简介

torch.autograd是PyTorch的自动微分引擎（自动求导），为神经网络训练提供动力。torch.autograd需要对现有代码进行最少的更改——声明需要计算梯度的Tensor的属性requires_grad=True。截至目前，PyTorch仅支持 FloatTensor类型（half、float、double和bfloat16）和 ComplexTensor（cfloat、cdouble）的autograd。【信息来自官网】

叶子节点

叶子结点是离散数学中的概念。一棵树当中没有子结点（即度为0）的结点称为叶子结点，简称“叶子”。叶子是指出度为0的结点，又称为终端结点。

在pytorch中，什么是叶子节点？根据官方定义理解如下。

所有requires_grad为False的张量，都约定俗成地归结为叶子张量
requires_grad为True的张量, 如果他们是由用户创建的,则它们是叶张量(leaf Tensor), 表明不是运算的结果，因此grad_fn=None

示例1

def test_training_pipeline2():
    input_data = [[4, 4, 4, 4],
                  [9, 9, 9, 9]]  # 2x4
    input = torch.tensor(input_data, dtype=torch.float32)  # requires_grad=False
    output = torch.sqrt(input)
    
    target_data = [1, 2, 3, 4]
    target = torch.tensor(target_data, dtype=torch.float32)  # requires_grad=False
    loss_fn = torch.nn.MSELoss()
    loss = loss_fn(input=output, target=target)
    print("\ninput.is_leaf:", input.is_leaf)
    print("output.requires_grad:", output.requires_grad)
    print("output.is_leaf:", output.is_leaf)
    print("target.is_leaf:", target.is_leaf)
    print("loss.requires_grad:", loss.requires_grad)
    print("loss.is_leaf:", loss.is_leaf)

样例2

def test_training_pipeline2():
    input_data = [[4, 4, 4, 4],
                  [9, 9, 9, 9]]  # 2x4
    input = torch.tensor(input_data, dtype=torch.float32)  # requires_grad=False
    output = torch.sqrt(input)
    output.requires_grad_(True) # requires_grad=True
    
    target_data = [1, 2, 3, 4]
    target = torch.tensor(target_data, dtype=torch.float32)  # requires_grad=False
    loss_fn = torch.nn.MSELoss()
    loss = loss_fn(input=output, target=target)
    
    print("\ninput.is_leaf:", input.is_leaf)
    print("output.requires_grad:", output.requires_grad)
    print("output.is_leaf:", output.is_leaf)
    print("target.is_leaf:", target.is_leaf)
    print("loss.requires_grad:", loss.requires_grad)
    print("loss.is_leaf:", loss.is_leaf)

样例3

def test_training_pipeline5():
    input = torch.rand(1, requires_grad=True)
    output = torch.unique(
        input=input, 
        sorted=True, 
        return_inverse=False, 
        return_counts=False, 
        dim=None
    )
    
    print("\ninput.is_leaf:", input.is_leaf)
    print("output.requires_grad:", output.requires_grad)
    print("output.is_leaf:", output.is_leaf)
    
    output.backward()

样例4

def test_training_pipeline3():
    input_data = [[4, 4, 4, 4],
                  [9, 9, 9, 9]]  # 2x4
    input_a = torch.tensor(input_data, dtype=torch.float32, requires_grad=True)
    input_b = torch.tensor(input_data, dtype=torch.float32, requires_grad=True)
    output = torch.ne(input_a, input_b)

    print("\ninput_a.is_leaf:", input_a.is_leaf)
    print("input_b.is_leaf:", input_b.is_leaf)
    print("output.dtype:", output.dtype)
    print("output.requires_grad:", output.requires_grad)
    print("output.is_leaf:", output.is_leaf)

    output.backward()   # 报错

样例5

def test_training_pipeline7():
    input_data = [[4, 4, 4, 4],
                  [9, 9, 9, 9]]  # 2x4
    input_a = torch.tensor(input_data, dtype=torch.float32, requires_grad=True)
    input_b = torch.tensor(input_data, dtype=torch.float32)    
    output = torch.add(input_a, input_b)
    print("\ninput_a.requires_grad:", input_a.requires_grad)
    print("input_b.requires_grad:", input_b.requires_grad)
    print("output.requires_grad:", output.requires_grad)
    print("output.is_leaf:", output.is_leaf)
    
    grad = torch.ones_like(output)
    
    input_b[0][0] = 10 
    input_a[0][0] = 10 
    output.backward(grad)

样例6

def test_training_pipeline9():
    x = torch.tensor([1.0], requires_grad=True)
    y = x + 2
    z = 2 * y		# <-- dz/dy=2
    y[0] = -2.0
    
    print("\nx.is_leaf:", x.is_leaf)
    print("y.is_leaf:", y.is_leaf)
    print("z.is_leaf:", z.is_leaf)
    print("\nx.requires_grad:", x.requires_grad)
    print("y.requires_grad:", y.requires_grad)
    print("z.requires_grad:", z.requires_grad)
    z.backward()


def test_training_pipeline9():
    x = torch.tensor([1.0], requires_grad=True)
    y = x + 2
   z = y * y  # <-- dz/dy= 2*y
    y[0] = -2.0
    
    print("\nx.is_leaf:", x.is_leaf)
    print("y.is_leaf:", y.is_leaf)
    print("z.is_leaf:", z.is_leaf)
    print("\nx.requires_grad:", x.requires_grad)
    print("y.requires_grad:", y.requires_grad)
    print("z.requires_grad:", z.requires_grad)
    z.backward()