Problems of reproducing the transfering results with SimSiam/MoCo-v2 on COCO

facebookresearch / simsiam

PyTorch implementation of SimSiam https//arxiv.org/abs/2011.10566

Other

1.15k stars 176 forks source link

Problems of reproducing the transfering results with SimSiam/MoCo-v2 on COCO #26

Closed thss15fyt closed 2 years ago

thss15fyt commented 2 years ago

Hi, thanks for your contribution to this interesting work!

I am trying to reproduce the transferring results on COCO, based on Detectron2 and MoCo codebase. I reached the 38.5 AP_box and 33.7 AP_mask of MoCO-v2, which leaves a gap compared with the reported 39.2 AP_box and 34.3 AP_mask. I have noticed that SimSiam searched the learning rate for all the methods. Thus, could you please let me know the best configurations to reproduce the reported results on COCO? Thanks!

endernewton commented 2 years ago

I think it is listed in the paper already (with 10x the learning rate and weight decay changes).