_final_transformation of FullyCNN not set during initialization

[tztsai]

Directly passing a tensor to FullyCNN() will cause AttributeError: 'FullyCNN' object has no attribute 'final_transformation'. In the test of FullyCNN, the _final_transformation attribute is manually set. Perhaps this line should be moved into the __init__ method of FullyCNN.
Besides, there is a Mixin class related to the final transformation here, but the FullyCNN class does not inherit it. It instead implements the setter and getter methods of final_transformation in itself.

[raehik]

I have a feeling the test is setting a placeholder value for _final_transformation. I think it should be set when configuring the model object. See:

gz21_ocean_momentum/src/gz21_ocean_momentum/trainScript.py

Lines 299 to 314 in 5e5706e

	# Recover the model's class, based on the corresponding CLI parameters
	try:
	models_module = importlib.import_module(model_module_name)
	model_cls = getattr(models_module, model_cls_name)
	except ModuleNotFoundError as e:
	raise type(e)("Could not find the specified module for : " + str(e))
	except AttributeError as e:
	raise type(e)("Could not find the specified model class: " + str(e))
	net = model_cls(datasets[0].n_features, criterion.n_required_channels)
	try:
	transformation_cls = getattr(models.transforms, transformation_cls_name)
	transformation = transformation_cls()
	transformation.indices = criterion.precision_indices
	net.final_transformation = transformation
	except AttributeError as e:
	raise type(e)("Could not find the specified transformation class: " + str(e))

It's a little clunky due to the dynamic module loading, which we could strip and add back only if required (it was not used). Maybe it belongs as an __init__ parameter...?

I guess setting the identity function as the default may be sensible...? But having it a required parameter would seem clearer. @tztsai what do you think?

[tztsai]

I agree that adding final_transformation to the argument list would be clearer, but it needs to be a callable object, so it cannot be directly specified from the CMD arguments. Perhaps a map from names to callable transformations could be added so that the user can provide a string as an argument and it will be mapped to a transformation, e.g. {'identity': lambda x: x}?