How to use the babi.model.VectorReductionLayer function in babi

with tf.variable_scope("embedding"):
            A = VariableEmbedder(params, wd=wd, initializer=initializer, name='A')
            Aq = A(q, name='Aq')  # [N, S, J, d]
            Ax = A(x, name='Ax')  # [N, S, J, d]

        with tf.name_scope("encoding"):
            encoder = PositionEncoder(J, d)
            u = encoder(Aq, q_mask)  # [N, d]
            m = encoder(Ax, x_mask)  # [N, M, d]

        with tf.variable_scope("networks"):
            m_mask = tf.reduce_max(tf.cast(x_mask, 'int64'), 2, name='m_mask')  # [N, M]
            gate_mask = tf.expand_dims(m_mask, -1)
            m_length = tf.reduce_sum(m_mask, 1, name='m_length')  # [N]
            prev_u = tf.tile(tf.expand_dims(u, 1), [1, M, 1])  # [N, M, d]
            reg_layer = VectorReductionLayer(N, M, d) if use_vector_gate else ReductionLayer(N, M, d)
            gate_size = d if use_vector_gate else 1
            h = None  # [N, M, d]
            as_, rfs, rbs = [], [], []
            hs = []
             
            for layer_idx in range(L):
                with tf.name_scope("layer_{}".format(layer_idx)):
                    dr_prev_u = tf.nn.dropout(prev_u, 0.7) if params.use_dropout else prev_u
                    u_t = tf.tanh(linear([dr_prev_u, m], d, True, wd=wd, scope='u_t'))
                    a = tf.cast(gate_mask, 'float') * tf.sigmoid(linear([dr_prev_u * m], gate_size, True, initializer=initializer, wd=wd, scope='a') - att_forget_bias)
                    h = reg_layer(u_t, a, 1.0-a, scope='h')
                    if layer_idx + 1 < L:
                        if params.use_reset:
                            rf, rb = tf.split(2, 2, tf.cast(gate_mask, 'float') *
                                tf.sigmoid(linear([dr_prev_u * m], 2 * gate_size, True, initializer=initializer, wd=wd, scope='r')))
                        else: