how to calculate a gradient on a specific weight manually in pytorch

Question

# a part of my code, it works independently
import torch
class wmodel(torch.nn.Module):
    def __init__(self):
        super().__init__()
        self.model = torch.nn.Linear(8, 10)
        self.lossfunc = torch.nn.BCELoss()
        self.optimizer = optimizer(self.parameters(), lr=0.2)

rrr = wmodel()
datai = torch.tensor([1.,2,3,4,5,6,7,8])
target = torch.tensor([1.,2,3,4,5,6,7,8,9,10])
qan = rrr.model(datai)
loss = target-qan

I'm trying to calculate the loss[0]'s partial derivative to rrr.model.weight[0][0], one of the problem I'm facing is rrr.model.weight[0][0] does not participate in the calculation, which means I can't call grad on it, I'm searching for a variable that actually participated in the process

Answer 1

It may be helpful in pinpointing the exact problem if you post a minimal reproducible example. However, I suspect that it's possible that you're not calling your_model.backwards() after calculating your loss. The gradients won't propagate until you call backwards(), this is to account for things like RNNs or taking the average gradient over multiple batches. Each time you call backwards(), it adds (as in arithmetic addition) the new gradient to the gradient currently propagated, until you call your_model.zero_grad(), which sets your gradient back to zero.