Why no custom collate?

[alok]

I find them pretty handy in pytorch

[darsnack]

Don't see a reason why there can't be. We'll just need to update BatchViewCollated to accept a user collate function.

[lorenzoh]

As Kyle pointed out, this will not be quite as straightforward if we want to support inplace data loading for custom collate functions. Below is a sketch of a possible solution, depending on the use case for it.

Currently a batch is recursively defined as either:

• an AbstractArray with one dimension being the batch size
• a Tuple of batches
• a NamedTuple of batches
• a Dict of batches

Importantly, getobs! is a property of the data container, not the BatchViewCollated. Let's say we have a data container DC with observations of type O so we have: getobs(::DC, idx)::O and getobs!(::O, ::DC, idx)::O.

The question is what you want to achieve through a custom colaltefn. If you want to return custom data types as batches, then the following would work:

• have a custom collatefn that returns batches of type B
• define a method of DataLoaders.obsslices(::B, ::DataLoaders.BatchDim) that returns an iterator over views of type O. For example, if O is an array type, then it should return array views.

---

Of course, if we don't want to support buffering and custom collate functions (as is the case in PyTorch if I'm not mistaken), we could simply make buffered and collatefn arguments on DataLoader mutually exclusive.