Supervisor, Workers and WebUI need not run on the same host. The entire Hyperdock system can be distributed, i.e. They continually send status updates to the database to notify the user of progress. The Hyperdock Workers dequeues jobs from the work queue and then evaluates the target image with these parameters.
It determines what jobs (a specific hyperparameter combination) need to scheduled, which jobs have failed and should be restarted, and notifies users of results. The Hyperdock Supervisor is a background process that monitors all trials. All trials, their status and results are stored in a Mongo database. The Hyperdock Web UI is the main interface for the end-users from where they can specify trials (a target Docker image and the hyperparameter space to search over). The figure below shows the Hyperdock system and its main components. That means that the target can be written in any language, use any framework or run on any operating system as long as it can be made into a Docker image. Hyperdock is a framework for hyperparameter search that can be used to optimized any target as long as that target can be run in a Docker container. A simple framework for distributed hyperparameter optimization in Docker.
0 kommentar(er)
