Difference between the dataloader of training and precise bn

Hi author,

I have a question about the precise_bn used in multi-grid training.

According to the code, the dataloader used in multigrid training with both short-cycle and long-cycle is different from that of the precise bn.

```
if cfg.MULTIGRID.SHORT_CYCLE and split in ["train"] and not is_precise_bn:
# Create a sampler for multi-process training
sampler = utils.create_sampler(dataset, shuffle, cfg)
batch_sampler = ShortCycleBatchSampler(
sampler, batch_size=batch_size, drop_last=drop_last, cfg=cfg
)
# Create a loader
loader = torch.utils.data.DataLoader(
dataset,
batch_sampler=batch_sampler,
num_workers=cfg.DATA_LOADER.NUM_WORKERS,
pin_memory=cfg.DATA_LOADER.PIN_MEMORY,
worker_init_fn=utils.loader_worker_init_fn(dataset),
)
else:
# Create a sampler for multi-process training
sampler = utils.create_sampler(dataset, shuffle, cfg)
# Create a loader
loader = torch.utils.data.DataLoader(
dataset,
batch_size=batch_size,
shuffle=(False if sampler else shuffle),
sampler=sampler,
num_workers=cfg.DATA_LOADER.NUM_WORKERS,
pin_memory=cfg.DATA_LOADER.PIN_MEMORY,
drop_last=drop_last,
collate_fn=detection_collate if cfg.DETECTION.ENABLE else None,
worker_init_fn=utils.loader_worker_init_fn(dataset),
)
```

Therefore, the samples for training at each epoch is different from that for precise bn. The dataloader for training has a short-cycle sampler which varies the spatial size. Will this difference cause convergence loss since the stats for each bn is also not 'precise'.

Thanks!