Overview

Task is a self-contained job unit which runs as a separate executable and, in fact, in a separate Docker container. It reads inputs from and writes results to corresponding directories (with fixed structure) and prints messages to stderr/stdout.

For compatibility with Supervisely a task implementation should meet some requirements. So:

• For NN training, just use TaskHelperTrain class.

• For NN inference, just use TaskHelperInference class... and report when inference is finished with report_inference_finished

• For other tasks, use TaskPaths class to determine in/out directories.

Furthermore, it is useful

• to count and report some progress;

• to report metrics from training;