Albert Zeyer

I have put some initial code for a GitHub action using [adobe/sizewatcher](https://github.com/adobe/sizewatcher) but I'm not really sure whether that can work. Some issue: https://github.com/adobe/sizewatcher/issues/79 Maybe this can not really work...

I also asked for such a feature now [in the official GitHub community forum](https://github.com/github-community/community/discussions/18490).

Yes. We could extend the sources.txt by some config about this. Or maybe better: automatically parse that information from the message. And then just avoid having too much connections to...

I assume this is not really fixed? We get the same in `MaxPoolingNoMaskOp`.

@JackTemaki @michelwi I tagged (assigned) you because I want to get some comment by you on this. Do you agree?

I think order 0 is much faster than 1. And I don't know how much the difference is in final WER performance but I would assume none. Note that this...

I have not done systematic comparison experiments, so I don't know. I just assume this.

Maybe before we make the change, we really need to do some testing.

The current goal is that you can run `python3 rnn.py demos/demo-torch.config` and it starts the training.

Note that I created the config option `backend`, so you do `backend = "torch"` in the config to enable the PyTorch backend. I also created an initial `returnn.torch` package/directory, with...