如果以后有人找这个：我和特遣部队一起完成了这一切

# def create_custom_loss(constant=1.0):
def create_nearest_other_neighbour(mybatchsize):
# def custom_loss(y_true, y_pred):
    def nearest_other_neighbour(B, A):
        na = tf.reduce_sum(tf.square(A), 1)
        nb = tf.reduce_sum(tf.square(B), 1)

        # na as a row and nb as a column vectors
        na = tf.reshape(na, [-1, 1])
        nb = tf.reshape(nb, [1, -1])

        # return pairwise euclidean difference matrix
        dists = tf.sqrt(na - 2*tf.matmul(A, B, False, True) + nb)

        # extract diag of dist matrix (distance from predicted self to real self), and replace with dummy vals that won't interfere with minimums
        d = tf.linalg.tensor_diag_part(dists)

        newdiag = tf.repeat(tf.reduce_max(dists), mybatchsize)
        dists_nodiag = tf.linalg.set_diag(dists, newdiag)

        # get minimum of each column of new matrix, then divide diagonal by these mins
        # (done in return for fewer variables)
        # min_percol = tf.math.reduce_min(dists_nodiag, 1)
        # dist_ratio = tf.math.divide(d, min_percol)
        # return dist_ratio

        return tf.math.divide(d, tf.math.reduce_min(dists_nodiag, 1))
    
    return nearest_other_neighbour

这需要使用datasets来管理输入：train_dataset = tf.data.Dataset.from_tensor_slices((train_x, train_y)).batch(mybatch, drop_remainder=True)