batch_size control

[davidsebfischer]

Allow user to set batch size. Do this as part of training strategy? You can leave defaults as they are now.

[falexwolf]

We had a long discussion about the notions chunks and batches? In the end, we settled for chunks as this is won't be confused with experimental batches and is analogous with h5py, pytables, xarray and zarr naming conventions on the file level.

So, in anndata and Scanpy, there is a parameter chunk_size, not batch_size... Maybe a consideration...

[davidsebfischer]

Agreed, that makes sense.

[davidsebfischer]

Opened a separate issue for batch vs chunk naming.

[Hoeze]

The user can change the batch size by specifying it in __init__:
Estimator(input_data, batch_size=500)

Sadly, this cannot be changed at runtime, since this would require to edit the tensorflow graph.

Nevertheless, this is an option which should be accessible in diffxpy:
theislab/diffxpy#8

[davidsebfischer]

Yes, if this is set in TrainingStrategy, changes in batch size at run time are not necessary anyway.