Closed poodarchu closed 3 years ago
I've previously tested in cifar10 (800 epochs) without a static lr for the projection head and without zero_init_residual and it converged to 91.6% acc.
I'm currently running on imagenet with both tricks implemented and I'll update the repo as soon as it's done.
I've also reach the same performance as the paper on CIFAR 10, but it may be not the case as for Imagenet-1K
I update my results of SimSiam here: https://github.com/poodarchu/SelfSup/blob/master/examples/simsiam/SimSiam.res50.imagenet.256bs.224size.100e.lin_cls/README.md 100 epoch on ImageNet-1K can reach 67.6% (paper 67.7%, without constant lr in MLP head).
Can you share your results on the full Imagenet-1k dataset? I think the results on imagenet-100 is not sufficient to prove that the implementation is right.
My implementation here can reach 65% top-1 acc on ImageNet 1k (using mocov2's linear evaluation protocol), https://github.com/poodarchu/SelfSup/blob/master/examples/simsiam/SimSiam.res50.imagenet.256bs.224size.100e.lin_cls/README.md, as reported in the paper, this number should be ~66.7%。
Moreover, my results on CIFAR can match the performance in paper exactly.