Closed scihacker closed 6 years ago
I agree with your opinion and I found there is already a PR to fix it. But I'm also confused whether it's fair to train with val data.
Wanted to thank you for pointing this out! We tried to get the PR up as quickly as possible - though I forgot to merge it.
Just to clarify, this was something that accidentally slipped through when we were refactoring for the public code release, and we didn't train on validation data on the experiments in our paper.
Have anyone got the new mAP on validation data?
@DouYishun I found that in the latest leaderboard of COCO2017, umich-vl method only scores 0.46 in AP.
@ahangchen Hi! Why did this happen? If the refinement step is removed, what AP will be? Thank you.
I found the loss decreasing when the network is passing through validation data. So I checked the code and found:
The trainer is built in the task path:
task/pose.py
, whereSo in my opinion, data in validation phase is also trained as those in the training phase. Could you explain this to me, please? Thanks.