What gpu did you use?

[YanShuo1992]

I write to reproduce the experiments.
I see the checkpoints have been trained for 100000 epoches.
But I don't know the batchsize.
Could you please provide more details?

Cheers!

[daibiaoxuwu]

The work was pretty ancient and at that time I used a single NVIDIA 2080Ti. The batch size was adjusted with SF because larger SF means larger memory usage in the classifier fully-connected layers. Other parameters like traning epoches and learning rates are also adjusted with SF, larger SF typically takes longer to learn.