Implement Row Convolution in PyTorch

[SeanNaren]

This will give us the ability to do a real time model, being able to predict just using forward only RNNs.

[EgorLakomkin]

I wanted to work on this if you do not mind?

[SeanNaren]

Go for it @EgorLakomkin! Let me know if you run into issues etc

[ryanleary]

@EgorLakomkin you may find this helpful: PaddlePaddle/Paddle#2373

[ryanleary]

Make any headway here @EgorLakomkin?

[EgorLakomkin]

Thx for the links @ryanleary ! I will push the vanilla implementation this week

[dlmacedo]

What is a row convolution?

[willfrey]

https://github.com/torch/nn/blob/master/doc/convolution.md#nn.TemporalRowConvolution

This.

[ryanleary]

@EgorLakomkin did you have any luck with this?

[dimatter]

bump!?

[dlmacedo]

bump!?

[jinserk]

Did you see this? Is it different from the DS2's row conv layer?
https://discuss.pytorch.org/t/how-to-use-temporalrowconvolution/4283

[SeanNaren]

It is! Thanks @jinserk that will be super useful.

[EgorLakomkin]

Sorry for late reply, recently been very busy. I am experimenting with convolutional-only acoustic models (like wav2letter) and I was thinking if it would be interesting to include them in this repo though it is not related to deepspeech?

[EgorLakomkin]

Bump! I implemented wav2letter-like architecture recently and it works quite good, but faster than rnn-based models. Do you think it could be a good addition to this repo?

[SeanNaren]

Make it a branch? I'll more than happily take it as a new branch :)

[ryanleary]

It would be cool to eventually support different model types within this repo if the model loading/options etc can be handled in a robust way. There is a lot of boilerplate code here that would be necessary for any speech system, so certainly good to reuse it.

[dlmacedo]

Amazing!!! Em 12 de out de 2017 4:09 PM, "Ryan Leary" <notifications@github.com> escreveu:

…

It would be cool to eventually support different model types within this repo if the model loading/options etc can be handled in a robust way. There is a lot of boilerplate code here that would be necessary for any speech system, so certainly good to reuse it. — You are receiving this because you commented. Reply to this email directly, view it on GitHub <#46 (comment)>, or mute the thread <https://github.com/notifications/unsubscribe-auth/AOTmHN3VnbyP6drvQMSYwarEjeHO2MBrks5srmP7gaJpZM4NQ8vu> .

[SeanNaren]

Added in #180