Jiali MA

After we have got the overall heatmap, how can we know which part each local branch is focusing at?

> Hi @milliema > You may modify the code by replacing the `sum` with the index to visualize the heatmap of any specific layer. Thanks a lot for the quick...

> @wangzyon 感谢您的反馈 > > 1. 每个通道一个权重，所以取均值，求和都可以的，结果一样 > 2.第二个问题，是bug,应该是feature[proposal_idx]，感谢反馈！ 第二个问题不存在bug，proposal_idx对应的是预测的类别，与feature没有关系。 self.feature可以打印尺寸，是[1,C,H,W]的， 所以feature = self.feature[0].cpu().data.numpy() # [C,H,W]相当于去掉了第一个batchsize维度。

I agree with @swigicat , `"I get the feeling that doing more iterations in general affects the finder, because the maximum learning rate is found faster. "`. I got similar...

> May I ask what should be the correct dimension of grads? > Suppose the output feature map size is 100, and the batch_size is 512, should the rep and...

Now I have a new question: the assumption in the paper is that, all the task branches use the same feature map from backbone model. What if each task only...

> Not sure if this would help, but I encountered the same issue and had to rollback to an earlier version of apex: > > ``` > git checkout f3a960f80244cf9e80558ab30f7f7e8cbf03c0a0...

@timelikeswind 你的loss好平滑啊。我用reg模式训练，loss从10很快掉到0.5左右，维持一段之间之后减小到负数，稳定在-0.5。整个模型崩坏测试精度几乎为0。 想问下，根据你的实验这个方法表现如何？有提升吗？promising吗？我试了cls和reg模式都不太好

@JieChang-M Hi! Tried to contact via email yet the above email address is no more valid, I'd like to check about the official source code, is it still available?

May I ask what is the proper way to replace losses in FCOS with GHM losses? Have you tried to replace the regression loss as well?