pytorch 动态图机制

动态图的含义是什么？

运算与搭建同时进行

---

动态图有什么好处？

灵活，易调节

---

动态图示例

torch.manual_seed(10)
W_h = torch.randn(size=(20,20),requires_grad=True)
W_x = torch.randn(size=(20,10),requires_grad=True)

x = torch.randn(size=(1,10))
prev_h = torch.randn(size=(1,20))

# print(prev_h.t().shape) t()的作用是转置
h2h = torch.mm(W_h,prev_h.t())
i2h = torch.mm(W_x,x.t())
next_h = h2h + i2h
next_h = next_h.tanh()

loss = next_h.sum()
print(loss)
loss.backward()
print(W_h.grad,W_x.grad)