Focal_Loss_Keras
Focal_Loss_Keras copied to clipboard
Loss is high in older implementation
Tough its not an issue but more of a confusion. in the paper alpha is set to .25 you have set it to 4 When i try other available implementations, using same alpha and gamma value i get high loss value tough it end up curbing impurity prediction in minority class quite a bit using your implementation tough loss value is low but in accuracy it tends to increase true minority class prediction but to small extent increases yhe false predictions of majority class