affine:为True时表示γ和β是可学习的参数,为False表示γ和β是不可学习的参数,此时γ=1,β=0; track_running_stats=True;整个batch的方差和均值
As you've noted, the num_features arg to InstanceNorm1d is only used for the affine=True or track_running_stats=True cases. While we could tighten up the error checking outside of these cases, I'm a bit worried about people relying on this lax behavior at the moment (i.e. it ...