Make the framework using a consistent strategy when lazy-loading val_set

[WenjieDu]

PyPOTS/pypots/imputation/saits/model.py

Lines 253 to 277 in d457629

	if val_set is not None:
	if isinstance(val_set, str):
	with h5py.File(val_set, "r") as hf:
	# Here we read the whole validation set from the file to mask a portion for validation.
	# In PyPOTS, using a file usually because the data is too big. However, the validation set is
	# generally shouldn't be too large. For example, we have 1 billion samples for model training.
	# We won't take 20% of them as the validation set because we want as much as possible data for the
	# training stage to enhance the model's generalization ability. Therefore, 100,000 representative
	# samples will be enough to validate the model.
	val_set = {
	"X": hf["X"][:],
	"X_intact": hf["X_intact"][:],
	"indicating_mask": hf["indicating_mask"][:],
	}
	# check if X_intact contains missing values
	if np.isnan(val_set["X_intact"]).any():
	val_set["X_intact"] = np.nan_to_num(val_set["X_intact"], nan=0)
	logger.warning(
	"X_intact shouldn't contain missing data but has NaN values. "
	"PyPOTS has imputed them with zeros by default to start the training for now. "
	"Please double-check your data if you have concerns over this operation."
	)
	val_set = BaseDataset(val_set, return_labels=False, file_type=file_type)

E.g. in imputation models, when val_set is set as an h5 file path to enable lazy loading, the framework still loads all data from the file. Although people usually don't have a large validation set, this may increase memory pressure if so. And we also expect the framework to have a consistent behavior with train_set. Therefore we need to make PyPOTS to apply the same strategy on val_set.