The linear layer's weight dosen't update during training, am I wrong?

Hi there,
I came across a problem.I try to train a net but only some of my parameters are updated and I don’t know why.
Here is my code.

class Model(nn.Module):
    def __init__(self,input_word_size,embedding_size,hidden_size,padding_index,embedding = None,use_cuda = False):
        self.input_word_size = input_word_size
        self.embedding_size = embedding_size
        self.hidden_size = hidden_size
        self.use_cuda = use_cuda

        self.embedding = nn.Embedding(input_word_size,embedding_size,padding_idx=padding_index)
        if embedding is not None:
            self.embedding.weight = embedding
        self.context_gru = nn.GRU(embedding_size,hidden_size,batch_first = True)
        self.response_gru = nn.GRU(embedding_size,hidden_size,batch_first = True)
        self.w = nn.Linear(hidden_size,hidden_size)
    def forward(self,x,y):
        x_emb = self.embedding(x)# x_emb:(batch_size,len,embedding_size)
        y_emb = self.embedding(y)# y_emb:(batch_size,len,embedding_size)
        batch_size = x_emb.shape[0]

        x_hidden = torch.randn(1,x.shape[0],self.hidden_size)
        y_hidden = torch.randn(1,x.shape[0],self.hidden_size)
        if self.use_cuda:
            x_hidden = x_hidden.cuda()
            y_hidden = y_hidden.cuda()
        x_hiddens = []
        y_hiddens = []
        x_output,x_hidden = self.context_gru(x_emb,x_hidden)
        y_output,y_hidden = self.response_gru(y_emb,y_hidden)

        context = x_hidden.transpose(0,1)
        response = y_hidden.transpose(0,1).transpose(1,2)

        # the operation of linear layer is below
        result = self.w(context)
        result = torch.bmm(result,response)
        result = F.sigmoid(result)
        return result

and after I do the operation below:

a = model.w.weight.clone()
result = model(x,y,label)
loss = loss_function(result,label) # loss function is BCELoss
b = model.w.weight.clone()
print(torch.equal(a,b)) # I got True here, which means the weight has not changed
# and the value of model.w.weight.grad is not None

What should I do to fix this?Should I change my network?

Try using optimizer.zero_grad() before loss.backward(). (BTW, I am assuming you are returning result in forward and not result3)

thanks to your reply but sorry,there is some typing mistakes.I do return result rather result3 and I also zero grad before backward.but it still doesn’t work