Hello, can I ask you about the accuracy of cifar-10, ImageNet100?
minkyujeon opened this issue · 3 comments
I experimented in cifar10, ImageNet100 in ResNet18 by using this code, but the accuracy of linear classifier is
cifar10 : 82.77 (pcl_r : 512, num_cluster : [1000,2000,4000] / pcl_r 1024, num_cluster : [1500,3000,4500])
ImageNet100 : 64.40 (pcl_r : 1024, num_cluster : [2500,5000,10000] / pcl_r : 4096, num_cluster : [6000,8000,12000])
Even I adjusted the hyper-parameter, It can not be achieved other method's performance (byol, etc..). If you know the reason, could you tell me why? Thank you
Hi, could you get good performance with moco? PCL can be built on-top-of existing SSL methods by adding the clustering loss, so it could be better to start with a set of hyper-parameters that is good for moco or simclr.
Okay I'll check it. Thank you for replying!
@jeonminkyu
What were the final results you got on CIFAR10? and what config did you use? also did you use SIMCLR or MOCO?