Open DianCh opened 3 years ago
we follow most common resnet-50 implementation (e.g. tensorflow, pytorch). There is a small change to original resnet for all these implementations (see ResNet-B in https://arxiv.org/pdf/1812.01187.pdf for explanation), but that shouldn't affect param count.
Hi, I loaded the r50_1x_sk0 checkpoint for ResNet-50, and it seems that the network architecture isn't exactly the original one in ResNet paper. However, I didn't see this is mentioned in the SimCLRv2 paper. Am I missing anything? If not, why did you modify the ResNet in this way?
Thank you!
(Just posting the printed architecture difference here)