Add option to choose dataset shuffle in dataloders

[icedoom888]

Is your feature request related to a problem? Please describe.

Currently the codebase is hardcoded:
training has shuffle= True
validation has shuffle = False
test has shuffle = False

Describe the solution you'd like

These should be extended to allow more flexibility in experiments:

@cached_property
    def ds_train(self) -> MultiDataset:
        """Create multi-dataset for training."""
        return self._get_dataset(
            self.train_dataloader_config,
            shuffle=True,
            label="training",
        )

    @cached_property
    def ds_valid(self) -> MultiDataset:
        """Create multi-dataset for validation."""
        return self._get_dataset(
            self.valid_dataloader_config,
            shuffle=False,
            val_rollout=self.config.dataloader.validation_rollout,
            label="validation",
        )

    @cached_property
    def ds_test(self) -> MultiDataset:
        """Create multi-dataset for testing."""
        return self._get_dataset(self.test_dataloader_config, shuffle=False, label="test")

Describe alternatives you've considered

No response

Additional context

No response

Organisation

No response

[mchantry]

Thanks for the issue @icedoom888 are there specific other configurations that you'd be particularly interested to be able to use?

[icedoom888]

Hey @mchantry, i had this fixed before the multi dataset PR in another branch but now with the new way the schemas are build it doesn't apply anymore, so decided to open an issue to highlight this, in case someone finds another solution!

At some point i found myself in the need of not shuffling the training set: wanted to limit batches to 1 and overfit on one sample for some experiments. Found this limitation ;)

[radiradev]

Why would you want to shuffle validation/test data?

[icedoom888]

One case is when you don't want to run the full val set every time (you could limit with limit_btaches but then always get the same), but still plot some random elements of the val set.

But ultmately, still nice to give the option no?