nnet3 architecture - SPEAKER RECOGNITION FOR MULTI-SPEAKER CONVERSATIONS USING X-VECTORS

[Bar Madar]

Hey,

I am reading the paper - "SPEAKER RECOGNITION FOR MULTI-SPEAKER CONVERSATIONS USING X-VECTORS" by David Snyder et al.

The nnet architecture that used in the paper is different from the one that we train on the kaldi recipe "sitw/v2". 

Is there a recipe or code on kaldi project that train the same nnet as shown in the paper?

If not, how should I modify the "run_xvector_1a.sh" to this architecture? just adding layers to the network.xconfig file or should I do more changes on other files?

Thanks,

Bar

[David Snyder]

Hi Bar,

You will probably find this discussion on Kaldi help to be useful: https://kaldi-asr.org/forums.html?place=msg%2Fkaldi-help%2FWLCtGOaT6Uc%2FjIWHIhF4BAAJ . It's about getting this architecture to work in Kaldi.

Basically, you need to add the extra layers as specified in the paper, and increase the amount of training data. You can do the latter by retaining all of the augmented training data (it will take up a lot of space on disk, and will take awhile to train). Also, it helps if you reduce the archive size. I believe you can do that by reducing the number of frames per iter option in get_egs.sh. As I recall, having  around 800 training archives where each has around 200,000 examples per archive will produce good results. You'll have to play with the options to get_egs.sh to figure out how obtain this. I believe I left comments in the scripts that should provide some guidance. 

Best,

Davd

[Bar Madar]

Hey, thanks!

I added extra layers due to the article, and changed the frame per iter in get_egs.sh to 160000000 so the number of training archives is around 800.

I am using the 6M voxceleb augmented data to train the nnet.

The problem is that the get_egs.sh is failed after few hours because of out of memory on the disk. I have a SSD hard disk of 4T and before I run the get_egs.sh there are 2.2T free, but before the script is finished, it used all the 2.2T and failed.

Do I need more space for this run? or should I make some changes in the script so it will be more cost effective in memory?

Thanks alot!

Bar 

‫בתאריך יום ה׳, 21 בנוב׳ 2019 ב-4:12 מאת ‪David Snyder‬‏ <‪david.ry...@gmail.com‬‏>:‬

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.
To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/13bcc4e4-d7e4-4ea3-a0e4-53471ad82aab%40googlegroups.com.

[Daniel Povey]

I just checked the code of nnet3-xvector-get-egs.cc, and it looks like it is writing the egs without compression,

which would make them a factor of 4 larger than they need to be.

In line 188:

    SequentialBaseFloatMatrixReader feat_reader(feature_rspecifier);

it should be of type  SequentialGeneralMatrixReader.  Other parts of the code would need to use GeneralMatrix as

well.

Look in nnet3-get-egs.cc, where it uses  ExtractRowRangeWithPadding().  It would be necessary to use

that function in nnet3-xvector-get-egs.cc:WriteExamples(), which would of course require figuring out the

exact interface and how to convert the code to use it and have equivalent behavior.

It would be great if you could contribute the code back, if you do this.

Dan

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CAFHC87HUsU53rsC%3D0_Lv7WPm5YN_iWexL1HMMJJmHnL9TQVpNw%40mail.gmail.com.

[Bar Madar]

OK, thank you!

Just to clarify before I will make the changes and contribute the code back:

1. On nnet3-xvector-get-egs.cc, the function SequentialBaseFloatMatrixReader should be replaced to SequentialGeneralMatrixReader.

2. What are the other parts in nnet3-xvector-get-egs.cc that use "BaseFloatMatrix" and should be changed to GeneralMatrix? (line 188 is the only part using the SequentialBaseFloatMatrixReader).

3. I don't understand how to use the ExtractRowRangeWithPadding() in nnet3-xvector-get-egs.cc:WriteExamples(), may I have more explanations on that?

Thanks,

Bar

 

‫בתאריך יום א׳, 24 בנוב׳ 2019 ב-11:38 מאת ‪Daniel Povey‬‏ <‪dpo...@gmail.com‬‏>:‬

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CAEWAuyS9BHeXWVM%2BTRoQjF-eLGOabUOOVGtpJ81JdM7VCfYt8Q%40mail.gmail.com.

[Daniel Povey]

OK, thank you!

Just to clarify before I will make the changes and contribute the code back:

1. On nnet3-xvector-get-egs.cc, the function SequentialBaseFloatMatrixReader should be replaced to SequentialGeneralMatrixReader.

Yeah.. but it's a class name, not a function. 

 

2. What are the other parts in nnet3-xvector-get-egs.cc that use "BaseFloatMatrix" and should be changed to GeneralMatrix? (line 188 is the only part using the SequentialBaseFloatMatrixReader).

You can let it be dictated by compilation errors.

 

3. I don't understand how to use the ExtractRowRangeWithPadding() in nnet3-xvector-get-egs.cc:WriteExamples(), may I have more explanations on that?

That code is extracting a range of rows of a matrix, and should be replaced with a single function call to
ExtractRowRangeWithPadding(); that handles compressed matrices.

You may have to ask for help from someone with more C++ experience.

Dan

 

 

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CAFHC87E%3DggtnKTSMBmVoxVJDxs4NiWNdorKo1jBE7S-jwUPszA%40mail.gmail.com.

[Bar Madar]

Ok,

In the meantime, I reduced the number of archives to 650 and now I have enough memory space for the egs, so the get_egs finished successfully.

I will try to adapt the nnet3-xvector-get-egs.cc to working with compressed matrix as you explained and let you know.

Thanks!

Bar Madar

‫בתאריך יום א׳, 24 בנוב׳ 2019 ב-13:03 מאת ‪Daniel Povey‬‏ <‪dpo...@gmail.com‬‏>:‬

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CAEWAuyQ1MDgph1rpOn9eNHChQnxVMbzpMe2zUE1a3CDpPveq9A%40mail.gmail.com.

[David Snyder]

How many examples do you have per archive? I'd recommend having around 200,000 per archive. You can work out this information from the ranges.* files present in the egs/temp/ directory. 

Thanks,

Bar

 

‫בתאריך יום ה׳, 21 בנוב׳ 2019 ב-4:12 מאת ‪David Snyder‬‏ <‪david.r...@gmail.com‬‏>:‬

Hi Bar,

You will probably find this discussion on Kaldi help to be useful: https://kaldi-asr.org/forums.html?place=msg%2Fkaldi-help%2FWLCtGOaT6Uc%2FjIWHIhF4BAAJ . It's about getting this architecture to work in Kaldi.

Basically, you need to add the extra layers as specified in the paper, and increase the amount of training data. You can do the latter by retaining all of the augmented training data (it will take up a lot of space on disk, and will take awhile to train). Also, it helps if you reduce the archive size. I believe you can do that by reducing the number of frames per iter option in get_egs.sh. As I recall, having  around 800 training archives where each has around 200,000 examples per archive will produce good results. You'll have to play with the options to get_egs.sh to figure out how obtain this. I believe I left comments in the scripts that should provide some guidance. 

Best,

Davd

On Wednesday, November 20, 2019 at 5:25:41 AM UTC-5, Bar Madar wrote:

Hey,

I am reading the paper - "SPEAKER RECOGNITION FOR MULTI-SPEAKER CONVERSATIONS USING X-VECTORS" by David Snyder et al.

The nnet architecture that used in the paper is different from the one that we train on the kaldi recipe "sitw/v2". 

Is there a recipe or code on kaldi project that train the same nnet as shown in the paper?

If not, how should I modify the "run_xvector_1a.sh" to this architecture? just adding layers to the network.xconfig file or should I do more changes on other files?

Thanks,

Bar

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/13bcc4e4-d7e4-4ea3-a0e4-53471ad82aab%40googlegroups.com.

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CAFHC87HUsU53rsC%3D0_Lv7WPm5YN_iWexL1HMMJJmHnL9TQVpNw%40mail.gmail.com.

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CAEWAuyS9BHeXWVM%2BTRoQjF-eLGOabUOOVGtpJ81JdM7VCfYt8Q%40mail.gmail.com.

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CAFHC87E%3DggtnKTSMBmVoxVJDxs4NiWNdorKo1jBE7S-jwUPszA%40mail.gmail.com.

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi...@googlegroups.com.

[Bar Madar]

Hey,

Can you please explain to me how to get this information (num of example per archive) from the ranges.* file?

Or maybe link me to documentation about this?

Thanks!

Bar 

‫בתאריך יום ד׳, 27 בנוב׳ 2019 ב-18:06 מאת ‪David Snyder‬‏ <‪david.ry...@gmail.com‬‏>:‬

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/d53050e8-a504-436c-9a7a-40e669777ebb%40googlegroups.com.

[David Snyder]

Read the comment before the get_egs.sh script is called: https://github.com/kaldi-asr/kaldi/blob/master/egs/voxceleb/v1/local/nnet3/xvector/tuning/run_xvector_1a.sh#L39

The global-ark-indx corresponds to the archive number (1 through 650 in your case). 

On Wednesday, November 27, 2019 at 12:35:20 PM UTC-5, Bar Madar wrote:

Hey,

Can you please explain to me how to get this information (num of example per archive) from the ranges.* file?

Or maybe link me to documentation about this?

Thanks!

Bar 

‫בתאריך יום ד׳, 27 בנוב׳ 2019 ב-18:06 מאת ‪David Snyder‬‏ <‪david.r...@gmail.com‬‏>:‬

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/d53050e8-a504-436c-9a7a-40e669777ebb%40googlegroups.com.

[David Snyder]

The simplest thing to do is to wc -l <path-to-your-egs>/temp/ranges.* and divide that number by the number of archives, which is 650.

[Bar Madar]

Thanks,

I have 257635 examples per archive...is it ok?

‫בתאריך יום ד׳, 27 בנוב׳ 2019 ב-19:49 מאת ‪David Snyder‬‏ <‪david.ry...@gmail.com‬‏>:‬

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/8333136e-85aa-4a26-a4b2-f9ba15d271a0%40googlegroups.com.

[David Snyder]

Yeah, I think that will be OK. I imagine you should be able to get results that are fairly similar to those in the paper. Keep in mind that we train for more epochs than usual (I think the example uses 6 epochs) and the minibatch size is larger, which seems to help (check the paper for those details). 

On Wednesday, November 27, 2019 at 1:50:20 PM UTC-5, Bar Madar wrote:

Thanks,

I have 257635 examples per archive...is it ok?

‫בתאריך יום ד׳, 27 בנוב׳ 2019 ב-19:49 מאת ‪David Snyder‬‏ <‪david.r...@gmail.com‬‏>:‬

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/8333136e-85aa-4a26-a4b2-f9ba15d271a0%40googlegroups.com.

[David Snyder]

I just checked the code of nnet3-xvector-get-egs.cc, and it looks like it is writing the egs without compression,

which would make them a factor of 4 larger than they need to be.

Dan, is it not sufficient to call the Compress() method on the NnetExample object, as we do here? https://github.com/kaldi-asr/kaldi/blob/master/src/nnet3bin/nnet3-xvector-get-egs.cc#L118

On Sunday, November 24, 2019 at 4:38:26 AM UTC-5, Dan Povey wrote:

I just checked the code of nnet3-xvector-get-egs.cc, and it looks like it is writing the egs without compression,

which would make them a factor of 4 larger than they need to be.

In line 188:

    SequentialBaseFloatMatrixReader feat_reader(feature_rspecifier);

it should be of type  SequentialGeneralMatrixReader.  Other parts of the code would need to use GeneralMatrix as

well.

Look in nnet3-get-egs.cc, where it uses  ExtractRowRangeWithPadding().  It would be necessary to use

that function in nnet3-xvector-get-egs.cc:WriteExamples(), which would of course require figuring out the

exact interface and how to convert the code to use it and have equivalent behavior.

It would be great if you could contribute the code back, if you do this.

Dan

On Sun, Nov 24, 2019 at 5:29 PM Bar Madar <mad...@post.bgu.ac.il> wrote:

Hey, thanks!

I added extra layers due to the article, and changed the frame per iter in get_egs.sh to 160000000 so the number of training archives is around 800.

I am using the 6M voxceleb augmented data to train the nnet.

The problem is that the get_egs.sh is failed after few hours because of out of memory on the disk. I have a SSD hard disk of 4T and before I run the get_egs.sh there are 2.2T free, but before the script is finished, it used all the 2.2T and failed.

Do I need more space for this run? or should I make some changes in the script so it will be more cost effective in memory?

Thanks alot!

Bar 

‫בתאריך יום ה׳, 21 בנוב׳ 2019 ב-4:12 מאת ‪David Snyder‬‏ <‪david.r...@gmail.com‬‏>:‬

Hi Bar,

You will probably find this discussion on Kaldi help to be useful: https://kaldi-asr.org/forums.html?place=msg%2Fkaldi-help%2FWLCtGOaT6Uc%2FjIWHIhF4BAAJ . It's about getting this architecture to work in Kaldi.

Basically, you need to add the extra layers as specified in the paper, and increase the amount of training data. You can do the latter by retaining all of the augmented training data (it will take up a lot of space on disk, and will take awhile to train). Also, it helps if you reduce the archive size. I believe you can do that by reducing the number of frames per iter option in get_egs.sh. As I recall, having  around 800 training archives where each has around 200,000 examples per archive will produce good results. You'll have to play with the options to get_egs.sh to figure out how obtain this. I believe I left comments in the scripts that should provide some guidance. 

Best,

Davd

On Wednesday, November 20, 2019 at 5:25:41 AM UTC-5, Bar Madar wrote:

Hey,

I am reading the paper - "SPEAKER RECOGNITION FOR MULTI-SPEAKER CONVERSATIONS USING X-VECTORS" by David Snyder et al.

The nnet architecture that used in the paper is different from the one that we train on the kaldi recipe "sitw/v2". 

Is there a recipe or code on kaldi project that train the same nnet as shown in the paper?

If not, how should I modify the "run_xvector_1a.sh" to this architecture? just adding layers to the network.xconfig file or should I do more changes on other files?

Thanks,

Bar

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi...@googlegroups.com.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/13bcc4e4-d7e4-4ea3-a0e4-53471ad82aab%40googlegroups.com.

--
Go to http://kaldi-asr.org/forums.html find out how to join
---
You received this message because you are subscribed to the Google Groups "kaldi-help" group.

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi...@googlegroups.com.

[Daniel Povey]

Oh.. you're right!

So the egs are not 4x larger than they should be.

There is a small advantage in doing it the way I said, that you would save time and additional compression inaccuracy from uncompressing and then compressing; but it wouldn't affect the size of the egs.

Dan

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/feb9e11f-2fc4-494d-8cf9-ba17476c4f04%40googlegroups.com.

[Bar Madar]

Hey,

It seems that the training process has finished successfully.

but now I have another problem while I am trying to extract x-vectors to part of the voxceleb data.

This is my x-config:

# The frame-level layers
input dim=23 name=input
relu-batchnorm-layer name=tdnn1 input=Append(-2,-1,0,1,2) dim=512
relu-batchnorm-layer name=tdnn2 dim=512
relu-batchnorm-layer name=tdnn3 input=Append(-2,0,2) dim=512
relu-batchnorm-layer name=tdnn4 dim=512
relu-batchnorm-layer name=tdnn5 input=Append(-3,0,3) dim=512
relu-batchnorm-layer name=tdnn6 dim=512
relu-batchnorm-layer name=tdnn7 input=Append(-4,0,4) dim=512
relu-batchnorm-layer name=tdnn8 dim=512
relu-batchnorm-layer name=tdnn9 dim=512
relu-batchnorm-layer name=tdnn10 dim=1500

# The stats pooling layer. Layers after this are segment-level.
# In the config below, the first and last argument (0, and 10000)
# means that we pool over an input segment starting at frame 0
# and ending at frame 10000 or earlier.  The other arguments (1:1)
# mean that no subsampling is performed.
stats-layer name=stats config=mean+stddev(0:1:1:10000)

# This is where we usually extract the embedding (aka xvector) from.
relu-batchnorm-layer name=tdnn11 dim=512 input=stats

# This is where another layer the embedding could be extracted
# from, but usually the previous one works better.
relu-batchnorm-layer name=tdnn12 dim=512
output-layer name=output include-log-softmax=true dim=7361

And on the extract log I have got Errors like this:

LOG (nnet3-xvector-compute[5.5.0~1-c449]:ExplainWhyNotComputable():nnet-computation-graph.cc:171) *** cindex output(0, 0, 0) is not computable for the following reason: ***
output(0, 0, 0) is kNotComputable, dependencies: tdnn10.affine(0, 0, 0)[kNotComputable], 
tdnn10.affine(0, 0, 0) is kNotComputable, dependencies: tdnn10.affine_input(0, 0, 0)[kNotComputable], 
tdnn10.affine_input(0, 0, 0) is kNotComputable, dependencies: tdnn9.relu(0, 0, 0)[kNotComputable], 
tdnn9.relu(0, 0, 0) is kNotComputable, dependencies: tdnn9.relu_input(0, 0, 0)[kNotComputable], 
tdnn9.relu_input(0, 0, 0) is kNotComputable, dependencies: tdnn9.affine(0, 0, 0)[kNotComputable], 
tdnn9.affine(0, 0, 0) is kNotComputable, dependencies: tdnn9.affine_input(0, 0, 0)[kNotComputable], 
tdnn9.affine_input(0, 0, 0) is kNotComputable, dependencies: tdnn8.relu(0, 0, 0)[kNotComputable], 
tdnn8.relu(0, 0, 0) is kNotComputable, dependencies: tdnn8.relu_input(0, 0, 0)[kNotComputable], 
tdnn8.relu_input(0, 0, 0) is kNotComputable, dependencies: tdnn8.affine(0, 0, 0)[kNotComputable], 
tdnn8.affine(0, 0, 0) is kNotComputable, dependencies: tdnn8.affine_input(0, 0, 0)[kNotComputable], 
tdnn8.affine_input(0, 0, 0) is kNotComputable, dependencies: tdnn7.relu(0, 0, 0)[kNotComputable], 
tdnn7.relu(0, 0, 0) is kNotComputable, dependencies: tdnn7.relu_input(0, 0, 0)[kNotComputable], 
tdnn7.relu_input(0, 0, 0) is kNotComputable, dependencies: tdnn7.affine(0, 0, 0)[kNotComputable], 
tdnn7.affine(0, 0, 0) is kNotComputable, dependencies: tdnn7.affine_input(0, 0, 0)[kNotComputable], 
tdnn7.affine_input(0, 0, 0) is kNotComputable, dependencies: tdnn6.relu(0, -4, 0)[kNotComputable], tdnn6.relu(0, 0, 0)[kNotComputable]tdnn6.relu(0, 4, 0)[kNotComputable], 
tdnn6.relu(0, -4, 0) is kNotComputable, dependencies: tdnn6.relu_input(0, -4, 0)[kNotComputable], 
tdnn6.relu(0, 0, 0) is kNotComputable, dependencies: tdnn6.relu_input(0, 0, 0)[kNotComputable], 
tdnn6.relu(0, 4, 0) is kNotComputable, dependencies: tdnn6.relu_input(0, 4, 0)[kNotComputable], 
tdnn6.relu_input(0, -4, 0) is kNotComputable, dependencies: tdnn6.affine(0, -4, 0)[kNotComputable], 
tdnn6.relu_input(0, 0, 0) is kNotComputable, dependencies: tdnn6.affine(0, 0, 0)[kNotComputable], 
tdnn6.relu_input(0, 4, 0) is kNotComputable, dependencies: tdnn6.affine(0, 4, 0)[kNotComputable], 
tdnn6.affine(0, -4, 0) is kNotComputable, dependencies: tdnn6.affine_input(0, -4, 0)[kNotComputable], 
tdnn6.affine(0, 0, 0) is kNotComputable, dependencies: tdnn6.affine_input(0, 0, 0)[kNotComputable], 
tdnn6.affine(0, 4, 0) is kNotComputable, dependencies: tdnn6.affine_input(0, 4, 0)[kNotComputable], 
tdnn6.affine_input(0, -4, 0) is kNotComputable, dependencies: tdnn5.relu(0, -4, 0)[kNotComputable], 
tdnn6.affine_input(0, 0, 0) is kNotComputable, dependencies: tdnn5.relu(0, 0, 0)[kNotComputable], 
tdnn6.affine_input(0, 4, 0) is kNotComputable, dependencies: tdnn5.relu(0, 4, 0)[kNotComputable], 
tdnn5.relu(0, -4, 0) is kNotComputable, dependencies: tdnn5.relu_input(0, -4, 0)[kNotComputable], 
tdnn5.relu(0, 0, 0) is kNotComputable, dependencies: tdnn5.relu_input(0, 0, 0)[kNotComputable], 
tdnn5.relu(0, 4, 0) is kNotComputable, dependencies: tdnn5.relu_input(0, 4, 0)[kNotComputable], 
tdnn5.relu_input(0, -4, 0) is kNotComputable, dependencies: tdnn5.affine(0, -4, 0)[kNotComputable], 
tdnn5.relu_input(0, 0, 0) is kNotComputable, dependencies: tdnn5.affine(0, 0, 0)[kNotComputable], 
tdnn5.relu_input(0, 4, 0) is kNotComputable, dependencies: tdnn5.affine(0, 4, 0)[kNotComputable], 
tdnn5.affine(0, -4, 0) is kNotComputable, dependencies: tdnn5.affine_input(0, -4, 0)[kNotComputable], 
tdnn5.affine(0, 0, 0) is kNotComputable, dependencies: tdnn5.affine_input(0, 0, 0)[kNotComputable], 
tdnn5.affine(0, 4, 0) is kNotComputable, dependencies: tdnn5.affine_input(0, 4, 0)[kNotComputable], 
tdnn5.affine_input(0, -4, 0) is kNotComputable, dependencies: tdnn4.relu(0, -7, 0)[kNotComputable], tdnn4.relu(0, -4, 0)[kNotComputable]tdnn4.relu(0, -1, 0)[kNotComputable], 
tdnn5.affine_input(0, 0, 0) is kNotComputable, dependencies: tdnn4.relu(0, -3, 0)[kNotComputable], tdnn4.relu(0, 0, 0)[kNotComputable]tdnn4.relu(0, 3, 0)[kNotComputable], 
tdnn5.affine_input(0, 4, 0) is kNotComputable, dependencies: tdnn4.relu(0, 1, 0)[kNotComputable], tdnn4.relu(0, 4, 0)[kNotComputable]tdnn4.relu(0, 7, 0)[kNotComputable], 
tdnn4.relu(0, -7, 0) is kNotComputable, dependencies: tdnn4.relu_input(0, -7, 0)[kNotComputable], 
tdnn4.relu(0, -4, 0) is kNotComputable, dependencies: tdnn4.relu_input(0, -4, 0)[kNotComputable], 
tdnn4.relu(0, -1, 0) is kNotComputable, dependencies: tdnn4.relu_input(0, -1, 0)[kNotComputable], 
tdnn4.relu(0, -3, 0) is kNotComputable, dependencies: tdnn4.relu_input(0, -3, 0)[kNotComputable], 
tdnn4.relu(0, 0, 0) is kNotComputable, dependencies: tdnn4.relu_input(0, 0, 0)[kNotComputable], 
tdnn4.relu(0, 3, 0) is kNotComputable, dependencies: tdnn4.relu_input(0, 3, 0)[kNotComputable], 
tdnn4.relu(0, 1, 0) is kNotComputable, dependencies: tdnn4.relu_input(0, 1, 0)[kNotComputable], 
tdnn4.relu(0, 4, 0) is kComputable, dependencies: tdnn4.relu_input(0, 4, 0), 
tdnn4.relu(0, 7, 0) is kComputable, dependencies: tdnn4.relu_input(0, 7, 0), 
tdnn4.relu_input(0, -7, 0) is kNotComputable, dependencies: tdnn4.affine(0, -7, 0)[kNotComputable], 
tdnn4.relu_input(0, -4, 0) is kNotComputable, dependencies: tdnn4.affine(0, -4, 0)[kNotComputable], 
tdnn4.relu_input(0, -1, 0) is kNotComputable, dependencies: tdnn4.affine(0, -1, 0)[kNotComputable], 
tdnn4.relu_input(0, -3, 0) is kNotComputable, dependencies: tdnn4.affine(0, -3, 0)[kNotComputable], 
tdnn4.relu_input(0, 0, 0) is kNotComputable, dependencies: tdnn4.affine(0, 0, 0)[kNotComputable], 
tdnn4.relu_input(0, 3, 0) is kNotComputable, dependencies: tdnn4.affine(0, 3, 0)[kNotComputable], 
tdnn4.relu_input(0, 1, 0) is kNotComputable, dependencies: tdnn4.affine(0, 1, 0)[kNotComputable], 
tdnn4.affine(0, -7, 0) is kNotComputable, dependencies: tdnn4.affine_input(0, -7, 0)[kNotComputable], 
tdnn4.affine(0, -4, 0) is kNotComputable, dependencies: tdnn4.affine_input(0, -4, 0)[kNotComputable], 
tdnn4.affine(0, -1, 0) is kNotComputable, dependencies: tdnn4.affine_input(0, -1, 0)[kNotComputable], 
tdnn4.affine(0, -3, 0) is kNotComputable, dependencies: tdnn4.affine_input(0, -3, 0)[kNotComputable], 
tdnn4.affine(0, 0, 0) is kNotComputable, dependencies: tdnn4.affine_input(0, 0, 0)[kNotComputable], 
tdnn4.affine(0, 3, 0) is kNotComputable, dependencies: tdnn4.affine_input(0, 3, 0)[kNotComputable], 
tdnn4.affine(0, 1, 0) is kNotComputable, dependencies: tdnn4.affine_input(0, 1, 0)[kNotComputable], 
tdnn4.affine_input(0, -7, 0) is kNotComputable, dependencies: tdnn3.relu(0, -7, 0)[kNotComputable], 
tdnn4.affine_input(0, -4, 0) is kNotComputable, dependencies: tdnn3.relu(0, -4, 0)[kNotComputable], 
tdnn4.affine_input(0, -1, 0) is kNotComputable, dependencies: tdnn3.relu(0, -1, 0)[kNotComputable], 
tdnn4.affine_input(0, -3, 0) is kNotComputable, dependencies: tdnn3.relu(0, -3, 0)[kNotComputable], 
tdnn4.affine_input(0, 0, 0) is kNotComputable, dependencies: tdnn3.relu(0, 0, 0)[kNotComputable], 
tdnn4.affine_input(0, 3, 0) is kNotComputable, dependencies: tdnn3.relu(0, 3, 0)[kNotComputable], 
tdnn4.affine_input(0, 1, 0) is kNotComputable, dependencies: tdnn3.relu(0, 1, 0)[kNotComputable], 
tdnn3.relu(0, -7, 0) is kNotComputable, dependencies: tdnn3.relu_input(0, -7, 0)[kNotComputable], 
tdnn3.relu(0, -4, 0) is kNotComputable, dependencies: tdnn3.relu_input(0, -4, 0)[kNotComputable], 
tdnn3.relu(0, -1, 0) is kNotComputable, dependencies: tdnn3.relu_input(0, -1, 0)[kNotComputable], 
tdnn3.relu(0, -3, 0) is kNotComputable, dependencies: tdnn3.relu_input(0, -3, 0)[kNotComputable], 
tdnn3.relu(0, 0, 0) is kNotComputable, dependencies: tdnn3.relu_input(0, 0, 0)[kNotComputable], 
tdnn3.relu(0, 3, 0) is kNotComputable, dependencies: tdnn3.relu_input(0, 3, 0)[kNotComputable], 
tdnn3.relu(0, 1, 0) is kNotComputable, dependencies: tdnn3.relu_input(0, 1, 0)[kNotComputable], 
tdnn3.relu_input(0, -7, 0) is kNotComputable, dependencies: tdnn3.affine(0, -7, 0)[kNotComputable], 
tdnn3.relu_input(0, -4, 0) is kNotComputable, dependencies: tdnn3.affine(0, -4, 0)[kNotComputable], 
tdnn3.relu_input(0, -1, 0) is kNotComputable, dependencies: tdnn3.affine(0, -1, 0)[kNotComputable], 
tdnn3.relu_input(0, -3, 0) is kNotComputable, dependencies: tdnn3.affine(0, -3, 0)[kNotComputable], 
tdnn3.relu_input(0, 0, 0) is kNotComputable, dependencies: tdnn3.affine(0, 0, 0)[kNotComputable], 
tdnn3.relu_input(0, 3, 0) is kNotComputable, dependencies: tdnn3.affine(0, 3, 0)[kNotComputable], 
tdnn3.relu_input(0, 1, 0) is kNotComputable, dependencies: tdnn3.affine(0, 1, 0)[kNotComputable], 
tdnn3.affine(0, -7, 0) is kNotComputable, dependencies: tdnn3.affine_input(0, -7, 0)[kNotComputable], 
tdnn3.affine(0, -4, 0) is kNotComputable, dependencies: tdnn3.affine_input(0, -4, 0)[kNotComputable], 
tdnn3.affine(0, -1, 0) is kNotComputable, dependencies: tdnn3.affine_input(0, -1, 0)[kNotComputable], 
tdnn3.affine(0, -3, 0) is kNotComputable, dependencies: tdnn3.affine_input(0, -3, 0)[kNotComputable], 
tdnn3.affine(0, 0, 0) is kNotComputable, dependencies: tdnn3.affine_input(0, 0, 0)[kNotComputable], 
tdnn3.affine(0, 3, 0) is kNotComputable, dependencies: tdnn3.affine_input(0, 3, 0)[kNotComputable], 
tdnn3.affine(0, 1, 0) is kNotComputable, dependencies: tdnn3.affine_input(0, 1, 0)[kNotComputable], 
tdnn3.affine_input(0, -7, 0) is kNotComputable, dependencies: tdnn2.relu(0, -9, 0)[kNotComputable], tdnn2.relu(0, -7, 0)[kNotComputable]tdnn2.relu(0, -5, 0)[kNotComputable], 
tdnn3.affine_input(0, -4, 0) is kNotComputable, dependencies: tdnn2.relu(0, -6, 0)[kNotComputable], tdnn2.relu(0, -4, 0)[kNotComputable]tdnn2.relu(0, -2, 0)[kNotComputable], 
tdnn3.affine_input(0, -1, 0) is kNotComputable, dependencies: tdnn2.relu(0, -3, 0)[kNotComputable], tdnn2.relu(0, -1, 0)[kNotComputable]tdnn2.relu(0, 1, 0)[kNotComputable], 
tdnn3.affine_input(0, -3, 0) is kNotComputable, dependencies: tdnn2.relu(0, -5, 0)[kNotComputable], tdnn2.relu(0, -3, 0)[kNotComputable]tdnn2.relu(0, -1, 0)[kNotComputable], 
tdnn3.affine_input(0, 0, 0) is kNotComputable, dependencies: tdnn2.relu(0, -2, 0)[kNotComputable], tdnn2.relu(0, 0, 0)[kNotComputable]tdnn2.relu(0, 2, 0)[kNotComputable], 
tdnn3.affine_input(0, 3, 0) is kNotComputable, dependencies: tdnn2.relu(0, 1, 0)[kNotComputable], tdnn2.relu(0, 3, 0)[kNotComputable]tdnn2.relu(0, 5, 0)[kNotComputable], 
tdnn3.affine_input(0, 1, 0) is kNotComputable, dependencies: tdnn2.relu(0, -1, 0)[kNotComputable], tdnn2.relu(0, 1, 0)[kNotComputable]tdnn2.relu(0, 3, 0)[kNotComputable], 
tdnn2.relu(0, -9, 0) is kNotComputable, dependencies: tdnn2.relu_input(0, -9, 0)[kNotComputable], 
tdnn2.relu(0, -7, 0) is kNotComputable, dependencies: tdnn2.relu_input(0, -7, 0)[kNotComputable], 
tdnn2.relu(0, -5, 0) is kNotComputable, dependencies: tdnn2.relu_input(0, -5, 0)[kNotComputable], 


ERROR (nnet3-xvector-compute[5.5.0~1-c449]:CreateComputation():nnet-compile.cc:59) Not all outputs were computable, cannot create computation.

Do you know what was wrong?

thanks!

Bar

‫בתאריך יום ד׳, 27 בנוב׳ 2019 ב-21:08 מאת ‪David Snyder‬‏ <‪david.ry...@gmail.com‬‏>:‬

To unsubscribe from this group and stop receiving emails from it, send an email to kaldi-help+...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/f931682f-815d-4a9e-9fce-8f257fb850a7%40googlegroups.com.

[David Snyder]

My guess is that you didn't update the extract.config file. This is the file that tells the nnet3 code which layer we want to extract the embeddings from. In your case, that should be tdnn11.affine (or something like that). 

On Thursday, December 5, 2019 at 6:12:22 AM UTC-5, Bar Madar wrote:

Hey,

‫בתאריך יום ד׳, 27 בנוב׳ 2019 ב-21:08 מאת ‪David Snyder‬‏ <‪david.r...@gmail.com‬‏>:‬

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/f931682f-815d-4a9e-9fce-8f257fb850a7%40googlegroups.com.