PyTensor Distributions

0.1.3 verified Fri May 01 auth: no python

PyTensor-powered probability distributions for probabilistic programming and Bayesian inference. Version 0.1.3, released 2025, maintenance level with monthly patches.

pip install pytensor-distributions

Common errors

error ImportError: cannot import name 'Normal' from 'pytensor_distributions' ↓

cause Trying to import distribution class directly from top-level package.

fix

Use 'from pytensor_distributions.distributions import Normal' instead.

error TypeError: 'Normal' object is not callable ↓

cause Attempting to instantiate distribution with constructor call like Normal(mu, sigma) instead of Normal.dist(mu, sigma).

fix

Use Normal.dist(mu, sigma) to create a distribution instance.

error NotImplementedError: Sampling not implemented ↓

cause Calling .sample() on a distribution instance.

fix

Do not use .sample(); use external sampling libraries like PyMC for random draws.

Warnings

breaking The distribution constructors use `.dist()` class method instead of direct instantiation (e.g., `Normal.dist(mu, sigma)` not `Normal(mu, sigma)`). Direct instantiation raises a TypeError. ↓

fix Always call `.dist()` on distribution classes to create distribution instances.

gotcha Random sampling is not supported yet. Calling `dist.sample()` raises NotImplementedError. Samples must be drawn using external sampling methods (e.g., MCMC via PyMC). ↓

fix Use PyMC or other sampling libraries; do not rely on pytensor-distributions for random draws.

deprecated Support for Python 3.10 and below is dropped. Requires Python >=3.11. ↓

fix Upgrade to Python 3.11 or later.

Imports

Normal
wrong
```
from pytensor_distributions import Normal
```
correct
```
from pytensor_distributions.distributions import Normal
```
Direct import of Normal from top-level package fails because distributions are in a submodule.
Distributions
wrong
```
import pytensor_distributions as dist
```
correct
```
from pytensor_distributions import distributions
```
The top-level module does not expose distributions directly; must import the submodule.

Quickstart

Define a Normal distribution and compute log-probability.

import pytensor
import pytensor.tensor as pt
from pytensor_distributions.distributions import Normal

x = pt.dmatrix('x')
mu = pt.dvector('mu')
sigma = pt.dscalar('sigma')
dist = Normal.dist(mu, sigma)
logp = dist.log_prob(x)

# Evaluate
f = pytensor.function([x, mu, sigma], logp)
print(f([[1.0, 2.0], [3.0, 4.0]], [0.0, 1.0], 0.5))