expose enable_transform for all methods.

[janfb]

the current API does not allow to not perform sampling in unconstrained space.
In scenarios where one has already transformed parameters to unconstrained space, one workaround is to manually build potential functions and pass them to the posterior object.

But it would be nice to enable this for build_posterior as well, as it offers more flexibility to switch between sampling methods etc.

This PR exposes the enable_transform argument from mcmc_transform at the top level to allow this for all methods.
Note: To keep things parallel for all three methods, this PR reverts #714, which enabled custom transforms for NPE.

[codecov-commenter]

Codecov Report

Merging #756 (b4acede) into main (a240efb) will increase coverage by 0.19%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##             main     #756      +/-   ##
==========================================
+ Coverage   73.87%   74.07%   +0.19%     
==========================================
  Files          78       78              
  Lines        6071     6067       -4     
==========================================
+ Hits         4485     4494       +9     
+ Misses       1586     1573      -13     

Flag	Coverage Δ
unittests	74.07% <100.00%> (+0.19%)	⬆️

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

Impacted Files	Coverage Δ
sbi/inference/posteriors/direct_posterior.py	98.21% <ø> (ø)
sbi/inference/posteriors/importance_posterior.py	29.03% <ø> (ø)
sbi/inference/posteriors/mcmc_posterior.py	82.02% <ø> (ø)
sbi/inference/posteriors/rejection_posterior.py	82.05% <ø> (ø)
sbi/inference/posteriors/vi_posterior.py	78.83% <ø> (ø)
sbi/inference/snle/snle_base.py	91.58% <ø> (ø)
sbi/inference/snre/snre_base.py	91.45% <ø> (ø)
sbi/samplers/importance/sir.py	23.07% <ø> (ø)
sbi/utils/posterior_ensemble.py	27.72% <ø> (ø)
sbi/utils/potentialutils.py	100.00% <ø> (ø)
... and 9 more

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

[michaeldeistler]

Thanks for implementing this! I like the idea of exposing enable_transform in likelihood_estimator_based_potential() (etc).

I'm less sure about exposing it also in build_posterior() (instead of just leaving it as a feature of the sampler interface), but I could see pros and cons for this so I'm fine with merging this.

[janfb]

I see your point and agree, we can expose it only in the sampler interface. build_posterior will then default to enable_transform=True.
The DirectPosterior still needs the enable_transform arg in the constructor because it builds the potential function internally and it should be possible to turn that off.

I encountered the additional problem that TransformedDistribution would result in an AttributeError in mcmc_transform, which is fixed now as well. So in that case it is possible to still use the convenience of build_posterior.

[michaeldeistler]

LGTM, thanks!