Sampling from Distributions

Use the rvs method of the scipy distribution stored in dist attribute

distribution.dist.rvs(...)

Scalar parameters

If the parameters are scalars, then just pass the number of samples to the rvs method!

from conjugate.distributions import Exponential

lam = 3.5
true_distribution = Exponential(lam=lam)

n_samples = 10
samples = true_distribution.dist.rvs(n_samples)

Vector parameter

If the parameter is a vector, then there will be a broadcast issue from the scipy distribution.

import numpy as np

lam = np.array([
    [1, 2], 
    [0.5, 5], 
])

true_distribution = Exponential(lam=lam)

n_samples = 100
try: 
    true_distribution.dist.rvs(n_samples)
except ValueError: 
    print("The number of samples doesn't broadcast with the shape of parameters!")

However, this is easy to fix by prepending the number of samples to the shape of the model parameter shape

size = (n_samples, *lam.shape)
samples = true_distribution.dist.rvs(size=size, random_state=rng)

Vector parameters

If there are many parameters in your model, then use the np.broadcast_shapes function in order to get the correct shape before sampling

from conjugate.distributions import Normal

mu = np.array([1, 2, 3])
sigma = np.array([2.5, 5])[:, None]

true_distribution = Normal(mu=mu, sigma=sigma)

shape = np.broadcast_shapes(mu.shape, sigma.shape)
size = (n_samples, *shape)
samples = true_distribution.dist.rvs(size=size, random_state=rng)

Comments