nnet2 to nnet3 model converter

[vijayaditya]

There have been multiple requests for a tool to convert nnet2 models into nnet3 models.
There is no such tool currently available. However it would be simple to write and does not even need accessing the C++ parts of the nnet2 or nnet3 code.

The procedure is

1. Convert nnet2 model to text format using nnet-am-copy
2. Isolate the parameter matrices corresponding to components (e.g. NaturalGradientAffineComponent) into separate files
3. Use these files to initialize the components in the nnet3 model using the file-name argument in the config string. For this write a nnet3 config which replicates the nnet2 model architecture and nnet3-am-init to initialize this model and write it out.

This would be a simple python script involving some string parsing.

[jfainberg]

I'll start working on this (unless anyone else already is).

[vijayaditya]

Please go ahead, no one has taken this yet.
Btw it would be helpful if you discuss your algorithm before implementation
as some changes can be a bit involved e.g. SpliceComponent to Descriptor
conversion.

Vijay

On Aug 26, 2016 05:19, "Joachim" notifications@github.com wrote:

I'll start working on this (unless anyone else already is).

—
You are receiving this because you authored the thread.
Reply to this email directly, view it on GitHub
#886 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/ADtwoI5i9u-YxAa4hs1POkdfPkDc5T6Rks5qjq-ogaJpZM4JHUA_
.

[jfainberg]

Ok, this is getting there and I'll make WIP PR soon once it's a bit cleaned up.

I have one question: when initialising affine components with nnet3-init, it expects a matrix with column dim = input_dim + 1 (bias). The isolated parameter matrices from the nnet2 model are split into LinearParameters and Bias. I could either use something like paste to manually append the bias to the linear parameters, or modify ::InitFromConfig to take either matrix or linear_params and bias separately. The first is a bit of a hack, but perhaps ok for the intended use of this?

[danpovey]

Either way is OK with me.

On Tue, Sep 6, 2016 at 10:32 AM, Joachim notifications@github.com wrote:

Ok, this is getting there and I'll make WIP PR soon once it's a bit
cleaned up.

I have one question: when initialising affine components with nnet3-init,
it expects a matrix with column dim = input_dim + 1 (bias). The isolated
parameter matrices from the nnet2 model are split into and . I could either
use something like paste to manually append the bias to the linear
parameters, or modify ::InitFromConfig to take either matrix or
linear_params and bias separately. The first is a bit of a hack, but
perhaps ok for the intended use of this?

—
You are receiving this because you are subscribed to this thread.
Reply to this email directly, view it on GitHub
#886 (comment),
or mute the thread
https://github.com/notifications/unsubscribe-auth/ADJVu0Y933zseAiU30gQP0YMofa-AcdMks5qnXmOgaJpZM4JHUA_
.

[yujunlhz]

Dear Kaldi developers:

I'd like to know whether there is any update for this thread?

thanks

[danpovey]

I am not aware of any update but it would be nice of Joachim comes through with this. Actually this is not a priority for me at all because as time goes on, fewer people will be using nnet2 models.

…

On Wed, Apr 19, 2017 at 12:20 PM, yujunlhz ***@***.***> wrote: Dear Kaldi developers: I'd like to know whether there is any update for this thread? thanks — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#886 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVu4ztE68YdO8QQujKBlBLpjS9UxDSks5rxl56gaJpZM4JHUA_> .

[jfainberg]

Sorry, other things took priority and I forgot to come back to this. I have done some work on it; I believe it was nearly ready. I'll have a look at it the coming weekend and make a PR.

[KnowBetterHelps]

I use convert_nnet2_to_nnet3.py script in kaldi-master/egs/librispeech/s5 to transfrom my nnet2 models to nnet3, and the model came from /trunk/egs/librispeech/s5/exp/nnet2_online/nnet_a/final.mdl in http://kaldi-asr.org/downloads/build/6/trunk/egs/librispeech/s5/exp/nnet2_online/nnet_a/. But I got the following error:
ERROR (nnet3-init[5.4]:Check():nnet-nnet.cc:737) Dimension mismatch for network-node fixed-affine1: input-dim 2100 versus component-input-dim 700

seems that the feature dim is not matched, but how can I fixed this problem?
Any reply from you would be great appreciate!

[danpovey]

There is not a lot of reason to use that script-- we may have created it for testing purposes. If you have a nnet2 model it will generally be easier to just use nnet2 tools with it. Dan

…

On Tue, Jun 12, 2018 at 9:40 PM, hyuezhi ***@***.***> wrote: I use convert_nnet2_to_nnet3.py script in *kaldi-master/egs/librispeech/s5* to transfrom my nnet2 models to nnet3, and the model came from */trunk/egs/librispeech/s5/exp/nnet2_online/nnet_a/final.mdl* in *http://kaldi-asr.org/downloads/build/6/trunk/egs/librispeech/s5/exp/nnet2_online/nnet_a/ <http://kaldi-asr.org/downloads/build/6/trunk/egs/librispeech/s5/exp/nnet2_online/nnet_a/>*. But I got the following error: *ERROR (nnet3-init[5.4]:Check():nnet-nnet.cc:737) Dimension mismatch for network-node fixed-affine1: input-dim 2100 versus component-input-dim 700* seems that the feature dim is not matched, but how can I fixed this problem? Any reply from you would be great appreciate! — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#886 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVu82cjZbmwHSD-5emvEu4he9a-ZJzks5t8G2TgaJpZM4JHUA_> .

[KnowBetterHelps]

Thanks for your answer!

but maybe I have the reasons to use the nnet3 model (https://github.com/tbright17/kaldi-dnn-ali-gop), or I got the error like this:
ERROR (compute-dnn-gop[5.4]:ExpectToken():io-funcs.cc:212) Expected token "", got instead ""
more information: compute-dnn-gop --use-gpu=no exp_dnn/gop/tree exp_dnn/gop/final.mdl data_dnn/lang/L.fst 'ark,s,cs:apply-cmvn --norm-means=false --norm-vars=false --utt2spk=ark:data_dnn/eval/split1/1/utt2spk scp:data_dnn/eval/split1/1/cmvn.scp scp:data_dnn/eval/split1/1/feats.scp ark:- |' scp:data_dnn/eval/ivectors_hires/ivector_online.scp 'ark:utils/sym2int.pl --map-oov 2 -f 2- data_dnn/lang/words.txt data_dnn/eval/split1/1/text|' ark,t:exp_dnn/gop/gop.1 ark,t:exp_dnn/gop/align.1 ark,t:exp_dnn/gop/phoneme_ll.1

[danpovey]

Maybe you can ask those people to make available the model they used the tool with.

…

On Tue, Jun 12, 2018 at 10:27 PM, hyuezhi ***@***.***> wrote: Thanks for your answer! but maybe I have the reasons to use the nnet3 model (*https://github.com/tbright17/kaldi-dnn-ali-gop <https://github.com/tbright17/kaldi-dnn-ali-gop>*), or I got the error like this: *ERROR* (compute-dnn-gop[5.4]:ExpectToken():io-funcs.cc:212) Expected token "", got instead "" *more information: compute-dnn-gop --use-gpu=no exp_dnn/gop/tree exp_dnn/gop/final.mdl data_dnn/lang/L.fst 'ark,s,cs:apply-cmvn --norm-means=false --norm-vars=false --utt2spk=ark:data_dnn/eval/split1/1/utt2spk scp:data_dnn/eval/split1/1/cmvn.scp scp:data_dnn/eval/split1/1/feats.scp ark:- |' scp:data_dnn/eval/ivectors_hires/ivector_online.scp 'ark:utils/sym2int.pl <http://sym2int.pl> --map-oov 2 -f 2- data_dnn/lang/words.txt data_dnn/eval/split1/1/text|' ark,t:exp_dnn/gop/gop.1 ark,t:exp_dnn/gop/align.1 ark,t:exp_dnn/gop/phoneme_ll.1* — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#886 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/ADJVu2c0HEF6Nut6t4ZmPFqY1t5_QL7Lks5t8HiXgaJpZM4JHUA_> .

[KnowBetterHelps]

ok, thanks!