Cases where the length of the IterableDataset is not overriden properly

[MaksymDel]

_LazyInstances (allennlp's subclass of IterableDataset) returns 1 when the __len__ method is not overridden.
This, for example, causes a problem with SlantedTriangular which expects the length of the dataset in batches as constructor parameter if used programmatically.

See #4028 (comment) and #4028 (comment) for more context.

[dirkgr]

It is baffling behavior to return 1 for the length. We should return None for the length when we don't know, and deal with the fallout wherever it happens.

[JohnGiorgi]

@maksym-del If I did want to use SlantedTriangular while lazy=True in my DatasetReader, would I just manually set dataloader.batches_per_epoch in my config?

(not just asking to be annoying, I actually do want to use SlantedTriangular with a lazy reader).

[MaksymDel]

@JohnGiorgi yes, specifying batches per epoch in loader and passsing corresponding number to the scheduler should work.

[JohnGiorgi]

@maksym-del Awesome, thanks!