Albert Zeyer

Please clean up the commits (rebase).

It seems this is heavily copied from `ExternSprintDataset._read_next_raw` (but code comments stripped away?), and now you would end up with three (almost? or even exact?) duplicate copies of the same...

@Marvin84 or @ZhouW321 (or whoever else feels responsible for this, maybe @curufinwe ), can you clean this up and make ready for merge? We should not keep ugly custom patches...

Is this still needed, or outdated now? @JackTemaki ?

Basically, you just need to put `use_tensorflow = True` in your config. Note that not all layers are implemented currently in the TF backend. So far, the linear/forward layer and...

So [that config file](https://github.com/rwth-i6/returnn/blob/master/demos/mdlstm/IAM/config_real) I guess. Most of the layers are not yet available in TF, i.e. there is no 1Dto2D, conv2, mdlstm. 1Dto2D and conv2 should be easy to...

@lannp Where are you looking at? [Here](https://github.com/rwth-i6/returnn/tree/master/demos/mdlstm/IAM) are configs for training (`config_real`) as well as forwarding (`config_fwd`). Anyway, TensorFlow is not supported at all right now for MDLSTM. Feel free...

This is not about any or the official pure TF implementation. This is about our own CUDA kernel, which should be much faster.

@curufinwe maybe you know some more about this? Or @JackTemaki? Or who else could know more?

Some research should be made on this, and also some benchmarking.