AR1 time series distribution [WIP]

[mjhajharia]

• Add RV
• Add Distribution
• Add Logp
• aesara scan stuff
• Tests

closes #5291

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[codecov]

Codecov Report

Merging #5388 (75f865a) into main (eb5177a) will decrease coverage by 0.01%.
The diff coverage is 51.51%.

@@            Coverage Diff             @@
##             main    #5388      +/-   ##
==========================================
- Coverage   80.44%   80.43%   -0.02%     
==========================================
  Files          82       82              
  Lines       14132    14150      +18     
==========================================
+ Hits        11369    11381      +12     
- Misses       2763     2769       +6     

Impacted Files	Coverage Δ
pymc/distributions/timeseries.py	28.19% <51.51%> (+5.83%)	⬆️
pymc/parallel_sampling.py	86.71% <0.00%> (-1.00%)	⬇️

[mjhajharia]

@ricardoV94 any chance this error happens because I use a for loop in the logp instead of scan

Reply via ReviewNB

[ricardoV94]

Suggested change

	ndim_supp = 0
	ndim_supp = 1

[mjhajharia]

i saw this comment on grw pr as well and for some reason ndim_supp = 1 throws an error when i add observed, something like "dimension mismatch of 2 and 1" as in data is 1d or something, i dont know why does it happen though

[ricardoV94]

Yes, it's related to the comments below, you haven't yet told Aesara how to infer the shape of the distribution given the parameters (and size). It tries to do something by default, but usually fails with multivariate distributions where there is not a simple 1-1 map between parameter shapes and distribution shape

[ricardoV94]

@mjhajharia I would suggest you try to run the default "check_rv_size" and "check_pymc_params_match_rv_op" tests as in here:

pymc/pymc/tests/test_distributions_random.py

Lines 1179 to 1196 in 71b18e6

	class TestStickBreakingWeights(BaseTestDistribution):
	pymc_dist = pm.StickBreakingWeights
	pymc_dist_params = {"alpha": 2.0, "K": 19}
	expected_rv_op_params = {"alpha": 2.0, "K": 19}
	sizes_to_check = [None, 17, (5,), (11, 5), (3, 13, 5)]
	sizes_expected = [
	(20,),
	(17, 20),
	(
	5,
	20,
	),
	(11, 5, 20),
	(3, 13, 5, 20),
	]
	tests_to_run = [
	"check_pymc_params_match_rv_op",
	"check_rv_size",

You need to tell Aesara how to infer the shape of the distribution given the parameters, and those checks cover that

[ricardoV94]

Actually the recent StickBreakingWeights distribution is a very good template in general for this type of multivariate distributions:

pymc/pymc/distributions/multivariate.py

Line 2194 in 71b18e6

class StickBreakingWeightsRV(RandomVariable):

Edit: For the shape inference look here:

pymc/pymc/distributions/multivariate.py

Lines 2214 to 2219 in 71b18e6

	def _infer_shape(self, size, dist_params, param_shapes=None):
	alpha, K = dist_params
	size = tuple(size)
	return size + (K + 1,)

[canyon289]

Just a heads up. I havent forgotten about this. I know youre busy with Data Umbrella and also I still dont exactly know what im doing with new distributions. Once both of those challenges are over in the near future I'll revisit this