對pytroch中torch.autograd.backward的思考

反向傳遞法則是深度學習中最為重要的一部分，torch中的backward可以對計算圖中的梯度進行計算和累積

這裡通過一段程式來演示基本的backward操作以及需要注意的地方

>>> import torch
>>> from torch.autograd import Variable

>>> x = Variable(torch.ones(2,2), requires_grad=True)
>>> y = x + 2
>>> y.grad_fn
Out[6]: <torch.autograd.function.AddConstantBackward at 0x229e7068138>
>>> y.grad

>>> z = y*y*3
>>> z.grad_fn
Out[9]: <torch.autograd.function.MulConstantBackward at 0x229e86cc5e8>
>>> z
Out[10]: 
Variable containing:
 27  27
 27  27
[torch.FloatTensor of size 2x2]
>>> out = z.mean()
>>> out.grad_fn
Out[12]: <torch.autograd.function.MeanBackward at 0x229e86cc408>
>>> out.backward()     # 這裡因為out為scalar標量，所以引數不需要填寫
>>> x.grad
Out[19]: 
Variable containing:
 4.5000  4.5000
 4.5000  4.5000
[torch.FloatTensor of size 2x2]
>>> out   # out為標量
Out[20]: 
Variable containing:
 27
[torch.FloatTensor of size 1]

>>> x = Variable(torch.Tensor([2,2,2]), requires_grad=True)
>>> y = x*2
>>> y
Out[52]: 
Variable containing:
 4
 4
 4
[torch.FloatTensor of size 3]
>>> y.backward() # 因為y輸出為非標量，求向量間元素的梯度需要對所求的元素進行標註，用相同長度的序列進行標註
Traceback (most recent call last):
  File "C:UsersdellAnaconda3envsmy-pytorchlibsite-packagesIPythoncoreinteractiveshell.py", line 2862, in run_code
    exec(code_obj, self.user_global_ns, self.user_ns)
  File "<ipython-input-53-95acac9c3254>", line 1, in <module>
    y.backward()
  File "C:UsersdellAnaconda3envsmy-pytorchlibsite-packages	orchautogradvariable.py", line 156, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph, retain_variables)
  File "C:UsersdellAnaconda3envsmy-pytorchlibsite-packages	orchautograd\__init__.py", line 86, in backward
    grad_variables, create_graph = _make_grads(variables, grad_variables, create_graph)
  File "C:UsersdellAnaconda3envsmy-pytorchlibsite-packages	orchautograd\__init__.py", line 34, in _make_grads
    raise RuntimeError("grad can be implicitly created only for scalar outputs")
RuntimeError: grad can be implicitly created only for scalar outputs

>>> y.backward(torch.FloatTensor([0.1, 1, 10]))
>>> x.grad                #注意這裡的0.1,1.10為梯度求值比例
Out[55]: 
Variable containing:
  0.2000
  2.0000
 20.0000
[torch.FloatTensor of size 3]

>>> y.backward(torch.FloatTensor([0.1, 1, 10]))
>>> x.grad                # 梯度累積
Out[57]: 
Variable containing:
  0.4000
  4.0000
 40.0000
[torch.FloatTensor of size 3]

>>> x.grad.data.zero_()  # 梯度累積進行清零
Out[60]: 
 0
 0
 0
[torch.FloatTensor of size 3]
>>> x.grad              # 累積為空
Out[61]: 
Variable containing:
 0
 0
 0
[torch.FloatTensor of size 3]
>>> y.backward(torch.FloatTensor([0.1, 1, 10]))
>>> x.grad
Out[63]: 
Variable containing:
  0.2000
  2.0000
 20.0000
[torch.FloatTensor of size 3]