Huber

public class Huber

Computes the Huber loss between labels and predictions.

For each value x in error = labels - predictions: 

     loss = 0.5 * x^2                  if |x| <= d
     loss = 0.5 * d^2 + d * (|x| - d)  if |x| > d

where d is delta.

Standalone usage: 

    Operand<TFloat32> labels =
        tf.constant(new float[][] { {0.f, 1.f}, {0.f, 0.f} });
    Operand<TFloat32> predictions =
        tf.constant(new float[][] { {0.6f, 0.4f}, {0.4f, 0.6f} });
    Huber huberLoss = new Huber(tf);
    Operand<TFloat32> result = huberLoss.call(labels, predictions);
    // produces 0.155

Calling with sample weight: 

    Operand<TFloat32> sampleWeight = tf.constant(new float[] {1.f, 0.f});
    Operand<TFloat32> result = huberLoss.call(labels, predictions, sampleWeight);
    // produces 0.09f

Using SUM reduction type: 

    Huber huberLoss = new Huber(tf, Reduction.SUM);
    Operand<TFloat32> result = huberLoss.call(labels, predictions);
    // produces 0.32f

Using NONE reduction type: 

    Huber huberLoss = new Huber(tf, Reduction.NONE);
    Operand<TFloat32> result = huberLoss.call(labels, predictions);
    // produces [0.18f, 0.13f]

See Also

Constants

float DELTA_DEFAULT

Inherited Fields

Public Constructors

Huber(Ops tf)
Creates a Huber Loss using getSimpleName() as the loss name, DELTA_DEFAULT as the delta and a Loss Reduction of REDUCTION_DEFAULT
Huber(Ops tf, String name)
Creates a Huber Loss using DELTA_DEFAULT as the delta and a Loss Reduction of REDUCTION_DEFAULT
Huber(Ops tf, Reduction reduction)
Creates a Huber Loss using getSimpleName() as the loss name and and DELTA_DEFAULT as the delta
Huber(Ops tf, String name, Reduction reduction)
Creates a Huber Loss using DELTA_DEFAULT as the delta
Huber(Ops tf, String name, float delta, Reduction reduction)
Creates a Huber Loss

Public Methods

<T extends TNumber> Operand<T>
call(Operand<? extends TNumber> labels, Operand<T> predictions, Operand<T> sampleWeights)
Generates an Operand that calculates the loss.

Inherited Methods

Constants

public static final float DELTA_DEFAULT

Constant Value: 1.0

Public Constructors

public Huber (Ops tf)

Creates a Huber Loss using getSimpleName() as the loss name, DELTA_DEFAULT as the delta and a Loss Reduction of REDUCTION_DEFAULT

Parameters
tf the TensorFlow Ops

public Huber (Ops tf, String name)

Creates a Huber Loss using DELTA_DEFAULT as the delta and a Loss Reduction of REDUCTION_DEFAULT

Parameters
tf the TensorFlow Ops
name the name of the loss, if null then getSimpleName() is used.

public Huber (Ops tf, Reduction reduction)

Creates a Huber Loss using getSimpleName() as the loss name and and DELTA_DEFAULT as the delta

Parameters
tf the TensorFlow Ops
reduction Type of Reduction to apply to the loss.

public Huber (Ops tf, String name, Reduction reduction)

Creates a Huber Loss using DELTA_DEFAULT as the delta

Parameters
tf the TensorFlow Ops
name the name of the loss, if null then getSimpleName() is used.
reduction Type of Reduction to apply to the loss.

public Huber (Ops tf, String name, float delta, Reduction reduction)

Creates a Huber Loss

Parameters
tf the TensorFlow Ops
name the name of the loss, if null then getSimpleName() is used.
delta the point where the Huber loss function changes from quadratic to linear.
reduction Type of Reduction to apply to the loss.

Public Methods

public Operand<T> call (Operand<? extends TNumber> labels, Operand<T> predictions, Operand<T> sampleWeights)

Generates an Operand that calculates the loss.

Parameters
labels the truth values or labels
predictions the predictions
sampleWeights Optional sampleWeights acts as a coefficient for the loss. If a scalar is provided, then the loss is simply scaled by the given value. If SampleWeights is a tensor of size [batch_size], then the total loss for each sample of the batch is rescaled by the corresponding element in the SampleWeights vector. If the shape of SampleWeights is [batch_size, d0, .. dN-1] (or can be broadcast to this shape), then each loss element of predictions is scaled by the corresponding value of SampleWeights. (Note on dN-1: all loss functions reduce by 1 dimension, usually axis=-1.)
Returns
  • the loss