Researcher Library: Hyperparameter Optimization Support¶
The Run:ai Researcher Library is a python library you can add to your deep learning python code. The hyperparameter optimization(HPO) support module of the library is a helper library for hyperparameter optimization (HPO) experiments
Hyperparameter optimization (HPO) is the process of choosing a set of optimal hyperparameters for a learning algorithm. A hyperparameter is a parameter whose value is used to control the learning process. Example hyperparameters: Learning rate, Batch size, Different optimizers, number of layers.
To search for good hyperparameters, Researchers typically start a series of small runs with different hyperparameter values, let them run for a while, and then examine results to decide what works best.
With the reporter module, you can externalize information such as progress, accuracy, and loss over time/epoch, and more. In addition, you can externalize custom metrics of your choosing.
Run:ai HPO library is dependent on PyYAML. Install it using the command:
runai Python library using
pip using the following command:
Make sure to use the correct
pipinstaller (you might need to use
- Import the
- Initialize the Run:ai HPO library with a path to a directory shared between all cluster nodes (typically using an NFS server). We recommend specifying a unique name for the experiment, the name will be used to create a sub-directory on the shared folder. To do so, we recommend using the environment variables
JOB_UUIDwhich are injected to the container by Run:ai.
- Decide on an HPO strategy:
- Random search - randomly pick a set of hyperparameter values
- Grid search - pick the next set of hyperparameter values, iterating through all sets across multiple experiments
- Call the Run:ai HPO library to specify a set of hyperparameters and pick a specific configuration for this experiment.
- Use the returned configuration in your code. For example:
Metrics could be reported and saved in the experiment directory under the fule
runai.hpo.report. You should pass the epoch number and a dictionary with metrics to be reported. For example: