wh_loss is too large

[unlabeledData]

Hi, I implemented the centernet on TF and used my own backbone net.

Then I found that when I trained my net on Pascal VOC 07+12, the hm_loss and off_loss decreased normally but wh_loss was oscillating around 20. ( hm_loss weight=1.0, wh_loss weight=0.1, off_loss weight=1.0)

Could you share some suggestions about how to set loss weights and how to tune the network to make it converge?

[xingyizhou]

I will suggest using our backbone network and fully reproduce our results first. This makes sure all details are correctly implemented.

[fuqichen1998]

I have the same question when I just changed the backbone to a smaller and not pretrained backbone.

[ShawnNew]

Hi,

Does the initialization of IDAUp matters while training CenterNet?

I re-implemented one model started training from scratch. The loss-hm plateaus to 9.208 and the network stopped converging. To be specific, the negative part of the heatmap focal loss ends up all zero.

But when I initialized the model with your implementation, the network trained normally.

How do you suggest debugging it?

[Hello-Felix]

Hi,

Does the initialization of IDAUp matters while training CenterNet?

I re-implemented one model started training from scratch. The loss-hm plateaus to 9.208 and the network stopped converging. To be specific, the negative part of the heatmap focal loss ends up all zero.

But when I initialized the model with your implementation, the network trained normally.

How do you suggest debugging it?

I had the same problem. Do u have solve it?

[G-Apple1]

I had the same problem. I found that the boxmode of coco datasets is XYWH_abs. But the boxmode I set was XYXY_abs. I had solved the problem.

[vivcheng01]

Hi @G-Apple1 can you elaborate on your solution?

[G-Apple1]

Hi @G-Apple1 can you elaborate on your solution?

抱歉，我直接用中文回复。我当时的情况是生成coco数据集的时候出了问题，我用labelimg标注的软件是xyxy（左上角和右下角坐标），转成coco的时候是用的xywh（我以为是中心点坐标和宽高），实际上coco的是左上角坐标和宽高，改过来就没事了。

[vivcheng01]

OK，谢谢！