梯度下降法求解3个向量的均值(未知向量到已知向量的Minimum Distance)

Minimum Distance

Write a python function def min_dist$(u_1,u_2,u_3)$
the functions takes in 3 lists of floats, each of length k, representing k-dimensional vectors. The function returns a list representing a k-dimensional vector v, such that
$|v-u_1{|}^2+|v-u_2{|}^2+|v-u_3{|}^2$ is as small as possible.
(Recall that the norm of a vector |w| is sqrt($w_1^2+w_2^2+...+w_k^2$).)
You MUST use gradient descent in order to complete this task. Make any reasonable algorithmic choices.

面试时遇到的题，有$u1,u2,u3$这3个向量，用梯度下降法找到向量$v$，使得$|v-u_1{|}^2+|v-u_2{|}^2+|v-u_3{|}^2$尽可能小。实际上$v$就是3个向量的均值，但需要用梯度下降法进行求解。

import numpy as np
import torch
np.random.seed(2333)
k = 10
u1 = torch.randn([k,1])
u2 = torch.randn([k,1])
u3 = torch.randn([k,1])
v = torch.zeros([k,1],requires_grad=True)

# 损失函数
def f(u1,u2,u3,v):
    return (v-u1) ** 2 + (v-u2) ** 2 + (v-u3) ** 2

# 优化方法
def sgd(params, lr):
    with torch.no_grad():
        for param in params:
            param -= lr * param.grad
            param.grad.zero_()

lr = 0.1
for epoch in range(10):
    y = f(u1,u2,u3,v)
    y.sum().backward()
    sgd([v], lr)
print(v.data)
print((u1+u2+u3)/3)

tensor([[-0.6027],
        [ 0.1890],
        [ 0.9928],
        [-0.1126],
        [-0.1857],
        [-0.1601],
        [ 0.3190],
        [-0.3286],
        [-0.0229],
        [ 0.4333]])
tensor([[-0.6027],
        [ 0.1890],
        [ 0.9929],
        [-0.1126],
        [-0.1857],
        [-0.1601],
        [ 0.3190],
        [-0.3287],
        [-0.0229],
        [ 0.4333]])

Process finished with exit code 0