which eval metric should I use to decide stop train?
captainfffsama opened this issue · 1 comments
Thx for you brilliant idea, I have a question about that which eval metric should I refer to know model is over fitting or not fitting?
Here is the eval metrics change during I set img size 960*544 and batchsize 1 train on coco train dataset, eval is on HPatch and img size is 320*240:
It seems that after epoch 3, M Score and Repeatability is all the way down not up...
If I train on my own image dataset ,which eval metric should I refer to stop train?
As you train for longer Loc_error should be going down, and corr_1 should be going up. As for stopping training, I think corr1 could be a good metric as it's quite strict. But it also depends on potential downstream applications, for example pose estimation, that you could use to check the performance of the network.