About the performance

[Whalesong-zrs]

I use the same setting,include number of GPUS, batch size(all the same as the yaml in the config),hoe_sigma=4,however,my result is like this
acc5 acc15 acc22.5 acc30 acc45 68.2 89.7 93.1 96.1 97.5 and I only focus on the orientation problem so I don't use joint point.Is the random seed changes the performance?

[ChenyanWu]

Your performance is close to the paper - acc5 68.6 acc15 90.7 acc30 96.9. I think it is in the range of random noise. CUDA version, GPU difference, Pytorch version, and random seed may all slightly influence the results. You may tune the batch size or learning rate a little bit to achieve the optimal performance in your computing environment.