Florian Boyer

`enc-init-mods`, `dec-init-mods`, `freeze-mods` are defined like that in ESPnet1, for all tasks. It's not a bug in regards to the documentation (we say "separated by a comma" only) but I...

He was talking about [this line in run.sh](https://github.com/supernaiter/cameraspeller_egs2/blob/main/run.sh#L24) which means you used [this yaml config](https://github.com/supernaiter/cameraspeller_egs2/blob/main/conf/train_asr_transformer.yaml) for your training. However, it's not equivalent to what is reported here. For example, `num_blocks`...

@kamo-naoyuki There is another docstring style I'm using which was used at one point (dunno why we discarded it?): ``` """time-frequency L1 loss. Arguments: ref (torch.Tensor): (B, T, F) or...

Hi, > Hi there, not sure why you use for loop over batch for memory reduction? Is this tradeoff between speed and memory? It is, it shouldn't impact training speed...

It is on my to do list but I can't give an estimated time yet.

Hi, > I have implemented the section 3.1 with CUDA binding and make better perf. Nice! Incorporating the approach into our `warp-transducer` fork was the next step but it's further...

Thank you for the benchmark! It seems speed/memory consumption trade-off is all over the place between the implementation :smile: I'll take a look at your implementation when I have time....

> @b-flo that is what I mentioned. do you think how merge into your part? Sorry, I totally forgot this PR until I received emails about your last commits... I'm...

Hi, Looking at the [tutorial for fuse modules](https://pytorch.org/tutorials/recipes/fuse.html) and docs, the second parameter for `fuse_modules` is a "list of list of module names (to fuse)". Also, there are no modules...

> Could you please help me with a little more detail on where I am going to add the code? Hum, I'm not really sure what's the better approach for...