JackAndCole / focal-loss-implementation-in-keras

Binary and Categorical Focal loss implementation in Keras.

Focal Loss

This is the keras implementation of focal loss proposed by Lin et. al. in their Focal Loss for Dense Object Detection paper.

Usage

Compile your model with focal loss as follows:

for binary

model.compile(loss=[binary_focal_loss(gamma=2)], metrics=["accuracy"], optimizer="adam")

for categorical

model.compile(loss=[categorical_focal_loss(gamma=2)], metrics=["accuracy"], optimizer="adam")

alpha setting:

model.fit(class_weight={0:alpha0, 1:alpha1, ...}, ...) the class_weight is the alpha of focal loss.

References

The implementation is based @umbertogriffo