Settings on whether to use affine

[Yuan1z0825]

The author's contribution toward eliminating artifacts is commendable; however, there are a few points that remain unclear. Specifically, I found that the affine setting of dense normalization during training appears to be ineffective. Yet, the weights and biases associated with this affine setting are still utilized during transfer. Could you clarify the rationale behind this discrepancy?