Issue with calculating gradients during back propagation

[RamakrishnaChaitanya]

Hi, I'm encountering the following error when calling loss.backward() in the train function:

  File "/file1/Ramakrishna/Projects/VAD/Github_repos/silero-vad/rk_stuff/utils.py", line 212, in train
    loss.backward()

  File "/file1/Ramakrishna/envs/silero_vad_env/lib/python3.10/site-packages/torch/_tensor.py", line 648, in backward
    torch.autograd.backward(

  File "/file1/Ramakrishna/envs/silero_vad_env/lib/python3.10/site-packages/torch/autograd/__init__.py", line 353, in backward
    _engine_run_backward(

  File "/file1/Ramakrishna/envs/silero_vad_env/lib/python3.10/site-packages/torch/autograd/graph.py", line 824, in _engine_run_backward
    return Variable._execution_engine.run_backward(# Calls into the C++ engine to run the backward pass)

RuntimeError: Trying to backward through the graph a second time (or directly access saved tensors after they have already been freed). Saved intermediate values of the graph are freed when you call .backward() or autograd.grad(). Specify retain_graph=True if you need to backward through the graph a second time or if you need to access saved tensors after calling backward.

---

It seems like the issue might be related to missing optimizer.zero_grad() before the backward pass. Could someone please help me understand the root cause and suggest how to resolve this error?

I'm using the decoder like this

class Decoder(nn.Module):
    def __init__(self):
        super(Decoder, self).__init__()
        self.rnn = nn.LSTMCell(32, 32)
        self.decoder = nn.Sequential(nn.Dropout(0.1),
                                     nn.ReLU(),
                                     nn.Conv1d(32, 1, kernel_size=1, padding=0, dilation=1),
                                     nn.Sigmoid())

    def forward(self, x, state=torch.zeros(0)):
        x = x.squeeze(-1)
        if len(state):
            h, c = self.rnn(x, (state[0], state[1]))
        else:
            h, c = self.rnn(x)

        x = h.unsqueeze(-1).float()
        state = torch.stack([h, c]
        x = self.decoder(x)
        return x, state

However, after replacing

state = torch.stack([h,c])_ with _state = torch.stack([h.detach(), c.detach()]) resolved the issue. But I'm not sure whether this breaks anything!