org.nd4j.autodiff.loss.LossReduce Maven / Gradle / Ivy
package org.nd4j.autodiff.loss;
/**
* The LossReduce enum specifies how (or if) the values of a loss function should be reduced to a single value.
* See the javadoc comments on the individual enumeration constants for details.
*
* @author Alex Black
*/
public enum LossReduce {
/**
* No reduction. In most cases, output is the same shape as the predictions/labels.
* Weights (if any) are applied
* Example Input: 2d input array with mean squared error loss.
* Example Output: squaredDifference(predictions,labels), with same shape as input/labels
*/
NONE,
/**
* Weigted sum across all loss values, returning a scalar.
*/
SUM,
/**
* Weighted mean: sum(weights * perOutputLoss) / sum(weights) - gives a single scalar output
* Example: 2d input, mean squared error
* Output: squared_error_per_ex = weights * squaredDifference(predictions,labels)
* output = sum(squared_error_per_ex) / sum(weights)
*
* NOTE: if weights array is not provided, then weights default to 1.0 for all entries - and hence
* MEAN_BY_WEIGHT is equivalent to MEAN_BY_NONZERO_WEIGHT_COUNT
*/
MEAN_BY_WEIGHT,
/**
* Weighted mean: sum(weights * perOutputLoss) / count(weights != 0)
* Example: 2d input, mean squared error loss.
* Output: squared_error_per_ex = weights * squaredDifference(predictions,labels)
* output = sum(squared_error_per_ex) / count(weights != 0)
*
* NOTE: if weights array is not provided, then weights default to scalar 1.0 and hence MEAN_BY_NONZERO_WEIGHT_COUNT
* is equivalent to MEAN_BY_WEIGHT
*/
MEAN_BY_NONZERO_WEIGHT_COUNT
}