API simplify DensityEstimator base class

[tomMoral]

Simplify the DensityEstimator base class and make it an abstract class.
Remove unnecessary functions

• add Estimator base class
• doc shape conventions
• remove broadcasting stuff from log_prob and sample methods
• add VectorFieldEstimator
• add RatioEstimator
• sync with @michaeldeistler 's Shape conventions for all DensityEstimators #1066

A step in the direction of #1046

[codecov]

Codecov Report

Attention: Patch coverage is 84.84848% with 5 lines in your changes are missing coverage. Please review.

Project coverage is 76.86%. Comparing base (0b5f931) to head (d47f4a7).

❗ Current head d47f4a7 differs from pull request most recent head c6ff223. Consider uploading reports for the commit c6ff223 to get more accurate results

Additional details and impacted files

@@            Coverage Diff             @@
##             main    #1072      +/-   ##
==========================================
- Coverage   85.13%   76.86%   -8.27%     
==========================================
  Files          90       89       -1     
  Lines        6651     6558      -93     
==========================================
- Hits         5662     5041     -621     
- Misses        989     1517     +528     

Flag	Coverage Δ
unittests	76.86% <84.84%> (-8.27%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

Files	Coverage Δ
sbi/inference/snpe/snpe_a.py	65.15% <100.00%> (-25.07%)	⬇️
sbi/neural_nets/density_estimators/nflows_flow.py	98.21% <100.00%> (+35.46%)	⬆️
sbi/neural_nets/density_estimators/zuko_flow.py	97.91% <100.00%> (+33.47%)	⬆️
sbi/neural_nets/density_estimators/base.py	65.00% <66.66%> (+9.44%)	⬆️
sbi/utils/user_input_checks.py	85.36% <62.50%> (+1.60%)	⬆️

... and 51 files with indirect coverage changes

[tomMoral]

Finally it's green ^^

[manuelgloeckler]

Great :). I just checked with @michaeldeistler, who also works on the Density estimator functionality in #1066; so, we will delay merging this until #1066 is done.

[tomMoral]

#1138 conflicts with this PR.
As this PR modifies the base class, merging it fast would be a good idea otherwise we should drop it as it will always need to resolve conflicts.

I am not sure how to solve the conflict here:

• From our discussion during the sprint, I thought the goal was to avoid having constraints on the base class and went on removing the attribute that were not used in it (net/_condition_shape)
• Give all DensityEstimators an input_shape #1138 adds an attribute which is not used in the class.

I think if we have input/condition_shape attributes, it would make sense to add a mechanism to check these shapes in all the class function. something like:

def _check_shape(self, x=None, theta=None):
    if x is not None:
        check_input_shape(x, self._input_shape)
    if theta is not None:
        check_condition_shape(theta, self.condition_shape)

maybe this could also handle some of the reshaping from #1066?

[janfb]

Yes, this has a lot of conflicts with the changes to the shaping we made and we want this PR to be merged asap.

@manuelgloeckler planned to continuing this PR. I think he will either merge this branch into a new feature branch from main, or to start from scratch and cherry pick the changes from this PR.

I see the point about the _check_shape function, but I think we have that already, at least for the condition shape:

sbi/sbi/neural_nets/density_estimators/base.py

Lines 132 to 157 in 005aeac

	def _check_condition_shape(self, condition: Tensor):
	r"""This method checks whether the condition has the correct shape.
	Args:
	condition: Conditions of shape (*batch_shape, *condition_shape).
	Raises:
	ValueError: If the condition has a dimensionality that does not match
	the expected input dimensionality.
	ValueError: If the shape of the condition does not match the expected
	input dimensionality.
	"""
	if len(condition.shape) < len(self.condition_shape):
	raise ValueError(
	f"Dimensionality of condition is to small and does not match the\
	expected input dimensionality {len(self.condition_shape)}, as provided\
	by condition_shape."
	)
	else:
	condition_shape = condition.shape[-len(self.condition_shape) :]
	if tuple(condition_shape) != tuple(self.condition_shape):
	raise ValueError(
	f"Shape of condition {tuple(condition_shape)} does not match the \
	expected input dimensionality {tuple(self.condition_shape)}, as \
	provided by condition_shape. Please reshape it accordingly."
	)

And we also have other functions that are used by the SBI methods to check the correctness of the shape, e.g.,

sbi/sbi/neural_nets/density_estimators/shape_handling.py

Line 53 in 005aeac

def reshape_to_batch_event(theta_or_x: Tensor, event_shape: torch.Size) -> Tensor: