Difference between code and paper

[lizleo]

When

so

then

but in your code https://github.com/zhangxiaosong18/FreeAnchor/blob/master/maskrcnn_benchmark/modeling/rpn/free_anchor_loss.py#L161,
you just use (matched_cls_prob in your code) as ,
that means you just ignore the other predicted classes which not matching the target class, and I think it's different with retinanet_cls_loss defined in https://github.com/zhangxiaosong18/FreeAnchor/blob/master/maskrcnn_benchmark/modeling/rpn/retinanet_loss.py#L142.

I try to rewrite the code calculating matched_cls_prob as blew:

labels_mul = torch.zeros([len(labels_), self.num_classes])
for i in range(len(labels_)):
    labels_mul[i, labels_[i]] = 1

labels_mul = labels_mul.unsqueeze(1).repeat(1, self.pre_anchor_topk, 1)

loss_mul_class = nn.BCELoss(reduction="none")(cls_prob_[matched], labels_mul).sum(dim=-1)
matched_cls_prob = (-loss_mul_class).exp()

Did I get it wrong ？ @zhangxiaosong18

[liuyang-ict]

You know if b_{i}^{cls}[k] =1, the only way it can happen is if k = i, because b_{i}^{cls} is a one-hot label.