search

Duankaiwen / CenterNet

Codes for our paper "CenterNet: Keypoint Triplets for Object Detection" .
MIT License
1.87k stars 384 forks source link

Are these losses normal when I train in coco #100

Open xuyongxi opened 5 years ago

xuyongxi commented 5 years ago

focal loss at iteration 30995: 7.344306945800781 pull loss at iteration 30995: 0.05117112770676613 push loss at iteration 30995: 0.10568550229072571 regr loss at iteration 30995: 0.21536898612976074 training loss at iteration 31000: 7.482662200927734 focal loss at iteration 31000: 7.119882583618164 pull loss at iteration 31000: 0.044373102486133575 push loss at iteration 31000: 0.1216396912932396 regr loss at iteration 31000: 0.19676724076271057 validation loss at iteration 31000: 8.355660438537598

The focal loss is much bigger than other losses, when I train in coco, the total loss can't converge. Who know why?

fuqichen1998 commented 4 years ago

I have the same question, could @Duankaiwen explain if it is normal?

Duankaiwen commented 4 years ago

Hi, the losses are normal. These losses converge slowly. BTW, what's your batch size? @ xuyongxi @fuqichen1998

fuqichen1998 commented 4 years ago

Thanks! My batch size is 16 for CenterNet-52.

JoJoliking commented 4 years ago

@fuqichen1998 I have set batch size 16 for CenterNet52. And the learning rate is changed into 2.5e-5 in 1000 epoch. (Using the CenterNet-480000 for pretianed model). Have you achieve the author's result?