Extract_batch_size and Custom Batch Objects That Implement __len__ and __iter__.

[jonathanking]

According to the way extract_batch_size works, I am unable to use a custom batch object with PyTorch Lightning without explicitly passing the length of the batch to every call to self.log() (and there are many such calls). Instead, the following error message appears:

pytorch_lightning.utilities.exceptions.MisconfigurationException: We could not infer the batch_size from the batch. Either simplify its structure or provide the batch_size as self.log(..., batch_size=batch_size).

My custom batch object is an iterable and supports the len() and iter() operations. The batch object yields custom data objects (not tensors) when iterated over, and so PyTorch lightning does not work in this case.

I believe this is something that should be supported by PyTorch Lightning, especially when custom batch objects implement iter() or len(). Do you agree?