Skip to content

TaskHelperTrain

TaskHelperTrain class

A TaskHelperTrain object contains useful things that are needed to run training as Supervisely task.

class TaskHelperTrain:
    def __init__(self):

Fields

paths

TaskPaths object with paths to input and output directories.

task_settings

Task input settings (Python dictionary).

For training task it is exactly config typed by user in web panel. The config should be a JSON object, but content of the config is entirely dependent on model implementation.

in_project_meta

Meta (metadata) of input project. The meta contains, among others, list of project classes and is required to read annotations in Supervisely format now.

checkpoints_saver

TrainCheckpoints object which is used to save trained model weights.

Methods

model_dir_is_empty(self)

Checks if model dir is empty or missing. Empty model dir means that there are no initial weights for training, and model should be trained "from scratch".