Albert Zeyer

> > Also, the flag was not removed? > > It was removed. It now only exists within the `OpCodeCompiler` but not within the `NativeCodeCompiler`anymore. Why is that relevant? You...

Sorry, I forgot to answer here. We currently do not have a setup for this published, so if you want to have that right now, it requires some own work....

Some notes about `tf.data` (mostly for myself) (maybe I will move that somewhere later, e.g. to `TFDataPipeline.py`, or to the wiki...): Why to use `tf.data` at all? What were the...

(Note, I assigned you to this issue, but this does not really mean an assignment. This is rather that you might be interested in this, and maybe might to take...

@curufinwe Ha, the `tf.data.Dataset` has this function: ``` def shard(self, num_shards, index): """Creates a `Dataset` that includes only 1/`num_shards` of this dataset. This dataset operator is very useful when running...

[Lingvo](https://github.com/tensorflow/lingvo) does not seem to use `tf.data`. I wonder why. Maybe just because the project is older than `tf.data`. (Maybe @kazuki-irie knows?) ([Question on StackOverflow](https://stackoverflow.com/questions/61958140/why-does-lingvo-not-use-tf-data).)

(Sorry, clicked wrong button. Actually I also wanted to add @patrick-wilken but somehow that's not possible? Maybe because he is not part of the GitHub RETURNN team? I added you...

Btw, update: The plan on the pipeline is mostly finished (see initial comment). Please review. What's missing is mostly the design of how you would configure that in the config....

> Before I start to review your Ideas, I want to clarify if I understood the current situation correctly (for TF backend): > > * Currently only the FeedDictDataProvider is...

Yes, but you are again discussing minor implementation details. I even mentioned this already in the draft ("RETURNN `Dataset` wrapped as a TF `tf.data.Dataset`"). This is really the simple and...