Albert Zeyer

According to [the docs](https://cloud.google.com/dataflow/docs/reference/pipeline-options), there are also these options: * `max_num_workers` * `number_of_worker_harness_threads` * `num_workers`

I agree on `window` and `context_window`. `chunking` is a bit special. But it might anyway be decoupled from the dataset (#376). I would not change `chunking` from a user point...

> [HDF] can't be done via command line This is also slightly wrong (I think). I think you can just specify the dict as a string on command line (or...

> I've never used windowing and chunking, but the argument might be similar, right? Or are those things that would not make sense to be configured differently in different datasets?...

> > No, this is not similar for windowing. Windowing changes the dimension, so it must be consistent. > > Ok, makes sense then to have it global. No, this...

> > And it also make sense to logically decouple this > > But this is the problem here, the parameter "batching" has nothing to do with "batching" (in the...

> > You can easily have different options for train/dev/eval. > > Then I don't understand what you mean. Like having `train_seq_ordering`, `eval_seq_ordering` etc. as global parameters? Yes, just like...

> > I'm just arguing/discussing on the question here whether this should be an option to the dataset (in the train/dev dict in the config), or a separate option. Nothing...

I also tagged you as you might know people or setups which make use of it. Or not. I just wanted to have some confirmation.

See failing test, i.e. code style.