GPflow

2.10.0 verified Sat May 09 auth: no python

GPflow is a Gaussian process library in TensorFlow, designed for GP regression and advanced GP models. It provides a flexible, composable framework for building and training GP models using TensorFlow's automatic differentiation. Current version 2.10.0 targets TensorFlow 2.x and has a moderate release cadence (major/minor roughly every few months).

pip install gpflow

Common errors

error AttributeError: module 'gpflow' has no attribute 'kernels' ↓

cause GPflow 2.x lazy-loads submodules; you must import them explicitly.

fix

Use from gpflow import kernels instead of import gpflow.kernels.

error TypeError: Expected float64, got float32 ↓

cause GPflow defaults to float64; TensorFlow may produce float32 from default settings.

fix

Cast input data: tf.cast(X, tf.float64) or set global default to float64.

Warnings

breaking GPflow 2.x is a major rewrite with breaking changes from 1.x. The API, module structure, and TensorFlow version requirements (TF 2.x) are different. Models built for GPflow 1.x will not work without modification. ↓

fix Migrate code: use `from gpflow import ...` (not `import gpflow as gp` submodules), replace `gpflow.kernels.RBF` with `gpflow.kernels.SquaredExponential`, and update to TF 2.x.

deprecated GPflow 2.10.0 still supports TensorFlow 2.x, but future versions may require TensorFlow 2.10+ or higher. Check compatibility matrix. ↓

fix Keep TensorFlow updated and pin versions if needed.

gotcha When using `gpflow.models.GPR` (or other models), the training loss is a TensorFlow tensor, not a scalar. Use `.numpy()` to get a Python float if needed. ↓

fix Call `m.training_loss.numpy()` to get the value as a float.

Imports

gpflow.kernels
wrong
```
import gpflow.kernels
```
correct
```
from gpflow import kernels
```
In GPflow 2.x, submodules are not automatically imported; use `from gpflow import kernels` or import explicitly.
GPR
```
from gpflow.models import GPR
```
SVGP
```
from gpflow.models import SVGP
```

Quickstart

Basic GP regression with GPflow: create data, build GPR model with SquaredExponential kernel, optimise hyperparameters, and predict.

import gpflow
import numpy as np
import tensorflow as tf

# Create synthetic data
X = np.random.randn(20, 1)
Y = np.sin(X) + 0.1 * np.random.randn(20, 1)

# Build model
k = gpflow.kernels.SquaredExponential()
m = gpflow.models.GPR(data=(X, Y), kernel=k, noise_variance=1.0)

# Optimise
opt = gpflow.optimizers.Scipy()
opt.minimize(m.training_loss, m.trainable_variables)

# Predict
Xtest = np.linspace(-3, 3, 100)[:, None]
mean, var = m.predict_f(Xtest)
print("Predictions:", mean.numpy().flatten()[:5])