Make 'RestrictedTransformForConditional' conform to 'torch.Transform' interface

[Baschdl]

RestrictedTransformForConditional is currently typed as torch.Transform having a transform as variable:

sbi/sbi/utils/conditional_density_utils.py

Lines 386 to 409 in bae6994

	class RestrictedTransformForConditional(torch_tf.Transform):
	"""
	Class to restrict the transform to fewer dimensions for conditional sampling.
	The resulting transform transforms only the free dimensions of the conditional.
	Notably, the `log_abs_det` is computed given all dimensions. However, the
	`log_abs_det` stemming from the fixed dimensions is a constant and drops out during
	MCMC.
	All methods work in a similar way: `full_theta`` will first have all entries of the
	`condition` and then override the entries that should be sampled with `theta`. In
	case `theta` is a batch of `theta` (e.g. multi-chain MCMC), we have to repeat
	`theta_condition`` to the match the batchsize.
	This is needed for the the MCMC initialization functions when conditioning and when
	transforming the samples back into the original theta space after sampling.
	"""
	def __init__(
	self,
	transform: torch_tf.Transform,
	condition: Tensor,
	dims_to_sample: List[int],
	) -> None:

The conditioning with theta makes the interface incompatible to torch.Transform as e.g. the normal inv() is called without arguments and our inv(theta) is called with an argument. This example could be fixed by renaming it to restricted_inv(theta).

[gmoss13]

The problem is caused by the fact that RestrictedTransformForConditional is a torch Transform, but also takes a torch Transform as an argument. This can be refactored, but I also think this is not priority for the release milestone as conditional_potential is not used by anything at the moment.

[janfb]

Thanks for the clarification.

Then this is related to the recent question here: #1223 and we will fix this when we fix the general handling of custom potentials.