Running with --eval val still runs evaluation on the test split

[joeloskarsson]

As described in

neural-lam/neural_lam/train_model.py

Lines 163 to 168 in 1c281a2

	parser.add_argument(
	"--eval",
	type=str,
	help="Eval model on given data split (val/test) "
	"(default: None (train model))",
	)

running with ---eval val should evaluate the model on the validation split. This used to be the case, but when the data module was introduced this was apparently changed so that evaluation always happens with the test set

neural-lam/neural_lam/weather_dataset.py

Lines 655 to 663 in 1c281a2

	if stage == "test" or stage is None:
	self.test_dataset = WeatherDataset(
	datastore=self._datastore,
	split="test",
	ar_steps=self.ar_steps_eval,
	standardize=self.standardize,
	num_past_forcing_steps=self.num_past_forcing_steps,
	num_future_forcing_steps=self.num_future_forcing_steps,
	)

This is both an error in documentation, but I think we should keep the possibility to run eval separately for the full validation set, as this is highly useful. Here is a possible fix that allows for this: joeloskarsson@deec455