请问一个can网络实现的细节

[chudouye]

if mode == "can":
        out_seq = []
        hh = []
        for i in range(orders):
            hh.append(his_items**(i+1))
        #hh = [sum(hh)]
        for i, h in enumerate(hh):
            if order_indep:
                weight, bias = weight_orders[i], bias_orders[i]
            else:
                weight, bias = weight_orders[0], bias_orders[0]
            for j, (w, b) in enumerate(zip(weight, bias)):
                h  = tf.matmul(h, w)
                if b is not None:
                    h = h + b
                if j != len(weight)-1:
                    h = tf.nn.tanh(h)
                out_seq.append(h)
        out_seq = tf.concat(out_seq, 2)

此处can的输出相当于包括了mlp每一层的输出，论文中是mlp最后一层的输出，请问应该以哪个为准呢

[wwwangzhch]

arXiv上最新的论文改成concat了