Refactor datamodule

[ArneBinder]

This is in preparation for ArneBinder/pytorch-ie-hydra-template-1#1.

Relevant changes:

• add parameters: train_split, val_split, and test_split
• prepare_split defaults to value of train_split
• taskmodule.prepare is only called if stage == "fit" or stage is None
• since all constructor parameters for pytorch_lightning.DataModule are deprecated, pass remaining keyword arguments passed to __init__ to the DataLoaders
• simplifies the setup method a bit
• add simple PIEDatasetDict = Dict[Union[str, Split], List[Document]], in analogy to huggingface DatasetDict, to the data.datasets package

[ArneBinder]

Honestly, this is beyond confusing. Why not just use either train_sizeor val_size, which is Union[int, float]? If you want to limit the number of training or validation examples, pytorch-lightning provides all the functionality.

I'm not so much experienced with pytorch-lightning yet, how would it work with that? E.g. if I want to have 5% or 3 examples from my train data each for training and for validation?

[ArneBinder]

And yes, I also don't like the random_train_val_split parameter.

[ChristophAlt]

Honestly, this is beyond confusing. Why not just use either train_sizeor val_size, which is Union[int, float]? If you want to limit the number of training or validation examples, pytorch-lightning provides all the functionality.

I'm not so much experienced with pytorch-lightning yet, how would it work with that? E.g. if I want to have 5% or 3 examples from my train data each for training and for validation?

One way to do this is described here -- you can limit the number of batches by either providing an int (num batches) or float (fraction of batches).

[ArneBinder]

So we have just a parameter val_size to split some part from the train data?

[ChristophAlt]

And yes, I also don't like the random_train_val_split parameter.

This will be done by the PIE Dataset / DatasetDict in the future, as it will provide the same interface as the HF implementation (see here).

[ArneBinder]

And yes, I also don't like the random_train_val_split parameter.

This will be done by the PIE Dataset / DatasetDict in the future, as it will provide the same interface as the HF implementation (see here).

OK, this really depends on which level the PIE Dataset / DatasetDict lives on. Is it about Documents or training instances (TaskEncodings)? Until now I thought a PIE Dataset is about Documents. EDIT: Note that all the logic related to random_train_val_split is about training instances (TaskEncodings).

[ArneBinder]

@ChristophAlt can you have a look again?