gpflux.losses

This module provides the LikelihoodLoss adapter to use GPflow’s Likelihood implementations as a tf.keras.losses.Loss.

Module Contents

unwrap_dist(dist: tfp.distributions.Distribution) → tfp.distributions.Distribution

Unwrap the given distribution, if it is wrapped in a _TensorCoercible.

class LikelihoodLoss(likelihood: gpflow.likelihoods.Likelihood)

Bases: gpflow.keras.tf_keras.losses.Loss

This class is a tf.keras.losses.Loss implementation that wraps a GPflow Likelihood instance.

When the prediction (last-layer output) is a Distribution q(f), calling this loss returns the negative variational expectation \(-\mathbb{E}_{q(f)}[\log p(y|f)]\). When the prediction is a tf.Tensor, calling this loss returns the negative log-probability \(-\log p(y|f)\).

When you use this loss function in training a Keras model, the value of this loss is not logged explicitly (in contrast, the layer-specific losses are logged, as is the overall model loss). To output this loss value explicitly, wrap this class in a tf.keras.metrics.Metric and add it to the model metrics.

Note

Use either this LikelihoodLoss (e.g. together with a tf.keras.Sequential model) or LikelihoodLayer (together with gpflux.models.DeepGP). Do not use both at once because this would add the loss twice.

Parameters:

likelihood –

the GPflow likelihood object to use.

Note

If you want to train any parameters of the likelihood (e.g. likelihood variance), you must include the likelihood as an attribute on a TrackableLayer instance that is part of your model. (This is not required when instead you use a gpflux.layers.LikelihoodLayer together with gpflux.models.DeepGP.)

call(y_true: gpflow.base.TensorType, f_prediction: gpflow.base.TensorType | tfp.distributions.MultivariateNormalDiag) → tf.Tensor

Note that we deviate from the Keras Loss interface by calling the second argument f_prediction rather than y_pred.