Plugins Reference


Guild supports a plugin API that is used to modify and extend Guild behavior.

Guild plugins can be used for the following functions:

  • Extend or modify Guild file configuration
  • Perform tasks during a run
  • Provide built-in operations

Enable Plugins for an Operation

By default, plugins are not enabled for operations. You can enable one or more plugins for an operation using the plugins operation attribute.

For example, to enable all summary-related plugins for operation train, use:

  plugins: summary

Important Summary plugins are not supported with Guild output scalars. You must explicitly log scalars as TensorBoard summaries from the operation script itself using one of the supported methods.

Built-In Plugins

Guild provides the following built-in plugins:

cpu Logs CPU usage per step
disk Logs disk usage per step
exec_script Support for running executable scripts
gpu Logs GPU usage per step
keras Support for running Keras scripts
memory Logs memory usage per step
perf Logs last step performance per step
python_script Support for running Python scripts
queue Support for queue operations
skopt Support for scikit-optimize based operations (e.g. Bayesian optimizers)

Custom Plugins

To provide a custom plugin, create a class that extends guild.plugins.Plugin and register it as a Python entry point.

Refer to Guild AI plugins source code for examples.

Below are specific examples for various uses.

Example Use Log additional scalar values for a given step. Support new operation types — e.g. handling operation specs such as a path to an R script, a Java JAR file, a URL, etc. Support new Python based frameworks. Add built-in operations including optimizers.