In forward process,i need loops multiple times. How can Backward update the only once loop

1 
class Fun(torch.autograd.Function):

    @staticmethod
    def forward(ctx, *args, **kwargs):


    @staticmethod
    def backward(ctx, *grad_outputs):

class NN(nn.Module):
    def __init__(self):
        super(NN, self).__init__()
        self.fc1 = nn.Linear(784, 800)
        self.fc2 = nn.Linear(800, 10)

    def forward(self, inpu, L=5):
        for step in rang(L):
            x = input.view(batch,-1)
            x = self.fc1(x)
            x = self.fc2(x)
        return x

nn = NN()
criterion = nn.MSELoss()
optimizer = torch.optim.Adam(nn.parameters(), lr=learning_rate,)

for epoch in range(e):
    for i, (x, y) in enumerate(train_loader):
        outputs = nn(x)
        y_ = torch.zeros(batch_size, 10).scatter_(1, y.view(-1, 1), 1)# ont-hot encoding
        loss = criterion(outputs.cpu(), y)
        loss.backward()

As shown above, I need to loop multiple times in the forward stage, but I want to update parameters only once in the Backward process, not all forward loops need to be updated.
What should I do, please.Thank you！

2

Hello , You can just remove the second loop. Now, the Neural Network will take all the data instead of a batched data. batch_size will now be the size of the whole dataset. Hope it helps