Refactor nsf

[janfb]

refactoring the usage of nsf. Importantly, in case of a one-dimensional x we are currently warning that in this case it wouldn't help to have more than one transform.
However, I think it still makes sense to have multiple transforms. In this case the spline parameters are learned from the conditioning variables only, because there is no "other" dimension in x to learn from. But we can still use multiple transforms to stack conditioners and get more flexibility.

Maybe I am missing something here? But in my current view this makes sense and we should remove the warning.

[codecov-commenter]

Codecov Report

Merging #537 (d2557b2) into main (4be9fa2) will decrease coverage by 0.02%.
The diff coverage is 32.25%.

@@            Coverage Diff             @@
##             main     #537      +/-   ##
==========================================
- Coverage   66.80%   66.77%   -0.03%     
==========================================
  Files          67       67              
  Lines        4187     4199      +12     
==========================================
+ Hits         2797     2804       +7     
- Misses       1390     1395       +5     

Flag	Coverage Δ
unittests	66.77% <32.25%> (-0.03%)	⬇️

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

Impacted Files	Coverage Δ
sbi/neural_nets/flow.py	41.09% <32.25%> (+3.39%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 4be9fa2...d2557b2. Read the comment docs.

[michaeldeistler]

Thanks for the refactor!

Regarding the warning: I think you're right, stacking multiple rational-quadratics will be non-rational-quadratic, so it might be more flexible. Out of interest: do you have empirical evidence that it improves performance?

[janfb]

thanks for the review @michaeldeistler and sorry, I extended the refactor just now.

Regarding empirical evidence, yes! when optimising the MNLE architectures we found that more transforms performed much better. Of course they were slower too, so we found a good trade-off of two transforms in the end.

[michaeldeistler]

Ah, interesting, thanks!

Feel free to merge once tests are passing.