Mikel Zhobro

>Why open a separate new PR then? You could have just updated #60. Please do that in the future. But anyway, leave it like this now. It was not possible....

> Btw, I see that you also added the dataset there. Please separate this (always separate things when they are logically separate). And I'm anyway not sure about this. I...

> (I don't understand what's the different between making a path or making an alignment) Naming is bad. `update_for_alignment_dumping` and `update_for_fixed_path_training` would be more exact. > But making (and dumping)...

> This PR still looks very much work-in-progress. Can you mark it as draft, until you consider it ready? For FixedPath training, different datasets, are to be handled differently. For...

> Why is this specific for `DummyDatasetMultipleDataKeys`? Why not for the base class `DummyDataset`? Or even better its base class `GeneratingDataset`. I don't see any reason why not to use...

The dummy dataset above is the same as `DummyDatasetMultipleSequenceLength`, only that I added a `Vocab` for the target labels. > I think also you want to have a separate tool...

I've added the tf implementation of the new `slice_nd2` in `util/basic` and some tests. I haven't yet cleaned the old ones out. Before implementing the RETURNN layer, I wanted to...

> (Just post the error with stack trace.) The [Error strace](https://github.com/rwth-i6/returnn/runs/2930341551) you asked. The thing is that the `slice_nd` we had, expects `start` to have only one batch axis `[B]`,...

[Here](https://distill.pub/2017/ctc/#d-footnote-1:~:text=is%20the-,CTC%20score%20of%20the%20subsequence) you say that alpha is the score of the subsequence Z and alpha is calculated assuming that exact Z as sequence. Wouldn't the calculation of alpha change if we...

@ibrahimLearning @robert-meiner any solution?