The Keras library for Python has a number of options for monitoring the progress of model training. Keras comes with a number of built-in callbacks to save checkpoint state, update the learning rate, and monitor progress. The built-in progress monitors include CSV, a remote endpoint, and Tensor Board. While TensorBoard provides some visualizations, the UI is not the easiest to navigate and provides limited options for interaction. To extend this, we developed a quick custom callback that logs Keras progress to Elasticsearch.
With just a few steps explained on the ComputationalHealth GitHub repository (and a working Elasticsearch/Kibana install), you can quickly collect your ML progress data within Elasticsearch and quickly visualize in real-time on Kibana.
Keras ES Callback: Monitoring Machine Learning with Kibana
- May 28, 2017