ImageClassificationTrainer.EarlyStopping Class

Definition

Namespace:

Microsoft.ML.Vision

Assembly:

Microsoft.ML.Vision.dll

Package:

Microsoft.ML.Vision v3.0.1

Early Stopping feature stops training when monitored quantity stops improving'. Modeled after https://github.com/tensorflow/tensorflow/blob/00fad90125b18b80fe054de1055770cfb8fe4ba3/ tensorflow/python/keras/callbacks.py#L1143

public sealed class ImageClassificationTrainer.EarlyStopping

type ImageClassificationTrainer.EarlyStopping = class

Public NotInheritable Class ImageClassificationTrainer.EarlyStopping

Inheritance

Object

ImageClassificationTrainer.EarlyStopping

Constructors

ImageClassificationTrainer.EarlyStopping(Single, Int32, ImageClassificationTrainer+EarlyStoppingMetric, Boolean)

Properties

CheckIncreasing	Whether the monitored quantity is to be increasing (eg. Accuracy, CheckIncreasing = true) or decreasing (eg. Loss, CheckIncreasing = false).
MinDelta	Minimum change in the monitored quantity to be considered as an improvement.
Patience	Number of epochs to wait after no improvement is seen consecutively before stopping the training.

Methods

ShouldStop(ImageClassificationTrainer+TrainMetrics)

To be called at the end of every epoch to check if training should stop. For increasing metric(eg.: Accuracy), if metric stops increasing, stop training if value of metric doesn't increase within 'patience' number of epochs. For decreasing metric(eg.: Loss), stop training if value of metric doesn't decrease within 'patience' number of epochs. Any change in the value of metric of less than 'minDelta' is not considered a change.