allow more flexible proposals in atomic snpe-c

[michaeldeistler]

if one used, e.g., an MCMCPosterior, or simply a torch.distributions.MultivariateNormal, one would get an error that the propsals has no attribute posterior_estimator. This is fixed here.

Resolves #706

[codecov-commenter]

Codecov Report

Merging #732 (341ac90) into main (b7f0e9e) will not change coverage.
The diff coverage is n/a.

@@           Coverage Diff           @@
##             main     #732   +/-   ##
=======================================
  Coverage   74.25%   74.25%           
=======================================
  Files          79       79           
  Lines        6033     6033           
=======================================
  Hits         4480     4480           
  Misses       1553     1553           

Flag	Coverage Δ
unittests	74.25% <ø> (ø)

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

Impacted Files	Coverage Δ
sbi/inference/snpe/snpe_c.py	75.36% <ø> (ø)

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

[janfb]

how does this check change the behaviour for MCMCPosterior or MultivariateNormal?

[michaeldeistler]

previously, the first check was isinstance(proposal.posterior_estimator._distribution, mdn). This raised an error if, e.g., proposal was an MCMCPosterior.

Now, the first check is isinstance(proposal, DirectPosterior) which will be False if proposal is, e.g., an MCMCPosterior. Because everything below is an and case, these lines will never be executed.

[michaeldeistler]

See https://stackoverflow.com/questions/13960657/does-python-evaluate-ifs-conditions-lazily

[janfb]

the Python behaviour is clear to me, I was wondering about the sbi semantic. But true, we want the line to be False for MCMCPosterior and other, instead of throwing an error. so all good!

[janfb]

the Python behaviour is clear to me, I was wondering about the sbi semantic. But true, we want the line to be False for MCMCPosterior and other, instead of throwing an error. so all good!