xuzhiyuan1022

thank you for your reply. grad_alpha = (grad_output * (sign * i + (input_q - input) * (1 - i))).sum() when update alpha,why use (input_q-input)