Temporary Wrappers to fix MADE

[gmoss13]

Fixes #1394.

The fix here is to add a wrapper around MADE and MADEMoG which just initializes the networks with an additional dummy feature, and replaces the forward call of these methods by first adding a dummy input variable which gets thrown out of the output.

[janfb]

great, thanks for adding this!

can you maybe add a simple test for this patch in general? e.g., conditional density estimation with a linear gaussian test with multiple dimensions? I think this test is missing in nflows anyways so it would be good to have this.

[codecov]

Codecov Report

Attention: Patch coverage is 81.57895% with 7 lines in your changes missing coverage. Please review.

Project coverage is 78.13%. Comparing base (18f92b1) to head (7c7de6c).
Report is 5 commits behind head on main.

Files with missing lines	Patch %	Lines
sbi/utils/nn_utils.py	80.55%	7 Missing ⚠️

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #1398       +/-   ##
===========================================
- Coverage   89.31%   78.13%   -11.19%     
===========================================
  Files         119      119               
  Lines        8779     8891      +112     
===========================================
- Hits         7841     6947      -894     
- Misses        938     1944     +1006     

Flag	Coverage Δ
unittests	78.13% <81.57%> (-11.19%)	⬇️

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

Files with missing lines	Coverage Δ
sbi/neural_nets/net_builders/flow.py	93.12% <100.00%> (ø)
sbi/utils/nn_utils.py	83.01% <80.55%> (-5.22%)	⬇️

... and 39 files with indirect coverage changes

[gmoss13]

great, thanks for adding this!

can you maybe add a simple test for this patch in general? e.g., conditional density estimation with a linear gaussian test with multiple dimensions? I think this test is missing in nflows anyways so it would be good to have this.

I added MADEMoG to the list of SNPE linear_gaussian tests (which are slow but pass locally for me), and the density_estimators tests. This is because MADEMoG is what we actually use as build_made. The custom MADEWrapper is currently not used, but is included in anticipation for #1269. In this case, it should be tests through CategoricalMADE tests.

[janfb]

Great! Makes sense now and the new made test nicely fails without your fix. Thanks! 🙏