Pymc Distributions | Skills Pool

Pymc Distributions | Skills Pool

# Truncated distribution
pm.Truncated("x", pm.Normal.dist(mu=0, sigma=1), lower=0, upper=10)

# Censored distribution
pm.Censored("y", pm.Normal.dist(mu=0, sigma=1), lower=None, upper=5)

# Define custom distribution
def custom_logp(value, mu):
    return -0.5 * ((value - mu) ** 2)

pm.CustomDist("custom", mu, logp=custom_logp, observed=data)

pm.Simulator("sim", fn=simulator_function, params=[param1, param2], observed=data)

# Log-probability
dist.logp(value)

# Log cumulative distribution
dist.logcdf(value)

# Inverse CDF (quantile function)
dist.icdf(q)

# Random sampling (outside of pm.sample)
dist.random(size=100, rng=rng)

# Old (PyMC3)
import pymc3 as pm

# New (latest PyMC version)
import pymc as pm

# Some distributions have parameter renames
# Check documentation for specific changes

# Old: sd parameter
pm.Normal('x', mu=0, sd=1)  # PyMC3

# New: sigma parameter
pm.Normal('x', mu=0, sigma=1)  # latest PyMC version

# Shape handling is more explicit in latest PyMC version
# Use dims and coords for cleaner shape management

with pm.Model(coords={"subject": subjects, "time": times}) as model:
    x = pm.Normal("x", mu=0, sigma=1, dims=("subject", "time"))

# Creating distributions without adding to model
# Old
dist = pm.Normal.dist(mu=0, sd=1)

# New
dist = pm.Normal.dist(mu=0, sigma=1)

import pymc as pm
import numpy as np

# Generate data
true_mu = 5
true_sigma = 2
data = np.random.normal(true_mu, true_sigma, size=100)

# Build model
with pm.Model() as model:
    # Priors
    mu = pm.Normal("mu", mu=0, sigma=10)
    sigma = pm.HalfNormal("sigma", sigma=5)

    # Likelihood
    y = pm.Normal("y", mu=mu, sigma=sigma, observed=data)

    # Sample
    trace = pm.sample(1000, tune=1000)