Reproducing MoCo on ImageNet-1k.

[hobbitlong]

Reproducing MoCo on ImageNet-1k.

[HobbitLong]

Hi, @yzou2,

Thanks for your kind words! I have reproduced his results (60.6 with 64 negatives) on ImageNet. I strictly followed his pre-training recipe (8GPUs, 256 batch size, 200 epochs, blabla), and used a Distributed Dataparallel training code (I will release this once I got some time).

[yzou2]

Thanks a lot, Yonglong! Hope to see the code soon.

[happywu]

Hi @HobbitLong ,

Did you modify other parts than this distributed data parallel training (will this distributed training influence accuracy)? Could you please tell that If I directly change the configuration to (8Gpus, 256 batch size, 200 epochs) without changing current code, would this achieve comparable performance?

Thanks!

[HobbitLong]

Hi, @yzou2, @happywu,

Maybe this is a bit late, but just let you know that the DDP implementation for MoCo (as well as InstDis, MoCo v2, PIRL, ...) have been released here.

[yzou2]

Hi, @yzou2, @happywu,

Maybe this is a bit late, but just let you know that the DDP implementation for MoCo (as well as InstDis, MoCo v2, PIRL, ...) have been released here.

It's really great! Thanks!