KL-Loss very large

[EternityZY]

The kl-loss has three parts :bbox_pred_std_abs_mulw_loss, bbox_pred_std_abs_logw_loss,loss_bbox.

When I add it, bbox_pred_std_abs_logw_loss will be a very large negative number, resulting in a final loss=nan. If only loss_bbox is optimized, then log_loss will become a very large positive number, making the final loss_bbox almost zero. How to calculate kl-loss I can reproduce in the code, but how can you train me in the end, can you help me?

[ethanhe42]

All three parts are needed in training, u cannot just optimize loss_bbox. bbox_pred_std_abs_mulw_loss, bbox_pred_std_abs_logw_loss produces gradients for std, loss_bbox produces gradients for bbox.
Are u training on ur own dataset or COCO?
If u train on COCO as the README suggests https://github.com/yihui-he/KL-Loss#training , u will get the same results as in the paper.
If u train on ur own dataset, there's nothing I can help.

[EternityZY]

Thanks for your reply, I am training COCO + YOLOv3, now bbox_pred_std_abs_logw_loss is a very small negative number, and it will get smaller and smaller.
Is these three items simple to add?

[ethanhe42]

yes, but as I said, bbox_pred_std_abs_mulw_loss, bbox_pred_std_abs_logw_loss produces gradients for std, loss_bbox produces gradients for bbox.

[csu-qwjy]

but, bbox_pred_std_abs_mulw_loss gradients will also go to bbox and std, because of bbox_pred_std_abs_mulw_loss=bbox_pred_std_abs*('bbox_pred'- 'bbox_targets'), resulting in a final loss=nan

[ethanhe42]

@csu-xiewei No, stopgradient prevents gradient goes to bbox

https://github.com/yihui-he/KL-Loss/blob/9242d088860c1d9990bbfb414f0800e7488b37c8/detectron/modeling/fast_rcnn_heads.py#L136