Custom batch sampler fails to re-instantiate in `_dataloader_init_kwargs_resolve_sampler`

[Kamichanw]

Outline & Motivation

I created a batch sampler of batches which is used to sample a larger batch. For example, assume we have a PyTorch BatchSampler which yields a batch of batch_size=3. I can use my custom batch sampler whose batch_size=5 to sample 5 times from underlying batch sampler to yield a large batch of batch_size=15.

However, in function _dataloader_init_kwargs_resolve_sampler, it tries to inject a normal sampler (only yield one batch) to my custom batch sampler.

Pitch

I want to know what is the correct approach in my situation.

Additional context

No response

cc @justusschock @awaelchli

[Kamichanw]

I solved this problem with following approach. (Perhaps nobody will meet the same problem as me)

1. Add use_distributed_sampler=False in Trainer when instantiate it.
2. Replace all sampler that yield one batch (such as RandomSampler, SequentialSamper etc) with DistributedSampler when trainer.world_size > 1.

[hanyangyu1021]

我通过以下方法解决了这个问题。（也许没有人会遇到和我一样的问题）

1. use_distributed_sampler=False实例化的时候添加Trainer。
2. 将所有产生一个批次的采样器（例如RandomSampler等SequentialSamper）替换DistributedSampler为trainer.world_size > 1。
   Can you tell me where to inject the customed sampler into the lightning framework. I want to use the distributional sampler and set shuffle to false. But i don't know where to inject. Thanks