Question about loss function

[Ivanyan0516]

In the class loss, you use the sigmoid focal loss with the binary cross-entropy loss， why not choose the softmax focal loss? And what may be the reason why the multi-classes loss doesn't converge？