Cannot pass dims to idata_kwargs parameter in sample_numpyro_nuts

[jhrcook]

A minimal, self-contained, and reproducible example.

import pymc as pm
import pymc.sampling_jax
import arviz as az

coords = {"param": ["a", "b"]}
with pm.Model(coords=coords) as model:
    chol, corr, stds = pm.LKJCholeskyCov(
        "chol", n=2, eta=2.0, sd_dist=pm.Gamma.dist(2, 1)
    )
    trace = pm.sample(
        10, tune=10, cores=2, chains=2, idata_kwargs={"dims": {"chol_stds": ["param"]}}
    )

with model:
    trace = pymc.sampling_jax.sample_numpyro_nuts(
        10, tune=10, chains=2, idata_kwargs={"dims": {"chol_stds": ["param"]}}
    )

Full traceback.

---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
Input In [5], in <cell line: 1>()
      1 with model:
----> 2     trace = pymc.sampling_jax.sample_numpyro_nuts(10, tune=10, chains=2, idata_kwargs={"dims": {"chol_stds": ["param"]}})

File /usr/local/Caskroom/miniconda/base/envs/bluishred/lib/python3.10/site-packages/pymc/sampling_jax.py:564, in sample_numpyro_nuts(draws, tune, chains, target_accept, random_seed, initvals, model, var_names, progress_bar, keep_untransformed, chain_method, postprocessing_backend, idata_kwargs, nuts_kwargs)
    559 attrs = {
    560     "sampling_time": (tic3 - tic2).total_seconds(),
    561 }
    563 posterior = mcmc_samples
--> 564 az_trace = az.from_dict(
    565     posterior=posterior,
    566     log_likelihood=log_likelihood,
    567     observed_data=find_observations(model),
    568     constant_data=find_constants(model),
    569     sample_stats=_sample_stats_to_xarray(pmap_numpyro),
    570     coords=coords,
    571     dims=dims,
    572     attrs=make_attrs(attrs, library=numpyro),
    573     **idata_kwargs,
    574 )
    576 return az_trace

TypeError: arviz.data.io_dict.from_dict() got multiple values for keyword argument 'dims'

Details.

I am trying to add coordinate labels to the variables produced by LKJCholeskyCov() by passing the dimensions to the idata_kwargs parameter of the PyMC sampling function (as demonstrated in Oriol Abril's blog post (https://oriolabrilpla.cat/python/arviz/pymc/xarray/2022/06/07/pymc-arviz.html#2nd-example:-radon-multilevel-model). The method works when sampling with the default PyMC sampler, but fails with the Numpyro JAX backend. I have provided a full example in the code above, but please let me know if more details are needed

Versions and main components

• PyMC Version: 4.0.1
• Aesara Version: 2.7.3
• Python Version: 3.10.5
• Other relevant libraries: jax=v0.3.14, jaxlib=0.3.10, numpyro=0.9.2
• Operating system: macOS Monterey (v12.4)
• How did you install PyMC/PyMC3: conda

[bherwerth]

Thanks for reporting.

The same problem occurs with coords in idata_kwargs and when using sample_numpyro_nuts. This is shown by running the following code with py.test:

import pymc as pm
from functools import partial

from pymc import sampling_jax
import pytest

@pytest.mark.parametrize(
    "sampler",
    [
        partial(pm.sample, cores=1),
        sampling_jax.sample_blackjax_nuts,
        sampling_jax.sample_numpyro_nuts,
    ]
)
@pytest.mark.parametrize(
    "idata_kwargs",
    [
        {},
        {"dims": {"x": ["x_coord"]}},
        {"coords": {"x": ["a", "b"]}},
        {"coords": {"x": ["a", "b"]}, "dims": {"x": ["x_coord"]}},
    ]
)
def test_idata_kwargs(sampler, idata_kwargs):
    with pm.Model() as model:
        x = pm.Uniform("x", -10, 10, shape=(2,))
        trace = sampler(100, tune=100, idata_kwargs=idata_kwargs)

This produces the following output:

================================================================================================ short test summary info ================================================================================================
FAILED issue_5932.py::test_idata_kwargs[idata_kwargs1-sample_blackjax_nuts] - TypeError: arviz.data.io_dict.from_dict() got multiple values for keyword argument 'dims'
FAILED issue_5932.py::test_idata_kwargs[idata_kwargs1-sample_numpyro_nuts] - TypeError: arviz.data.io_dict.from_dict() got multiple values for keyword argument 'dims'
FAILED issue_5932.py::test_idata_kwargs[idata_kwargs2-sample_blackjax_nuts] - TypeError: arviz.data.io_dict.from_dict() got multiple values for keyword argument 'coords'
FAILED issue_5932.py::test_idata_kwargs[idata_kwargs2-sample_numpyro_nuts] - TypeError: arviz.data.io_dict.from_dict() got multiple values for keyword argument 'coords'
FAILED issue_5932.py::test_idata_kwargs[idata_kwargs3-sample_blackjax_nuts] - TypeError: arviz.data.io_dict.from_dict() got multiple values for keyword argument 'coords'
FAILED issue_5932.py::test_idata_kwargs[idata_kwargs3-sample_numpyro_nuts] - TypeError: arviz.data.io_dict.from_dict() got multiple values for keyword argument 'coords'
================================================================================== 6 failed, 6 passed, 3 warnings in 101.87s (0:01:41) ==================================================================================

I think that the problem that dims and coords is defined in the sampling functions and then passed to az.from_dict together with idata_kwargs:

pymc/pymc/sampling_jax.py

Lines 368 to 377 in 2583b7f

	az_trace = az.from_dict(
	posterior=posterior,
	log_likelihood=log_likelihood,
	observed_data=find_observations(model),
	constant_data=find_constants(model),
	coords=coords,
	dims=dims,
	attrs=make_attrs(attrs, library=blackjax),
	**idata_kwargs,
	)

A solution could be to pass ikwargs = dict(dims=dims, coords=coords, ...) after updating ikwargs with idata_kwargs, similar sampling.sample:

pymc/pymc/sampling.py

Lines 685 to 687 in 2583b7f

	ikwargs = dict(model=model, save_warmup=not discard_tuned_samples)
	if idata_kwargs:
	ikwargs.update(idata_kwargs)

[jhrcook]

The implementation to address this issue overwrites the dimensions extracted from the random variables in the model. Therefore, supplying some dimesions through idata_kwargs={"dims": {...}} causes all other variables to lose their information. Here is an example:

import arviz as az
import pymc as pm
import pymc.sampling_jax

coords = {
    "param": ["a", "b"],
    "animal": ["bear", "penguin", "lizard"],
}

with pm.Model(coords=coords) as model:
    a = pm.Normal("a", 0, 1, dims="animal")
    chol, corr, stds = pm.LKJCholeskyCov(
        "chol", n=2, eta=2.0, sd_dist=pm.Gamma.dist(2, 1)
    )
    trace_pm = pm.sample(
        10, tune=10, cores=2, chains=2, idata_kwargs={"dims": {"chol_stds": ["param"]}}
    )

with model:
    trace_numpyro = pymc.sampling_jax.sample_numpyro_nuts(
        10, tune=10, chains=2, idata_kwargs={"dims": {"chol_stds": ["param"]}}
    )

The information is retained for variable a in the PyMC NUTS posterior:

>>> trace_pm.posterior["a"].dims
('chain', 'draw', 'animal')
>>> trace_pm.posterior["a"].coords
Coordinates:
  * chain    (chain) int64 0 1
  * draw     (draw) int64 0 1 2 3 4 5 6 7 8 9
  * animal   (animal) <U7 'bear' 'penguin' 'lizard'

The information is lost for variable a in the Numpyro NUTS posterior:

>>> trace_numpyro.posterior["a"].dims
('chain', 'draw', 'a_dim_0')
>>> trace_numpyro.posterior["a"].coords
Coordinates:
  * chain    (chain) int64 0 1
  * draw     (draw) int64 0 1 2 3 4 5 6 7 8 9
  * a_dim_0  (a_dim_0) int64 0 1 2

I think it should be a simple enough fix, so I will try working on it and submit a PR. I'll also add cases to the tests added in the original solution.

[bherwerth]

Thank you for noting this and proposing a fix. When I wrote the original solution, I was assuming the dims and coords from idata_kwargs should be passed without change. However, I agree that updating existing dims and coords makes more sense and ensures consistency with pm.sample.

[jhrcook]

I think that's a perfectly fair interpretation. Both use cases could be valuable.