Trainer flag overfit_batches does not overwrite train dataloaders shuffle flag

[p-wein]

🐛 Bug

Setting the trainer flag overfit_batches (e.g. =10) does not overwrite the shuffle flag set in the training dataloader, even though the warning reads:
UserWarning: You requested to overfit but enabled training dataloader shuffling. We are turning it off for you.

To Reproduce

Steps to reproduce the behavior:

1. Create lightning module with method train_dataloader with flag shuffle=True:

   def train_dataloader(self) -> loading.DataLoader:
        dataset = ProstateX(train=True)
        batch_transforms, gpu_transforms, sample_transforms = self.get_transformations()
        dataloader = loading.DataLoader(dataset,
                                        batch_size=self.hparams.tr_batch_size,
                                        batch_transforms=batch_transforms,
                                        shuffle=True,
                                        sample_transforms= sample_transforms,
                                        gpu_transforms=gpu_transforms,
                                        pseudo_batch_dim=True,
                                        num_workers=self.hparams.num_workers)
        return dataloader 

( I use a rising dataloader, bug should also occur with pytorch dataloaders though)

1. Create main.py with:

mymodel = model.Model3D(cfg)
trainer = pl.Trainer(gpus=1, precision=16, overfit_batches=10)
trainer.fit(mymodel)`

2. Run main.py
3. Find out that your model does not converge.
4. set shuffle=False when creating Dataloader in train_dataloader
5. See that your model converges after some epochs.

(Or log the samples loaded by the dataloader and check if they are the same each epoch.)

Code sample

Expected behavior

Either model also converges with shuffle=True, since warning says that it got overwritten (assuming model converges with shuffle=False) or at least warning should read that user has to change shuffle to False.

Environment

• CUDA:
  - GPU:
  - GeForce GTX 1080 Ti
  - available: True
  - version: 10.1
• Packages:
  - numpy: 1.19.0
  - pyTorch_debug: False
  - pyTorch_version: 1.7.0.dev20200705+cu101
  - pytorch-lightning: 0.8.5
  - tensorboard: 2.2.2
  - tqdm: 4.47.0
• System:
  - OS: Linux
  - architecture:
  - 64bit
  -
  - processor: x86_64
  - python: 3.7.7
  - version: docs: enable syntax highlight #109-Ubuntu SMP Fri Jun 19 11:33:10 UTC 2020

Additional context

[github-actions]

Hi! thanks for your contribution!, great first issue!

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[denck007]

I am seeing the same issue when using --overfit_pct. From a comment in the code, I believe that option is to be removed in 1.0.0, but is it worth it to fix it anyways? The same code will do fix the issue just checking self.overfit_pct instead.