Quicker dataset experimentation

[AmitMY]

To catch errors early, I recommend to swap the order of the splits here:

multimodalhugs/multimodalhugs/data/datasets/pose2text.py

Lines 50 to 72 in adcc3ea

	return [
	datasets.SplitGenerator(
	name=datasets.Split.TRAIN,
	gen_kwargs={
	"metafile_path": self.config.train_metadata_dir,
	"split": f"{datasets.Split.TRAIN}"
	}
	),
	datasets.SplitGenerator(
	name=datasets.Split.VALIDATION,
	gen_kwargs={
	"metafile_path": self.config.validation_metadata_dir,
	"split": "val"
	}
	),
	datasets.SplitGenerator(
	name=datasets.Split.TEST,
	gen_kwargs={
	"metafile_path": self.config.test_metadata_dir,
	"split": f"{datasets.Split.TEST}"
	}
	),
	]

validation, test, then train

Since the validation and test sets are smaller, it is in my opinion most useful to have them run first - such that if they fail, we get a signal right away.
For me, the train split takes 8~ hours, and who knows, might fail at the end.