nn.Module是否要通过forward才能进行反向传播？

先说结论不需要，其他的自定义函数也行的

import torch
 
import torch
import torch.nn as nn
 
 
class g(nn.Module):
    def __init__(self):
        super(g, self).__init__()
        self.k = nn.Conv2d(in_channels=2, out_channels=1, kernel_size=1, padding=0, bias=False)
 
    def forward(self, z):
        return self.k(z)
    
    def encode(self, z):
        return self.k(z)
 

c = 2
h = 5
w = 5
z = torch.rand( (1,c , h , w)).float().view(1, c, h, w)*100
z.requires_grad = True
k = g()
r = k.encode(z)
r = r.sum()
loss = (r - 1) * (r - 1)
for name,v in k.named_parameters():
    print(name,v.grad) 
print("*********************")

loss.backward()
for name,v in k.named_parameters():
    print(name,v.grad)

结果

k.weight None

***************************

k.weight tensor([[[[426972.1875]],

[[383235.8438]]]])