PTGP

A Gaussian process library for building GP models that solve real-world problems.

Who this is for

PTGP is for practitioners who need flexible, well-supported GP modeling. The goal of PTGP is to be fully batteries-included and ready to work on real-world problems:

• Practical GP algorithms: exact GP, VFE with collapsed bound, SVGP with minibatch training, VFF (Variational Fourier Features)
• Full kernel library: ExpQuad, Matern52/32/12, RandomWalk, Gibbs, WarpedInput, categorical kernels for multi-class or categorical input variables, composition via + and *, active_dims for dimension selection
• Non-Gaussian likelihoods: Bernoulli, Poisson, NegativeBinomial, StudentT
• PyMC priors: set priors on any hyperparameter; use PyMC distributions for mean functions and noise models; MAP training by default
• Training tools: L-BFGS-B and Adam optimizers, per-parameter learning rates, staged optimization, frozen variables, inducing point initialization strategies, diagnostic-guided workflows; more are being added, such as carefully monitored training to help diagnose issues early
• Agent-readable docs: docs/agents/ ships LLM-readable guides for debugging training issues and folk wisdom (VFE training covered). See the Working with AI coding assistants section below.
• More coming: see the issues

Researchers benefit from the underlying design: PTGP is built on PyTensor's symbolic graph and rewrite system, so you write GP math directly (pt.linalg.inv(K), pt.linalg.slogdet(K)) and the compiler chooses efficient algorithms based on declared matrix structure. This makes it straightforward to implement new GP approximations and create custom models, and will eventually allow matrix structure like Kronecker, Toeplitz, and sparse to be taken advantage of automatically.

Models

Model	Scale	Best for
gp.Unapproximated	N < ~2,000	Exact inference, model comparison
gp.VFE	N < ~50,000	Medium-scale data with inducing points
gp.SVGP	N up to ~500,000	Large data, non-Gaussian likelihoods, minibatch training
FourierFeatures1D	1D Matern kernels	Structured Kuu via Fourier basis; no inducing point placement

Quick start

import numpy as np
import pymc as pm
import pytensor.tensor as pt
import ptgp as pg

X = np.random.randn(200, 1)
y = np.sin(X.ravel()) + 0.1 * np.random.randn(200)
Z_init = np.linspace(-2, 2, 20)[:, None]
Z_var = pt.matrix("Z", shape=(20, 1))

with pm.Model() as model:
    ls = pm.InverseGamma("ls", alpha=2.0, beta=1.0)
    eta = pm.Exponential("eta", lam=1.0)
    kernel = eta**2 * pg.kernels.Matern52(input_dim=1, ls=ls)

    svgp = pg.gp.SVGP(
        kernel=kernel,
        likelihood=pg.likelihoods.Gaussian(sigma=0.1),
        inducing_variable=pg.inducing.Points(Z_var, Z_init=Z_init),
        variational_params=pg.gp.init_variational_params(M=20),
    )
    fit = pg.fit(svgp, X, y, method="L-BFGS-B")

mean, var = pg.predict(svgp, np.linspace(-3, 3, 100)[:, None], fit)

pg.fit picks a default objective from the gp type (Unapproximated → marginal_log_likelihood, VFE → collapsed_elbo, SVGP → elbo) and returns a FitResult that pg.predict consumes. For stochastic mini-batch training, staged VFE, or per-group learning rates, drop down to pg.optim.compile_training_step / pg.optim.compile_scipy_objective — see notebooks/demo.ipynb:

X_var = pt.matrix("X")
y_var = pt.vector("y")

step, shared_params, shared_extras = pg.optim.compile_training_step(
    pg.objectives.elbo, svgp, X_var, y_var, model, learning_rate=1e-2
)

for i in range(500):
    loss = step(X, y)

predict_fn = pg.optim.compile_predict(
    svgp, pt.matrix("X_new"), model, shared_params, shared_extras=shared_extras
)
mean, var = predict_fn(np.linspace(-3, 3, 100)[:, None])

Training uses MAP by default: the PyMC log-prior is added to the objective. Pass include_prior=False for pure ELBO. For exact GPs and VFE, use compile_scipy_objective with L-BFGS-B instead. See notebooks/demo.ipynb for end-to-end examples covering all three models.

How it works

PTGP is built on PyTensor's symbolic graph. Kernels, likelihoods, and GP models return symbolic tensors with naive linear algebra like pt.linalg.inv(K) that PyTensor's rewrite system automatically lowers to efficient Cholesky-based code using declared matrix properties. All models compile their full forward+gradient step down to the minimum number of cubic factorizations.

PTGP tries to distill some of the approaches of existing GP libraries and make them more accessible, mainly GPJax, GPflow, and GPyTorch.

Working with AI coding assistants

PTGP is set up to work nicely with AI coding assistants:

• AGENTS.md — project-level instructions for AI coding assistants (architecture, conventions, where things live, how to run tests). Follows the AGENTS.md cross-tool convention used by Codex, Cursor, Aider, and others.
• docs/agents/ — backend-agnostic agent-skill docs covering folk wisdom and training-debug recipes. Currently includes ptgp-vfe (VFE diagnostic skill: pitfalls, escalation workflow, interpretation of VFEDiagnostics and GreedyVarianceDiagnostics).

Claude Code users

Claude Code reads CLAUDE.md, not AGENTS.md. Symlink so they stay in sync:

ln -s AGENTS.md CLAUDE.md

To install the VFE skill into a Claude Code skill directory (so Claude auto-discovers it when you mention VFE), run:

python scripts/install_claude_skills.py --project .   # ./.claude/skills/
python scripts/install_claude_skills.py --user        # ~/.claude/skills/

Install

pip install git+https://github.com/pymc-devs/ptgp.git

To hack on PTGP itself, clone and install in editable mode:

git clone https://github.com/pymc-devs/ptgp.git
cd ptgp
pip install -e .

Contributing

See the issues for what's being worked on. Feel free to propose issues, feature requests, or use cases you've been hoping could be made easier. PRs always welcome.