为什么 pytorch 没有为我最小化 x*x? [英] Why pytorch isn't minimizing x*x for me?

问题描述

我希望 x 收敛到 0，这是 x*x 的最小值.但这不会发生.我在这个小示例代码中做错了什么:

I expect x to converge to 0, which is minimum of x*x. But this doesn't happen. What am I doing wrong in this small sample code:

import torch
from torch.autograd import Variable
tns = torch.FloatTensor([3])
x = Variable(tns, requires_grad=True)
z = x*x
opt = torch.optim.Adam([x], lr=.01, betas=(0.5, 0.999))
for i in range(3000):
    z.backward(retain_graph=True) # Calculate gradients
    opt.step()
    print(x)

推荐答案

您遇到的问题是在计算每个循环时没有将梯度归零.相反，通过在循环的每一步设置 retain_graph=True 而不是调用 opt.zero_grad()，您实际上是将计算的梯度添加到 ALL 之前的梯度计算.因此，您不是在梯度下降中迈出一步，而是在所有累积梯度方面迈出了一步，这当然不是您想要的.

The problem you have is that you don't zero the gradients when you are calculating each loop. Instead, by setting retain_graph=True and not calling opt.zero_grad() at each step of the loop you are actually adding the gradients calculated to ALL previous gradients calculated. So instead of taking a step in gradient descent, you are taking a step with respect to all accumulated gradients which is certainly NOT what you want.

您应该确保在循环开始时调用 opt.zero_grad()，并将 z=x*x 移动到循环内，以便您不必retain_graph.

You should instead make sure to call opt.zero_grad() at the beginning of your loop, and move the z=x*x inside your loop so that you don't have to retain_graph.

我做了这些细微的修改:

I made these slight modifications:

import torch
from torch.autograd import Variable
tns = torch.FloatTensor([3])
x = Variable(tns, requires_grad=True)
opt = torch.optim.Adam([x], lr=.01, betas=(0.5, 0.999))
for i in range(3000):
    opt.zero_grad()
    z = x*x
    z.backward() # Calculate gradients
    opt.step()
    print(x)

我最终的 x 是 1e-25.

这篇关于为什么 pytorch 没有为我最小化 x*x?的文章就介绍到这了，希望我们推荐的答案对大家有所帮助，也希望大家多多支持IT屋！