Allow shuffle when auto-batching disabled in DataLoader

[huww98]

🚀 Feature

Allow DataLoader(batch_size=None, shuffle=True)

Motivation

My model handles one sample a time. So, I don't need auto-batching, but I want my dataset to be shuffled. However, when I use DataLoader(batch_size=None, shuffle=True), it raises batch_size=None option disables auto-batching and is mutually exclusive with shuffle, and drop_last.

But I can't see why they are mutually exclusive.

Pitch

Allow DataLoader(batch_size=None, shuffle=True).

Basically, just remove the shuffle check from this line:

pytorch/torch/utils/data/dataloader.py

Line 202 in ada6472

if shuffle or drop_last:

Alternatives

I can:

• Use batch_size=1 then remove the batching dimension afterwards
• Use sampler=RandomSampler(dataset)

But these makes me feel like I'm fighting with DataLoader. Why I cannot just use shuffle=True?

Additional context

cc @ssnl

[ydaiming]

Hi, there. If you haven't, I create a PR and work on it.

[ssnl]

@ydaiming Feel free to open a PR and tag me for review.