关于escm2算法的疑问

[huzuoliang]

def counterfact_ipw(self, loss_cvr, ctr_num, O, ctr_out_one):
    PS = paddle.multiply(
        ctr_out_one, paddle.cast(
            ctr_num, dtype="float32"))
    min_v = paddle.full_like(PS, 0.000001)
    PS = paddle.maximum(PS, min_v)
    IPS = paddle.reciprocal(PS)
    batch_shape = paddle.full_like(O, 1)
    batch_size = paddle.sum(paddle.cast(
        batch_shape, dtype="float32"),
                            axis=0)
    #TODO this shoud be a hyparameter
    IPS = paddle.clip(IPS, min=-15, max=15)  #online trick 
    IPS = paddle.multiply(IPS, batch_size)
    IPS.stop_gradient = True
    loss_cvr = paddle.multiply(loss_cvr, IPS)
    loss_cvr = paddle.multiply(loss_cvr, O)
    return paddle.mean(loss_cvr)

IPW方法为什么要乘batch_size，论文中也没发现需要做这个操作？IPS = paddle.multiply(IPS, batch_size)