I think you have a problem with the focalloss

[xiaomao19970819]

In the function, you set the weights of the plus and minus samples to alpha
loss_1 = - gt * (alpha * backend.pow((1 - pr), gamma) * backend.log(pr))
loss_0 = - (1 - gt) * (alpha * backend.pow((pr), gamma) * backend.log(1 - pr))

[xiaomao19970819]

If anyone sees this question, I hope to get your help and answer, thank you very much
Error code located in '/segmentation_models/base/functional.py'

[qubvel]

Hi @xiaomao19970819
What is the correct way from your point of view?

[xiaomao19970819]

Hi @xiaomao19970819
What is the correct way from your point of view?

In my opinion,alpha is used to equalize positive and negative samples
So the weights of the positive and negative samples should not all be equal to alpha
this is my point :
loss_1 = - gt * (alpha * backend.pow((1 - pr), gamma) * backend.log(pr))
loss_0 = - (1 - gt) * ( (1-alpha) * backend.pow((pr), gamma) * backend.log(1 - pr))

[LeeJayHui]

Hi @xiaomao19970819, I agree with you. The implementation here is different from the formula of Focalloss, and it seems that your proposed modification is correct.
Hi @qubvel, maybe it's time to update the code.

[qubvel]

fixed #259