PyMC

Warnings

• breaking PyMC v5.26.0 dropped support for Python 3.10 and NumPy versions older than 2.0. Ensure your Python environment is 3.11 or newer and NumPy is 2.0 or newer.
  Fix: Upgrade Python to >=3.11 and NumPy to >=2.0. If using conda, create a new environment with `conda create -c conda-forge -n pymc_env "pymc>=5" python=3.11`.
• breaking PyMC was renamed from `PyMC3` to `PyMC` with the release of version 4.0. All `import pymc3 as pm` statements must be updated to `import pymc as pm`.
  Fix: Replace `import pymc3 as pm` with `import pymc as pm`.
• breaking With PyMC v5.27.0, the `VarName` class was replaced by Python's built-in `str`. Code relying on `VarName` objects for variable introspection or manipulation may need adjustment.
  Fix: Review any code directly interacting with `VarName` objects and adapt to use string representations of variable names.
• breaking From PyMC v5.26.0, `Model.compile_logp` now expects *all* model variables as input, not just a subset of the logp terms. This changes the API for functions that infer graph inputs.
  Fix: Update calls to `Model.compile_logp` to provide all necessary model variables as inputs.
• gotcha PyMC relies on PyTensor's ability to compile C implementations for performance. If `g++` (or an equivalent C++ compiler) is not detected in your environment, PyTensor will fall back to slower Python implementations, severely degrading performance.
  Fix: Install a C++ compiler (e.g., `conda install m2w64-toolchain` on Windows with Conda, or `sudo apt-get install build-essential` on Debian/Ubuntu).
• gotcha Since PyMC3 version 3.9 (and all PyMC v4+), `pm.sample()` defaults to returning an `arviz.InferenceData` object instead of a `MultiTrace` object. Ensure your analysis code is adapted for `InferenceData`.
  Fix: Access sampling results using `idata.posterior`, `idata.sample_stats`, etc., and use ArviZ plotting and summary functions which expect `InferenceData`.

Install

• pip install pymc pip
• conda install -c conda-forge pymc Conda

Imports

• pm

  import pymc as pm

  PyMC was renamed from PyMC3 starting with version 4.0. The 'pymc3' import is for older versions and will not work for PyMC v4+ projects.
• Model

  from pymc import Model

  Commonly used within a 'with' statement for model definition.
• Normal

  from pymc import Normal

  Example distribution; many other distributions are available directly from the top-level 'pymc' package.

Quickstart

This quickstart demonstrates how to define a simple Bayesian linear regression model, perform Markov chain Monte Carlo (MCMC) sampling using `pm.sample()`, and analyze the posterior results with ArviZ.

import numpy as np
import pymc as pm
import arviz as az

# 1. Simulate some data
np.random.seed(42)
size = 100
alpha_true, beta_true = 1, [1, 2.5]
sigma_true = 1

X1 = np.random.randn(size)
X2 = np.random.randn(size) * 0.2
Y = alpha_true + beta_true[0] * X1 + beta_true[1] * X2 + np.random.normal(size=size) * sigma_true

# 2. Define the PyMC model
with pm.Model() as linear_model:
    # Priors for unknown model parameters
    alpha = pm.Normal('alpha', mu=0, sigma=10)
    beta = pm.Normal('beta', mu=0, sigma=10, shape=2)
    sigma = pm.HalfNormal('sigma', sigma=1)

    # Expected value of outcome
    mu = alpha + beta[0] * X1 + beta[1] * X2

    # Likelihood (sampling distribution) of observations
    Y_obs = pm.Normal('Y_obs', mu=mu, sigma=sigma, observed=Y)

    # 3. Perform MCMC sampling
    idata = pm.sample(draws=1000, tune=1000, cores=2, random_seed=42)

# 4. Analyze results (e.g., print summary)
print(az.summary(idata, var_names=['alpha', 'beta', 'sigma']))

# To visualize, you would typically use:
# import matplotlib.pyplot as plt
# az.plot_trace(idata)
# plt.show()