gpflux.callbacks#

This module implements a callback that enables GPflow’s gpflow.monitor.ModelToTensorBoard to integrate with Keras’s tf.keras.Model‘s fit method.

Module Contents#

class TensorBoard(log_dir: str = 'logs', *, keywords_to_monitor: List[str] = ['kernel', 'mean_function', 'likelihood'], max_size: int = 3, histogram_freq: int = 0, write_graph: bool = True, write_images: bool = False, update_freq: int | str = 'epoch', profile_batch: int = 2, embeddings_freq: int = 0, embeddings_metadata: Dict | None = None)[source]#

Bases: tf.keras.callbacks.TensorBoard

This class is a thin wrapper around a tf.keras.callbacks.TensorBoard callback that also calls GPflow’s gpflow.monitor.ModelToTensorBoard monitoring task.

Parameters:

  • log_dir – The path of the directory to which to save the log files to be read by TensorBoard.

  • keywords_to_monitor – A list of keywords to monitor. If the parameter’s name includes any of the keywords specified, it will be added to the TensorBoard.

  • max_size – The maximum size of arrays (inclusive) for which each element is written independently as a scalar to the TensorBoard.

For information on all other arguments, see TensorFlow’s TensorBoard callback docs.

log_dir: str[source]#

The path of the directory to which to save the log files to be read by TensorBoard. Files are saved in log_dir / "train", following the Keras convention.

keywords_to_monitor: List[str][source]#

Parameters whose name match a keyword in the keywords_to_monitor list will be added to the TensorBoard.

update_freq: int | str[source]#

Either an integer or "epoch". If using an integer n, write losses/metrics/parameters at every nth batch; when using "epoch", write them at the end of each epoch. Note that writing too frequently to TensorBoard can slow down the training.

set_model(model: tf.keras.Model) → None[source]#

Set the model (extends the Keras set_model method).

This method initialises KerasModelToTensorBoard and mimics Keras’s TensorBoard callback in writing the summary logs to log_dir / “train”.

on_train_batch_end(batch: int, logs: Mapping | None = None) → None[source]#

Write to TensorBoard if update_freq is an integer.

Parameters:

  • batch – The index of the batch within the current epoch.

  • logs – The metric results for this batch.

on_epoch_end(epoch: int, logs: Mapping | None = None) → None[source]#

Write to TensorBoard if update_freq equals "epoch".