Start adding GRW RV Op and distribution

[canyon289]

Starting to add time series support to v4 on step at a time (See what I did there?)

Scope of this PR is just to add the random variable, not the distribution as this might take me a while and I'll likely have a lot of basic questions

#5290

[ricardoV94]

Code inside the rng_fn should be pure non-symbolic numpy/python.

Anyway the idea with aeppl is that we don't need to build a new RandomVariable for every possible distribution, but use Aesara directly (as you are doing with that scan) to generate random draws.

Similarly, instead of needing a custom logp method for every distribution we use aeppl to generate the logp graph, given the corresponding Aesara random generating graph.

The thing that is needed is to make pm.GaussianRandomWalk return the right Aesara random generating graph. We still have to figure out the right API for this type of "distributions'. I have a WIP proposal in #5169 under the name of DerivedDistribution (feedback and better name appreciated)

[ricardoV94]

Also to illustrate the point of this approach, the only thing that is needed to create any other random walk is to change the distribution used in that scan graph

rv, updates = aesara.scan(
  fn=lambda prev_value: pm.StudentT.dist(prev_value + mu, sigma, nu),
  outputs_info=np.array(0.0),
  n_steps=size,
)
# ignoring: updates have to be managed differently when using RandomVariables via `.dist`

Instead of having to create them one by one we can offer all at once in a pm.RandomWalk by letting users specify which base distribution to use as one of the arguments.

my_srw = pm.RandomWalk("my_srw", dist=pm.StudentT.dist(mu, sigma, nu), size=10)

pm.GaussianRandomWalk can then be a short convenience function that calls pm.RandomWalk(name, dist=pm.Normal.dist(mu, sigma), size=size) under the hood.

[canyon289]

These are great pointers thank you!

[canyon289]

@ricardoV94, reviewed the other prs and thought through this some. Do you mind if I learn how to implement a "fixed" GRW RV first in this PR, get that working, then follow up with a RW RV generalization? Thatll help me break up unknown bits that I need to learn into smaller unknown bits for me

[ricardoV94]

@ricardoV94, reviewed the other prs and thought through this some. Do you mind if I learn how to implement a "fixed" GRW RV first in this PR, get that working, then follow up with a RW RV generalization? Thatll help me break up unknown bits that I need to learn into smaller unknown bits for me

If you want to take it in small steps I think limiting yourself to the case of Gaussian innovations might help. I would still suggest you sidestep writing the RandomVariable altogether and try to use the WIP DerivedDistribution from my branch.

Otherwise you'll just be practicing writing traditional PyMC distributions. That can be useful for temporary backwards compatibility of V4, but it will have a smaller impact than understanding how to make use of aeppl's flexibility.

[canyon289]

Great thanks for this advice, I'll start working off of that branch

[codecov]

Codecov Report

Merging #5298 (3af1b59) into main (02897d1) will increase coverage by 0.23%.
The diff coverage is 90.76%.

@@            Coverage Diff             @@
##             main    #5298      +/-   ##
==========================================
+ Coverage   88.58%   88.82%   +0.23%     
==========================================
  Files          75       75              
  Lines       13716    13736      +20     
==========================================
+ Hits        12151    12201      +50     
+ Misses       1565     1535      -30     

Impacted Files	Coverage Δ
pymc/distributions/timeseries.py	47.31% <90.76%> (+24.42%)	⬆️

[canyon289]

Why is precommit failing here? Same commit, different results locally than in Ci?

[canyon289]

@ricardoV94 Let me know if you like this GRW RV implementation.

Per our conversation for now we'll

1. Skip init distribution
2. Restrict innovations just to Gaussians to get something working
3. Implement an old style distribution

Next steps for me to to create the GRW Distribution in old style

After thats working I'll work on

1. Switching to Derived dist with more aeppl and aesara integration
2. Generalize distributions
3. Figure out init distribution (if we want it)

[ricardoV94]

Are you planning to do this all in one PR or separate ones?

[canyon289]

Are you planning to do this all in one PR or separate ones?

Separate ones ideally

i updated the comment to make it more clear what my plan is, let me know if you think its a reasonable course of action

[ricardoV94]

Are you planning to refactor the PyMC distribution to work with the explicit RV from this PR?

[ricardoV94]

small tweak to comments

[ricardoV94]

There's some indentation error going on

[canyon289]

Looks like the init dist is causing some test failure. I can look into this weekend, next 72 hours or so

https://github.com/pymc-devs/pymc/runs/5788430816?check_suite_focus=true#step:7:411

[canyon289]

Alright well this passes except black which I think is broken, and mypy which Im pretty sure is broken, and flaky codecov. Should I just merge and those get fixed up in other prs?