For context: This PR makes sbibm compatibe with the latest PyTorch version (1.8) necessary due to changes in their API of transformed distributions.
When using the identity transform in case of automatic_transforms_enabled=False, for SNPE, the support of the resulting TransformedDistribution becomes unbounded in PyTorch, even if the support of base distribution is bounded. This leads to wrong rejection sampling in sequential NPE with bounded support priors.
The PR fixes this by not wrapping the prior as TransformedDistribution in the run.py script.
(This has the additional advantage that the prior is not wrapped and will be recognised "as is" by sbi, which is required for running SNPE with mixture density networks.)
The PR additionally extends the info when using KDE with ABC methods, see #7 .
For context: This PR makes
sbibmcompatibe with the latest PyTorch version (1.8) necessary due to changes in their API of transformed distributions.When using the identity transform in case of
automatic_transforms_enabled=False, forSNPE, the support of the resultingTransformedDistributionbecomes unbounded inPyTorch, even if the support of base distribution is bounded. This leads to wrong rejection sampling in sequential NPE with bounded support priors.The PR fixes this by not wrapping the prior as
TransformedDistributionin therun.pyscript. (This has the additional advantage that theprioris not wrapped and will be recognised "as is" bysbi, which is required for runningSNPEwith mixture density networks.)The PR additionally extends the info when using KDE with ABC methods, see #7 .