具有非标量输入的PyTorch自动分级

Question

以下内容在RuntimeError: grad can be implicitly created only for scalar outputs通话中给了我autograd.grad：

import torch
import torch.autograd as autograd

# Note `x` and `other_x` are both torch tensors of shape torch.Size([1, 20]), with values [-1.0, 1.0]
# and `y` is a torch tensor of shape torch.Size([7, 1]), with values [-1.0, 1.0]

def my_func(x):
    # compute something that's a different shape from input
    return y

def closure():
    minimizer.zero_grad()
    y = my_func(x)
    autograd.backward(other_x, autograd.grad(y, other_x))
    return y

minimizer.step(closure)
return x.detach()

但是为什么呢？对于标量输入（张量均为torch.Size([1])形状），此操作按预期运行。我无法从autograd docs看出问题所在。预先感谢！