Albert Zeyer

@JackTemaki @patrick-wilken some thoughts? Agree?

> But what about importing models into RASR? Are the exact variable names important there? If you have a checkpoint from old version, and use the compile_tf_graph script from the...

Note, in case we try to address this, i.e. clean up or refactor some of this: We should also check and run the test cases of returnn-common, as there are...

One big problem is that we never really defined the logic around `get_for_batch_ctx` well, and it was added only at a later point. And similarly for some of the other...

Related: - #1153 - #920 - #634 - #589 - #391

A bit related is also dim tag math, and esp questions on equality in those cases. However, I think this was not too much of a problem so far, except...

@Zettelkasten any thoughts or feedback on this?

I wonder whether we should maybe avoid the whole `get_for_batch_ctx` logic completely. To reiterate why we have it: - Automatically carry around a version with beam (tiled) of the `dyn_size_ext`....

@curufinwe might have some experience on this. Do you really need TFLite? Normal TF is not an option? Why? Things like quantization, sparsification, etc, all should work just as well...

Of course, the code itself is one source of documentation, i.e. how the dataset is actually used (in `TFEngine`/`Runner`). (RETURNN is not intended for users who do not want to...