Closed qqwrz114 closed 6 years ago
I think you should try a smaller learning rate, or the gradient will explode.
Yes, the loss is not normalized to the image resolution. I would suggest either using a smaller learning rate or dividing the loss by the size of the output and using a universal (way smaller) loss.
Is this model used for specific image size(480P)? I tried 960x640 to train in osvos_demo.py, it failed. when i change the image size to 640x480,it work.