05,0.07,0.09,0.12 0.16,0.21,0.40,0.65,0.71 0.76,0.79,0.83,0.89), mean=0.422, stddev=0.212] component name=prefinal-xent.batchnorm1 type=BatchNormComponent, dim=1024, block-dim=1024, epsilon=0.001, target-rms=1, count=38208, test-mode=false, data-mean=[percentiles( 0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.04,0.08,0.10,0.15 0.18,0.23,1.2,2.5,3.0 3.4,3.9,4.3,7.1), mean=1.39, stddev=1.16], data-stddev=[percentiles(0,1,2,5 10,20,50,80,90 95 ,98,99,100)=(0.26,0.35,0.40,0.48 0.54,0.66,2.2,2.8,3.1 3.3,3.7,4.1,5.4), mean=1.88, stddev=1.06] component name=prefinal-xent.linear type=LinearComponent, input-dim=1024, output-dim=256, learning-rate=0.0012, max-change=0.75, params-rms=0.04972, params-row-norms=[percen tiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(1.6,1.6,1.6,1.6 1.6,1.6,1.6,1.6,1.6 1.6,1.6,1.6,1.6), mean=1.59, stddev=0.00169], params-col-norms=[percentiles(0,1,2,5 10,20,50, 80,90 95,98,99,100)=(0.42,0.45,0.49,0.53 0.56,0.62,0.82,0.91,0.96 1.0,1.1,1.1,1.1), mean=0.78, stddev=0.155], orthonormal-constraint=-1, use-natural-gradient=true, rank-in=2 0, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=prefinal-xent.batchnorm2 type=BatchNormComponent, dim=256, block-dim=256, epsilon=0.001, target-rms=1, count=38208, test-mode=false, data-mean=[percentiles(0, 1,2,5 10,20,50,80,90 95,98,99,100)=(-2e-07,-1e-07,-1e-07,-9e-08 -6e-08,-4e-08,4e-09,5e-08,8e-08 1e-07,1e-07,1e-07,2e-07), mean=5.98e-09, stddev=5.82e-08], data-stddev=[perce ntiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(1.6,1.6,1.6,1.7 1.7,1.7,1.8,1.9,1.9 2.0,2.0,2.1,2.1), mean=1.83, stddev=0.0978] component name=output-xent.affine type=NaturalGradientAffineComponent, input-dim=256, output-dim=6184, learning-rate=0.006, learning-rate-factor=5, max-change=1.5, linear-pa rams-rms=0.07729, linear-params-row-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(0.03,0.16,0.32,0.70 0.92,1.0,1.2,1.4,1.6 1.7,1.8,1.9,2.4), mean=1.2, stddev=0.30 2], linear-params-col-norms=[percentiles(0,1,2,5 10,20,50,80,90 95,98,99,100)=(5.1,5.2,5.3,5.4 5.6,5.7,6.1,6.4,6.5 6.6,6.7,6.8,7.0), mean=6.07, stddev=0.363], bias-{mean,std dev}=-1.954e-08,0.1214, rank-in=20, rank-out=80, num-samples-history=2000, update-period=4, alpha=4 component name=output-xent.log-softmax type=LogSoftmaxComponent, dim=6184 component name=relu11.renorm type=NormalizeComponent, input-dim=1536, output-dim=1536, target-rms=1, add-log-stddev=false

[alumae]

Are you sure it's the right model? Because I don't see any LDA component in it.

[yangxueruivs]

Yes, I didnt use LDA-component and it runs smoothly when offline.

[alumae]

OK, please paste your YAML file.

[yangxueruivs]

use-nnet2: True decoder:

All the properties nested here correspond to the kaldinnet2onlinedecoder GStreamer plugin properties.

# Use gst-inspect-1.0 ./libgstkaldionline2.so kaldinnet2onlinedecoder to discover the available properties
feature-type: fbank
add-pitch: True
online-pitch-config: models/conf/online_pitch.conf
nnet-mode: 3
use-threaded-decoder:  true
model: models/final.mdl
word-syms: models/words.txt
fst: models/HCLG.fst
fbank-config: models/conf/fbank.conf
ivector-extraction-config: models/conf/ivector_extractor.conf
max-active: 10000
beam: 10.0
lattice-beam: 6.0
acoustic-scale: 1
do-endpointing: false
endpoint-silence-phones : "1:2:3:4:5:6:7:8:9:10"
traceback-period-in-secs: 1
chunk-length-in-secs: 1
num-nbest: 10
# frame-subsampling-factor: 1
#Additional functionality that you can play with:
#lm-fst:  test/models/english/tedlium_nnet_ms_sp_online/G.fst
#big-lm-const-arpa: test/models/english/tedlium_nnet_ms_sp_online/G.carpa
#phone-syms: test/models/english/tedlium_nnet_ms_sp_online/phones.txt
#word-boundary-file: test/models/english/tedlium_nnet_ms_sp_online/word_boundary.int
#do-phone-alignment: true

out-dir: tmp

use-vad: False silence-timeout: 10

post-processor: perl -npe 'BEGIN {use IO::Handle; STDOUT->autoflush(1);} s/(.*)/\1./;'

full-post-processor: ./models/sample_full_post_processor.py

logging: version : 1 disable_existing_loggers: False formatters: simpleFormater: format: '%(asctime)s - %(levelname)7s: %(name)10s: %(message)s' datefmt: '%Y-%m-%d %H:%M:%S' handlers: console: class: logging.StreamHandler formatter: simpleFormater level: DEBUG root: level: DEBUG handlers: [console]

[alumae]

OK, It think it might be related to ivector extractor. Could it be that you used pitch features as input to the the i-vector extractor?

[yangxueruivs]

Yes, so pitch extracted i-vector is not supported?

[alumae]

Yes, see a related issue in Kaldi: https://github.com/kaldi-asr/kaldi/issues/2657

[yangxueruivs]

Thank you, I'll have a look.

[yangxueruivs]

Sorry I forgot when I extracted i-vector, I used train_sp_hires_nopitch data, so pitch info should not be in i-vector.