PyTorch - autograd自动微分

论文：Automatic Differentiation in Machine Learning: a Survey，自动微分

参考 PyTorch：AUTOMATIC DIFFERENTIATION WITH TORCH.AUTOGRAD

Loss是标量，雅可比向量积，JVP，

primitive operation(原始操作)：

torch.autograd()，计算梯度

import torch

x = torch.ones(5)  # input tensor, 输入向量
print(f"x: {x}")
y = torch.zeros(3)  # expected output, 标签
print(f"y: {y}")
w = torch.randn(5, 3, requires_grad=True)  # 开启自动微分
print(f"w: {w}")
b = torch.randn(3, requires_grad=True)
print(f"b: {b}")
z = torch.matmul(x, w)+b

loss = torch.nn.functional.binary_cross_entropy_with_logits(z, y)

使用梯度反向传播算法，back propagation

backward()是Tensor类的方法，loss是标量直接调用backward()，loss如果是张量，则backward()需要传入张量

print(f"Gradient function for z = {z.grad_fn}")
print(f"Gradient function for loss = {loss.grad_fn}")

loss.backward()
print(w.grad)
print(b.grad)

retain_graph=True，保留图，不保留的话，第2次调用会报错：RuntimeError: Trying to backward through the graph a second time

torch.no_grad()关闭自动求导：

z = torch.matmul(x, w)+b
print(z.requires_grad)

with torch.no_grad():
    z = torch.matmul(x, w)+b
print(z.requires_grad)

z = z.detach()之后，z.requires_grad是False

z = z.detach()
z.requires_grad

DAG：directed acyclic graph，有向无环图

张量loss，输入torch.ones_like(inp)，反向传播

retain_graph保留图，可以连续backward()

梯度置0，np.grad.zero_()

inp = torch.eye(5, requires_grad=True)
out = (inp+1).pow(2)
print(out)
out.backward(torch.ones_like(inp), retain_graph=True)
print(f"First call\n{inp.grad}")
out.backward(torch.ones_like(inp), retain_graph=True)
print(f"\nSecond call\n{inp.grad}")
inp.grad.zero_()
out.backward(torch.ones_like(inp), retain_graph=True)
print(f"\nCall after zeroing gradients\n{inp.grad}")