Issue with RepeatAugSampler: Augmented Versions May End Up in the Same Process

Hello,

I've been working with the RepeatAugSampler, and I have some doubts regarding the following comment in its code:

It ensures that each augmented version of a sample will be visible to a different process (GPU)

From my understanding, this statement does not always hold true—particularly in configurations where num_replicas < num_repeats. For example, if num_replicas = 2 and num_repeats = 3, multiple augmented versions of the same image may end up assigned to the same process. In such cases, these versions can appear consecutively in a batch.

This can become problematic when using augmentation techniques like Mixup or CutMix, as they may end up mixing different augmentations of the same image. I assume this is not ideal, since the intent of these techniques is to mix semantically different samples to improve generalization.

Questions:

1. Is my understanding correct that RepeatAugSampler does not prevent this behavior when num_replicas < num_repeats?
2. If so, is there a recommended strategy to avoid this? One idea I had was to permute the batch after sampling but before applying Mixup/CutMix, in a way that minimizes the chance of mixing augmented versions of the same original sample.

I would appreciate any insights or suggestions!

Thanks!