Status of pytorch integration plan

[saurabh vyas]

Hey,

I just wanted to know if you guys, have a expected released day of next version of kaldi which comes with pytorch integration ?

I saw some discussions regarding future version ( kaldi 10 ? ),  and some work on GitHub but I just want to confirm if work is going well 

[Daniel Povey]

It's going ahead but it will likely be at least 3 months.

--
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/CAHqPSsM4abLWdRr-Lj7g8ChF3PiWn1R%3D1xF0Aco6myv-Yxue4A%40mail.gmail.com.

[saurabh vyas]

Thank you Dan, I wish your team the best.

It would certainly be useful to use simplicity of pytorch to implement something like transformer architecture, and compare with current best tdnnf recipe.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/CAEWAuyRZnUWgfmUOSs1B7TZziAFeFvi9NB2xggF%3DxacQsaWLEQ%40mail.gmail.com.

[Arkadi]

Is there anything new about it?

On Tuesday, February 25, 2020 at 2:14:40 PM UTC+2, Dan Povey wrote:

It's going ahead but it will likely be at least 3 months.

On Tue, Feb 25, 2020 at 8:04 PM saurabh vyas <saurabh...@gmail.com> wrote:

Hey,

I just wanted to know if you guys, have a expected released day of next version of kaldi which comes with pytorch integration ?

I saw some discussions regarding future version ( kaldi 10 ? ),  and some work on GitHub but I just want to confirm if work is going well 

--
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]

I think Arkadi may have been talking about next-gen Kaldi.

Arkadi:

After we did some initial work, I decided that the job of wrapping existing Kaldi in Python would be too much, as 

the project is so big and there would be so much to support.  We are going in another direction, with the idea being

to add bits to PyTorch that supports certain kinds of differentiable operations on FSAs.. this will allow various

extensions of CTC-like models that will support things like RNNLMs and RNN-T-like things, and will also

give a natural way to implement towards flexible decoding algorithms for these kinds of models.

See the project 'k2' in my github, https://github.com/danpovey/k2.  We are also trying to figure out how to handle datasets in Python, see  https://github.com/pzelasko/lhotse.  The idea here is not to bundle things any more than necessary, but to

use separate repos for separable pieces.  There will likely eventually be a repo (name TBD) that contains

recipes, with dependencies on all the component repos-- or different repos for different recipes.

So basically it's more ground-up, without taking a dependency on any existing Kaldi code-- but of course, making use of what we've learned while developing Kaldi.

Dan

On Mon, Apr 27, 2020 at 8:35 PM Rishabh Kumar <cyfe...@gmail.com> wrote:

No, But I am writing the code. If you want then u can help me. I would be very thankful to you. Email id : kr.ris...@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/6d5fee5c-8d7c-479b-9e73-17b4a97b7409%40googlegroups.com.

[Rudolf A. Braun]

I thought the plan was to not wrap but replace core kaldi with pytorch, as was done in the pybind11 branch (defining the model completely in pytorch and using pytorch optimizers, so no kaldi components) and where it seemed you had managed to replicate results? Have you stopped working on that?

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/6d5fee5c-8d7c-479b-9e73-17b4a97b7409%40googlegroups.com.

[Daniel Povey]

Yes, that was the plan.  I was not sure that that was the right thing in the long term but pursued it for a few months because I didn't have a better plan and because I needed to gain familiarity with a bunch of things including PyTorch and we needed to have something relevant to work on.

The amount of work to create and maintain the Kaldi+PyTorch would be, I felt, too much (since it would be essentially a superset of Kaldi, and it's structured very differently from PyTorch); and the project was also not that interesting or exciting.  The idea now is to strip it way down, take out the phonetic context dependency from the graph (we can always reintroduce it via the lexicon and in other ways that I have plans for, that don't impact the decoding graph), and restructure it as a constellation of separable Python packages that are reusable for other purposes.  Starting from a blank slate also allows us to ditch any back-compatibility concerns.

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/337ed2d8-4468-4df3-aca0-0b5e3d9120fc%40googlegroups.com.

[jason]

Any updated plan for the next generation kaldi?

[Daniel Povey]

Things are moving forward on a constellation of Python-based projects, with PyTorch as the main engine.

The repos are k2, lhotse...

https://github.com/k2-fsa/k2

https://github.com/lhotse-speech/lhotse

.. and snowfall

https://github.com/k2-fsa/snowfall

which is just a draft of the recipes, icefall

https://github.com/k2-fsa/icefall

... which we are aiming to have an official first version of, around Sep 1st (so ~3 weeks, in time for the Interspeech tutorial).  

Of course things are continuing to develop; it's not as mature as Speechbrain or ESPNet yet.

Dan

Go to http://kaldi-asr.org/forums.html to find out how to join the kaldi-help group

---
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/0edff1cb-aa3e-47b9-a44f-773f2a48dff6n%40googlegroups.com.

[yhi...@gmail.com]

Hi Dan,

I just saw icefall on the web. Do u have a plan to support  TF/Keras in the near future? I guess that k2 may work with TF/Keras, right?

Thanks,

Yasser

[Daniel Povey]

Perhaps eventually, but we would need to spend some time to understand TensorFlow's C++ layer.

To view this discussion on the web visit https://groups.google.com/d/msgid/kaldi-help/31f89261-55b4-4148-bed7-1bfa84773e94n%40googlegroups.com.