I have trained the efficientnet in recalculated representations of UTD-MHAD (Inertial and Skeleton) provided by the project and UTD-MHAD(fused Inertial and Skeleton using nb_generate.py). However, the training results are inconsistent with the paper.Then, I checked the code finding that the transform config is "minimal" which means no transform used .So could you tell me what's the transform used in the paper.And I also want to know the approach_id and attention to fuse Inertial and Skeleton using nbgenerate.npy.Thanks for your help.