This Repo contains a simple starter project which wraps a model method and parses arguments according to Katib's format.
An indepth exploration of the project is provided in this Blog Post: End-to-End Hyperparameter Tuning with Katib, Tensorflow, Keras and Nvidia GPU on a gaming laptop
The user familiar with ML and Containers base concepts can use the sample method to implement their own models and Experiment on an existing Kubeflow deployment.
Some sample Katib YAML files can be found in the katib folder.
To run the base experiments on this project, one can simple run the following command on an existing kubernetes cluster with kubeflow deployment:
kubectl apply -f https://raw.githubusercontent.com/fabiononato/do-katib/master/katib/random-experiment.yaml
This project is set as a Depend-on-Docker automation that will help you build a container around the Katib wrapped method.
The main entrypoint to the project is the script containing the my_model method. Once you've added a call to your model into this script, you'll need to build the project image. Please refer to the Depend-on-Docker project documentation for more details.
The modifications specific to this projects are:
-
a test scrip that checks for the proper set up of the NVIDIA drivers, as this project rely on a GPU version of TF.
-
an environment file with a registry and project name specific to this project.
-
a tensorflow-gpu base image on our Dockerfile.