Register Op as subclass of Distributions with rv_type defined

[ricardoV94]

No description provided.

[codecov]

Codecov Report

Merging #6493 (92d62f5) into main (90b6bec) will decrease coverage by 23.00%.
The diff coverage is 88.88%.

Additional details and impacted files

@@             Coverage Diff             @@
##             main    #6493       +/-   ##
===========================================
- Coverage   94.74%   71.74%   -23.00%     
===========================================
  Files         148      148               
  Lines       27578    27790      +212     
===========================================
- Hits        26129    19939     -6190     
- Misses       1449     7851     +6402     

Impacted Files	Coverage Δ
pymc/distributions/distribution.py	96.01% <80.00%> (-1.46%)	⬇️
pymc/tests/distributions/test_distribution.py	98.73% <100.00%> (+0.03%)	⬆️
pymc/sampling_jax.py	0.00% <0.00%> (-100.00%)	⬇️
pymc/tests/sampling/test_jax.py	0.00% <0.00%> (-100.00%)	⬇️
pymc/tests/logprob/test_utils.py	0.00% <0.00%> (-100.00%)	⬇️
pymc/tests/logprob/test_cumsum.py	0.00% <0.00%> (-100.00%)	⬇️
pymc/tests/tuning/test_scaling.py	0.00% <0.00%> (-100.00%)	⬇️
pymc/tests/logprob/test_mixture.py	0.00% <0.00%> (-100.00%)	⬇️
pymc/tests/tuning/test_starting.py	0.00% <0.00%> (-100.00%)	⬇️
pymc/tests/logprob/test_abstract.py	0.00% <0.00%> (-100.00%)	⬇️
... and 72 more

[michaelosthege]

I'm not familiar with __new__ or clsdict. What's the purpose of this change?

One of the test failures will be get fixed by #6475. Just sayin' :)

[larryshamalama]

I'm not familiar with __new__ or clsdict. What's the purpose of this change?

My understanding is that symbolic Ops were not subclasses of Distribution prior to this PR, but it was the case for RV Ops. Is that correct @ricardoV94 ? However, I forgot why it is beneficial to have such Ops to be subclasses of Distribution...

[ricardoV94]

The reason for these changes is to have distributions that return RandomVariables and SymbolicRandomVariables behave the same.

The metaclass does 2 things:

1. Dispatching logp/logcdf/moment methods when implemented as methods of the class
2. Register the Op class as a subclass of the PyMC distribution

Before this PR, it only did those for classes that return RandomVariables Op (as those were all the cases in early V4)

I was recently writing a test in another library where I asserted we were creating the right variables by saying isinstance(rv.owner.op, pm.Normal) but this only works for regular distributions. It doesn't work for Censored for example, in which case I had to import the specific rv_type and do isinstance(rv.owner.op, CensoredRV)

There is no reason for this distinction. As long as we keep the Distribution pseudo-classes they should work the same regardless of the type of objects they return.