python - 用 PyTorch 绘制函数的导数？

Question

我有这个代码:

import torch
import matplotlib.pyplot as plt  
x=torch.linspace(-10, 10, 10, requires_grad=True)
y = torch.sum(x**2)
y.backward()
plt.plot(x.detach().numpy(), y.detach().numpy(), label='function')
plt.legend()

但是，我收到了这个错误:

ValueError: x and y must have same first dimension, but have shapes (10,) and (1,)

Answer 1

我认为主要问题是你们的尺寸不匹配。为什么你不想使用torch.sum？

这应该适合你:

# %matplotlib inline added this line only for jupiter notebook
import torch
import matplotlib.pyplot as plt  
x = torch.linspace(-10, 10, 10, requires_grad=True)

y = x**2      # removed the sum to stay with the same dimensions
y.backward(x) # handing over the parameter x, as y isn't a scalar anymore
# your function
plt.plot(x.detach().numpy(), y.detach().numpy(), label='x**2')
# gradients
plt.plot(x.detach().numpy(), x.grad.detach().numpy(), label='grad')
plt.legend()

虽然步骤更多，但您会得到更好的图片，我还将间隔稍微更改为 torch.linspace(-2.5, 2.5, 50, require_grad=True)。

编辑评论:

此版本为您绘制包含 torch.sum 的渐变:

# %matplotlib inline added this line only for jupiter notebook
import torch
import matplotlib.pyplot as plt  
x = torch.linspace(-10, 10, 10, requires_grad=True)

y = torch.sum(x**2) 
y.backward() 
print(x.grad)
plt.plot(x.detach().numpy(), x.grad.detach().numpy(), label='grad')
plt.legend()

输出:

tensor([-20.0000, -15.5556, -11.1111,  -6.6667,  -2.2222,   2.2222,
      6.6667,  11.1111,  15.5556,  20.0000])

剧情: