Closed weleen closed 6 years ago
https://github.com/chenyilun95/tf-cpn/blob/7c01aba69b8b9ad14fbb36d409f850bafd345dd5/models/COCO.res50.256x192.CPN/dataset.py#L156 I am curious about the generation of heatmap, why multiply 255 here? It seems that the network can not converge when i remove 255.
The prediction heatmap is expected to range in [0...255], which affects the network loss. I think it works if you increase learning rate properly and remove 255.
https://github.com/chenyilun95/tf-cpn/blob/7c01aba69b8b9ad14fbb36d409f850bafd345dd5/models/COCO.res50.256x192.CPN/dataset.py#L156 I am curious about the generation of heatmap, why multiply 255 here? It seems that the network can not converge when i remove 255.