pymc（pymc 命令行）

PyMC Bayesian Modeling

Probabilistic programming with MCMC - build and fit Bayesian models.

When to Use

• Bayesian regression and classification
• Hierarchical/multilevel models
• Uncertainty quantification
• Prior/posterior predictive checks
• Model comparison (LOO, WAIC)

Quick Start

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

# Build model
with pm.Model() as model:
    # Priors
    alpha = pm.Normal('alpha', mu=0, sigma=1)
    beta = pm.Normal('beta', mu=0, sigma=1)
    sigma = pm.HalfNormal('sigma', sigma=1)

    # Likelihood
    mu = alpha + beta * X
    y_obs = pm.Normal('y_obs', mu=mu, sigma=sigma, observed=y)

    # Sample
    idata = pm.sample(2000, tune=1000, chains=4)

# Analyze
az.summary(idata)
az.plot_posterior(idata)

Model Building

coords = {'predictors': ['var1', 'var2', 'var3']}

with pm.Model(coords=coords) as model:
    # Priors with dimensions
    alpha = pm.Normal('alpha', mu=0, sigma=1)
    beta = pm.Normal('beta', mu=0, sigma=1, dims='predictors')
    sigma = pm.HalfNormal('sigma', sigma=1)

    # Linear model
    mu = alpha + pm.math.dot(X, beta)

    # Likelihood
    y_obs = pm.Normal('y_obs', mu=mu, sigma=sigma, observed=y)

Common Models

Linear Regression

with pm.Model() as linear_model:
    alpha = pm.Normal('alpha', mu=0, sigma=10)
    beta = pm.Normal('beta', mu=0, sigma=10, shape=n_features)
    sigma = pm.HalfNormal('sigma', sigma=1)

    mu = alpha + pm.math.dot(X, beta)
    y = pm.Normal('y', mu=mu, sigma=sigma, observed=y_obs)

Logistic Regression

with pm.Model() as logistic_model:
    alpha = pm.Normal('alpha', mu=0, sigma=10)
    beta = pm.Normal('beta', mu=0, sigma=10, shape=n_features)

    logit_p = alpha + pm.math.dot(X, beta)
    y = pm.Bernoulli('y', logit_p=logit_p, observed=y_obs)

Hierarchical Model

with pm.Model(coords={'groups': group_names}) as hierarchical:
    # Hyperpriors
    mu_alpha = pm.Normal('mu_alpha', mu=0, sigma=10)
    sigma_alpha = pm.HalfNormal('sigma_alpha', sigma=1)

    # Group-level (non-centered parameterization)
    alpha_offset = pm.Normal('alpha_offset', mu=0, sigma=1, dims='groups')
    alpha = pm.Deterministic('alpha', mu_alpha + sigma_alpha * alpha_offset)

    # Likelihood
    sigma = pm.HalfNormal('sigma', sigma=1)
    y = pm.Normal('y', mu=alpha[group_idx], sigma=sigma, observed=y_obs)

Sampling

with model:
    # MCMC (default NUTS)
    idata = pm.sample(
        draws=2000,
        tune=1000,
        chains=4,
        target_accept=0.9,
        random_seed=42
    )

    # Variational inference (faster, approximate)
    approx = pm.fit(n=20000, method='advi')

Diagnostics

import arviz as az

# Check convergence
print(az.summary(idata, var_names=['alpha', 'beta', 'sigma']))

# R-hat should be < 1.01
# ESS should be > 400

# Trace plots
az.plot_trace(idata)

# Check divergences
print(f"Divergences: {idata.sample_stats.diverging.sum().values}")

Predictive Checks

with model:
    # Prior predictive (before fitting)
    prior_pred = pm.sample_prior_predictive(1000)
    az.plot_ppc(prior_pred, group='prior')

    # Posterior predictive (after fitting)
    pm.sample_posterior_predictive(idata, extend_inferencedata=True)
    az.plot_ppc(idata)

Model Comparison

# Fit models with log_likelihood
with model1:
    idata1 = pm.sample(idata_kwargs={'log_likelihood': True})

with model2:
    idata2 = pm.sample(idata_kwargs={'log_likelihood': True})

# Compare using LOO
comparison = az.compare({'model1': idata1, 'model2': idata2})
print(comparison)

Predictions

with model:
    pm.set_data({'X': X_new})
    post_pred = pm.sample_posterior_predictive(idata)

# Extract intervals
y_pred_mean = post_pred.posterior_predictive['y'].mean(dim=['chain', 'draw'])
y_pred_hdi = az.hdi(post_pred.posterior_predictive)

Best Practices

1. Standardize predictors for better sampling
2. Use weakly informative priors (not flat)
3. Non-centered parameterization for hierarchical models
4. Check diagnostics before interpretation (R-hat, ESS)
5. Prior predictive checks before fitting
6. Posterior predictive checks after fitting

Troubleshooting

• Divergences → Increase target_accept=0.95, use non-centered
• Low ESS → More draws, reparameterize
• High R-hat → Run longer chains

Resources

• Docs: https://www.pymc.io/projects/docs/
• Examples: https://www.pymc.io/projects/examples/