Overview
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:
train:
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 |
|---|---|
| cpu.py | Log additional scalar values for a given step. |
| python_script.py | Support new operation types — e.g. handling operation specs such as a path to an R script, a Java JAR file, a URL, etc. |
| keras.py | Support new Python based frameworks. |
| skopt.py | Add built-in operations including optimizers. |