pymc-modeling

Name: pymc-modeling
Author: pymc-labs

135installs10forks

How to install pymc-modeling?

pymc-labs/python-analytics-skills/pymc-modeling

$npx -y skills add pymc-labs/python-analytics-skills --skill pymc-modeling

Installs into the current project.

Summary

Load whenever the user is working on code that imports pymc, pytensor, or arviz, or asks about Bayesian modeling, MCMC, priors, posteriors, sampling, or model diagnostics. Covers PyMC 6+, PyTensor 3+, ArviZ 1.1+ (DataTree API), pymc-bart, pymc-extras, nutpie, and JAX/NumPyro back

Files · 1

View on GitHub

SKILL.md

1	# PyMC Modeling
2
3	Modern Bayesian modeling with PyMC 6+ on the ArviZ 1.1 / PyTensor 3 stack. Key defaults: nutpie sampler (2-5x faster; PyMC 6 selects it automatically when installed — no `nuts_sampler` argument needed), non-centered parameterization for hierarchical models, HSGP over exact GPs, coords/dims for readable DataTree output, and save-early workflow to prevent data loss from late crashes.
4
5	`pm.sample(...)` returns an `xarray.DataTree` — the `idata` name is kept by convention, but it is a DataTree, not the old `InferenceData`. Access groups by bracket: `idata["posterior"]`, `idata["sample_stats"]`, etc.
6
7	Modeling strategy: Build models iteratively — start simple, check prior
8	predictions, fit and diagnose, check posterior predictions, expand one piece at
9	a time. See [references/workflow.md](references/workflow.md) for the full workflow.
10
11	## Model Specification
12
13	### Basic Structure
14
15	```python
16	import pymc as pm
17	import arviz as az
18
19	with pm.Model(coords=coords) as model:
20	# Data containers (for out-of-sample prediction)
21	x = pm.Data("x", x_obs, dims="obs")
22
23	# Priors
24	beta = pm.Normal("beta", mu=0, sigma=1, dims="features")
25	sigma = pm.HalfNormal("sigma", sigma=1)
26
27	# Likelihood
28	mu = pm.math.dot(x, beta)
29	y = pm.Normal("y", mu=mu, sigma=sigma, observed=y_obs, dims="obs")
30
31	# Inference
32	idata = pm.sample(random_seed=42) # PyMC 6 uses nutpie automatically when installed
33	```
34
35	### Coords and Dims
36
37	Use coords/dims for an interpretable DataTree when the model has meaningful structure:
38
39	```python
40	coords = {
41	"obs": np.arange(n_obs),
42	"features": ["intercept", "age", "income"],
43	"group": group_labels,
44	}
45	```
46
47	Skip for simple models where overhead exceeds benefit.
48
49	### Parameterization
50
51	Prefer non-centered parameterization for hierarchical models with weak data:
52
53	```python
54	# Non-centered (better for divergences)
55	offset = pm.Normal("offset", 0, 1, dims="group")
56	alpha = mu_alpha + sigma_alpha * offset
57
58	# Centered (better with strong data)
59	alpha = pm.Normal("alpha", mu_alpha, sigma_alpha, dims="group")
60	```
61
62	## Inference
63
64	### Default Sampling (nutpie preferred)
65
66	In PyMC 6, `pm.sample` uses nutpie automatically whenever it is installed and the
67	model can be compiled — do not pass `nuts_sampler="nutpie"` explicitly:
68
69	```python
70	with model:
71	idata = pm.sample(
72	draws=1000, tune=1000, chains=4,
73	random_seed=42,
74	)
75	idata.to_netcdf("results.nc") # Save immediately after sampling
76	```
77
78	Important: For LOO-CV, model comparison, or LOO-PIT checks, ensure the
79	`log_likelihood` group exists. In PyMC 6, do not pass a top-level
80	`compute_log_likelihood=` argument to `pm.sample`. Either request it during
81	conversion with `idata_kwargs={"log_likelihood": True}` or compute it explicitly
82	after sampling:
83
84	```python
85	idata = pm.sample(idata_kwargs={"log_likelihood": True}, random_seed=42)
86	# or, after an existing sample:
87	pm.compute_log_likelihood(idata, model=model)
88	```
89
90	This applies to every sampler (nutpie, PyMC NUTS, NumPyro) — not just nutpie.
91
92	### When to Use PyMC's Default NUTS Instead
93
94	nutpie cannot handle discrete parameters or certain transforms (e.g., `ordered` transform with `OrderedLogistic`/`OrderedProbit`). PyMC 6 falls back automatically; to force the PyMC sampler explicitly, pass `nuts_sampler="pymc"`:
95
96	```python
97	idata = pm.sample(draws=1000, tune=1000, chains=4, nuts_sampler="pymc", random_seed=42)
98	```
99
100	Never change the model specification to work around sampler limitations.
101
102	If nutpie is not installed, install it (`pip install nutpie`) or fall back to `nuts_sampler="numpyro"`.
103
104	### Alternative MCMC Backends
105
106	See [references/inference.md](references/inference.md) for:
107	- NumPyro/JAX: GPU acceleration, vectorized chains
108
109	### Approximate Inference
110
111	For fast (but inexact) posterior approximations:
112	- ADVI/DADVI: Variational inference with Gaussian approximation
113	- Pathfinder: Quasi-Newton optimization for initialization or screening
114
115	## Diagnostics and ArviZ Workflow
116
117	Minimum workflow checklist — every model script should include:
118	1. Prior predictive check (`pm.sample_prior_predictive`)
119	2. Save results immediately after sampling (`idata.to_netcdf(...)`)
120	3. Divergence count + r_hat + ESS check
121	4.

People who liked this also viewed

6 picks

Type

pymc-modeling

135installs10forks

View on GitHub

How to install pymc-modeling?

pymc-labs/python-analytics-skills/pymc-modeling

$npx -y skills add pymc-labs/python-analytics-skills --skill pymc-modeling

Installs into the current project.

Summary

Files · 1

View on GitHub

SKILL.md

1	# PyMC Modeling
2
3	Modern Bayesian modeling with PyMC 6+ on the ArviZ 1.1 / PyTensor 3 stack. Key defaults: nutpie sampler (2-5x faster; PyMC 6 selects it automatically when installed — no `nuts_sampler` argument needed), non-centered parameterization for hierarchical models, HSGP over exact GPs, coords/dims for readable DataTree output, and save-early workflow to prevent data loss from late crashes.
4
5	`pm.sample(...)` returns an `xarray.DataTree` — the `idata` name is kept by convention, but it is a DataTree, not the old `InferenceData`. Access groups by bracket: `idata["posterior"]`, `idata["sample_stats"]`, etc.
6
7	Modeling strategy: Build models iteratively — start simple, check prior
8	predictions, fit and diagnose, check posterior predictions, expand one piece at
9	a time. See [references/workflow.md](references/workflow.md) for the full workflow.
10
11	## Model Specification
12
13	### Basic Structure
14
15	```python
16	import pymc as pm
17	import arviz as az
18
19	with pm.Model(coords=coords) as model:
20	# Data containers (for out-of-sample prediction)
21	x = pm.Data("x", x_obs, dims="obs")
22
23	# Priors
24	beta = pm.Normal("beta", mu=0, sigma=1, dims="features")
25	sigma = pm.HalfNormal("sigma", sigma=1)
26
27	# Likelihood
28	mu = pm.math.dot(x, beta)
29	y = pm.Normal("y", mu=mu, sigma=sigma, observed=y_obs, dims="obs")
30
31	# Inference
32	idata = pm.sample(random_seed=42) # PyMC 6 uses nutpie automatically when installed
33	```
34
35	### Coords and Dims
36
37	Use coords/dims for an interpretable DataTree when the model has meaningful structure:
38
39	```python
40	coords = {
41	"obs": np.arange(n_obs),
42	"features": ["intercept", "age", "income"],
43	"group": group_labels,
44	}
45	```
46
47	Skip for simple models where overhead exceeds benefit.
48
49	### Parameterization
50
51	Prefer non-centered parameterization for hierarchical models with weak data:
52
53	```python
54	# Non-centered (better for divergences)
55	offset = pm.Normal("offset", 0, 1, dims="group")
56	alpha = mu_alpha + sigma_alpha * offset
57
58	# Centered (better with strong data)
59	alpha = pm.Normal("alpha", mu_alpha, sigma_alpha, dims="group")
60	```
61
62	## Inference
63
64	### Default Sampling (nutpie preferred)
65
66	In PyMC 6, `pm.sample` uses nutpie automatically whenever it is installed and the
67	model can be compiled — do not pass `nuts_sampler="nutpie"` explicitly:
68
69	```python
70	with model:
71	idata = pm.sample(
72	draws=1000, tune=1000, chains=4,
73	random_seed=42,
74	)
75	idata.to_netcdf("results.nc") # Save immediately after sampling
76	```
77
78	Important: For LOO-CV, model comparison, or LOO-PIT checks, ensure the
79	`log_likelihood` group exists. In PyMC 6, do not pass a top-level
80	`compute_log_likelihood=` argument to `pm.sample`. Either request it during
81	conversion with `idata_kwargs={"log_likelihood": True}` or compute it explicitly
82	after sampling:
83
84	```python
85	idata = pm.sample(idata_kwargs={"log_likelihood": True}, random_seed=42)
86	# or, after an existing sample:
87	pm.compute_log_likelihood(idata, model=model)
88	```
89
90	This applies to every sampler (nutpie, PyMC NUTS, NumPyro) — not just nutpie.
91
92	### When to Use PyMC's Default NUTS Instead
93
94	nutpie cannot handle discrete parameters or certain transforms (e.g., `ordered` transform with `OrderedLogistic`/`OrderedProbit`). PyMC 6 falls back automatically; to force the PyMC sampler explicitly, pass `nuts_sampler="pymc"`:
95
96	```python
97	idata = pm.sample(draws=1000, tune=1000, chains=4, nuts_sampler="pymc", random_seed=42)
98	```
99
100	Never change the model specification to work around sampler limitations.
101
102	If nutpie is not installed, install it (`pip install nutpie`) or fall back to `nuts_sampler="numpyro"`.
103
104	### Alternative MCMC Backends
105
106	See [references/inference.md](references/inference.md) for:
107	- NumPyro/JAX: GPU acceleration, vectorized chains
108
109	### Approximate Inference
110
111	For fast (but inexact) posterior approximations:
112	- ADVI/DADVI: Variational inference with Gaussian approximation
113	- Pathfinder: Quasi-Newton optimization for initialization or screening
114
115	## Diagnostics and ArviZ Workflow
116
117	Minimum workflow checklist — every model script should include:
118	1. Prior predictive check (`pm.sample_prior_predictive`)
119	2. Save results immediately after sampling (`idata.to_netcdf(...)`)
120	3. Divergence count + r_hat + ESS check
121	4.

People who liked this also viewed

6 picks

Type