why don't use trunc_normal_ in TimeAttention. The location is https://github.com/ucaswangls/STFormer/blob/main/cacti/models/stformer.py#L104

[ChunLiu-Claire]

In the forward function of SpaceAttention class trunc_normal_ is used for relative_position_bias_table, why don't use trunc_normal_ in TimeAttention's forward function?