about training loss about pytorch-yolov3 HOT 6 CLOSED

I think that may because the base lr is too large
when i change to 0.0001, the loss is correct

from pytorch-yolov3.

Are you computing loss for w/h directly or sqrt(w)/sqrt(h)?

from pytorch-yolov3.

@alexandru-dinu I tried both, all got nan

from pytorch-yolov3.

In the paper, they use sqrt for width and height loss calculation, but in order to do that w/h values must be positive. Sqrt of a negative number will give you nan.

Did you apply any transformation on the w/h values predicted by the network? That is, did you compare transformed (bx, by, bw, bh) or raw (tx, ty, tw, th) values with the ground truths?

from pytorch-yolov3.

This should be solved in the latest commit. Model does not converge unfortunately but training is stable and shouldn't result in nan:s for batch size 8.

from pytorch-yolov3.

I think that may because the base lr is too large
when i change to 0.0001, the loss is correct

wher is baer lr ?

from pytorch-yolov3.