implementation details about the testing process

[Adasunnylily]

Hi, I have read your paper which is a great work. I also have tried to reproduce the linear probe results of baseline in Table2 in the cub dataset. However, the results I got are far below the number reported in paper, I got acc1 around 63.76 for Resnet50 pretrain while the results in paper is 68.17. For mocov2 pretrained results, I got acc around 59.92, but the results in paper is 68.30. Therefore I wonder if you could share your implementation details such as the parameters used in linear probe testing. Thanks a lot!

[GANPerf]

Hi, I have read your paper which is a great work. I also have tried to reproduce the linear probe results of baseline in Table2 in the cub dataset. However, the results I got are far below the number reported in paper, I got acc1 around 63.76 for Resnet50 pretrain while the results in paper is 68.17. For mocov2 pretrained results, I got acc around 59.92, but the results in paper is 68.30. Therefore I wonder if you could share your implementation details such as the parameters used in linear probe testing. Thanks a lot!

Hello! Did you try initializing the ResNet-50 architecture with ImageNet-trained weights, as described in our paper?

[Adasunnylily]

YES! I have done that~

[GANPerf]

YES! I have done that~

If that's the case, it is indeed a valid point to consider selecting the checkpoint with the best retrieval performance rather than relying solely on the last epoch during the pretraining process. It is possible that the checkpoint with the highest retrieval performance may exist during a mid-epoch, such as the 75th epoch, rather than the 100th epoch. Additionally, another factor that could contribute to the discrepancy could be the difference in classifier parameters. I will release them soon.

[Adasunnylily]

Thanks again! The release will be very helpful~

[bftan1949]

Hi, I have read your paper which is a great work. I also have tried to reproduce the linear probe results of baseline in Table2 in the cub dataset. However, the results I got are far below the number reported in paper, I got acc1 around 63.76 for Resnet50 pretrain while the results in paper is 68.17. For mocov2 pretrained results, I got acc around 59.92, but the results in paper is 68.30. Therefore I wonder if you could share your implementation details such as the parameters used in linear probe testing. Thanks a lot!

Hello, have you successfully reproduced the results of MoCo? (table2 68.30)