ImageClassificationTrainer.Options Class

Definition

Options class for ImageClassificationTrainer.

public sealed class ImageClassificationTrainer.Options : Microsoft.ML.Trainers.TrainerInputBaseWithLabel
type ImageClassificationTrainer.Options = class
    inherit TrainerInputBaseWithLabel
Public NotInheritable Class ImageClassificationTrainer.Options
Inherits TrainerInputBaseWithLabel
Inheritance
ImageClassificationTrainer.Options

Constructors

ImageClassificationTrainer.Options()

Fields

Arch

Specifies the model architecture to be used in the case of image classification training using transfer learning. The default Architecture is Resnet_v2_50.

BatchSize

Number of samples to use for mini-batch training. The default value for BatchSize is 10.

EarlyStoppingCriteria

Early stopping technique parameters to be used to terminate training when training metric stops improving. By default EarlyStopping is turned on and the monitoring metric is Accuracy.

Epoch

Number of training iterations. The default value for Epoch is 200.

FeatureColumnName

Column to use for features.

(Inherited from TrainerInputBase)
FinalModelPrefix

Final model and checkpoint files/folder prefix for storing graph files. The default prefix is "custom_retrained_model_based_on_".

LabelColumnName

Column to use for labels.

(Inherited from TrainerInputBaseWithLabel)
LearningRate

Learning rate to use during optimization. The default value for Learning Rate is 0.01.

LearningRateScheduler

A class that performs learning rate scheduling. The default learning rate scheduler is exponential learning rate decay.

MetricsCallback

Callback to report statistics on accuracy/cross entropy during training phase. Metrics Callback is set to null by default.

PredictedLabelColumnName

Name of the tensor that will contain the predicted label from output scores of the last layer when transfer learning is done. The default tensor name is "PredictedLabel".

ReuseTrainSetBottleneckCachedValues

Indicates to not re-compute cached bottleneck trainset values if already available in the bin folder. This parameter is set to false by default.

ReuseValidationSetBottleneckCachedValues

Indicates to not re-compute cached bottleneck validationset values if already available in the bin folder. This parameter is set to false by default.

ScoreColumnName

Name of the tensor that will contain the output scores of the last layer when transfer learning is done. The default tensor name is "Score".

TestOnTrainSet

Indicates to evaluate the model on train set after every epoch. Test on trainset is set to true by default.

TrainSetBottleneckCachedValuesFileName

Indicates the file name within the workspace to store trainset bottleneck values for caching, default file name is "trainSetBottleneckFile.csv".

ValidationSet

Validation set.

ValidationSetBottleneckCachedValuesFileName

Indicates the file name within the workspace to store validationset bottleneck values for caching, default file name is "validationSetBottleneckFile.csv".

ValidationSetFraction

When validation set is not passed then a fraction of train set is used as validation. To disable this behavior set ValidationSetFraction to null. Accepts value between 0 and 1.0, default value is 0.1 or 10% of the trainset.

WorkspacePath

Indicates the path where the image bottleneck cache files and trained model are saved, default is a new temporary directory.

Applies to