CUQIpy plugin for PyTorch
Project description
CUQIpy-PyTorch
CUQIpy-PyTorch is a plugin for the CUQIpy software package.
It adds a PyTorch backend to CUQIpy, allowing the user to use the PyTorch API to define models, distributions etc.
It also links to the Pyro No U-Turn Hamiltonian Monte Carlo sampler (NUTS) for efficient sampling from the joint posterior.
Installation
For optimal performance consider installing PyTorch using conda, then install CUQIpy-PyTorch using pip:
pip install cuqipy-pytorch
If PyTorch, Pyro or CUQIpy are not installed, they will be installed automatically from the above command.
Quickstart
Example for sampling from the eight schools model:
$$ \begin{align*} \mu &\sim \mathcal{N}(0, 10^2)\ \tau &\sim \log\mathcal{N}(5, 1)\ \boldsymbol \theta' &\sim \mathcal{N}(\mathbf{0}, \mathbf{I}_m)\ \boldsymbol \theta &= \mu + \tau \boldsymbol \theta'\ \mathbf{y} &\sim \mathcal{N}(\boldsymbol \theta, \boldsymbol \sigma^2 \mathbf{I}_m) \end{align*} $$
where $\mathbf{y}\in\mathbb{R}^m$ and $\boldsymbol \sigma\in\mathbb{R}^m$ are observed data.
import torch as xp
from cuqi.distribution import JointDistribution
from cuqipy_pytorch.distribution import Gaussian, LogGaussian
from cuqipy_pytorch.sampler import NUTS
# Observations
y_obs = xp.tensor([28, 8, -3, 7, -1, 1, 18, 12], dtype=xp.float32)
σ_obs = xp.tensor([15, 10, 16, 11, 9, 11, 10, 18], dtype=xp.float32)
# Bayesian model
μ = Gaussian(0, 10**2)
τ = LogGaussian(5, 1)
θp = Gaussian(xp.zeros(8), 1)
θ = lambda μ, τ, θp: μ+τ*θp
y = Gaussian(θ, cov=σ_obs**2)
# Posterior sampling
joint = JointDistribution(μ, τ, θp, y) # Define joint distribution
posterior = joint(y=y_obs) # Define posterior distribution
sampler = NUTS(posterior) # Define sampling strategy
samples = sampler.sample(N=500, Nb=500) # Sample from posterior
# Plot posterior samples
samples["θp"].plot_violin();
print(samples["μ"].mean()) # Average effect
print(samples["τ"].mean()) # Average variance
For more examples, see the demos folder.
Project details
Release history Release notifications | RSS feed
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
Source Distribution
Built Distribution
Hashes for CUQIpy-PyTorch-0.1.1.post0.dev17.tar.gz
Algorithm | Hash digest | |
---|---|---|
SHA256 | 9ca87ced3381bc1755a56f8d2fe253a5e24aaeff822e1349803db8e52c4e1f95 |
|
MD5 | 50954d53223a18f77312df940c2eabfb |
|
BLAKE2b-256 | 314f4bb1ba54a5e4b14afc355e65458aa479f53b9d905f5d6c5cd5aeec752640 |
Hashes for CUQIpy_PyTorch-0.1.1.post0.dev17-py3-none-any.whl
Algorithm | Hash digest | |
---|---|---|
SHA256 | 7c34c069b23d915a9cf33cf0466c2ac0e8bab4acc948136b82d685e4bd2ffc1d |
|
MD5 | 0a422ae11d3b14121233904823a58e3b |
|
BLAKE2b-256 | 5c7710e8865cb8912c0e85fa1645c982d29061a428c1dae45ef7273620953ba5 |