One of the variables needed for gradient computation has been modified by an inplace operation (backward pass is performed twice)

[KFrank]

Hi Srushti!

@Srushti_Patil

    with torch.autograd.set_detect_anomaly(True): 
        for batch_idx, (data, target) in enumerate(train_loader):

            output = model4(data) 
            loss = loss_fun(output, target)
            
            optimizer.zero_grad()
            loss.backward(retain_graph=True)
            optimizer.step()
        
            total_loss.append(loss.item()) 
        def closure():
            optimizer_qnn.zero_grad()
            loss = loss_fun(output, target)  
            loss.backward()
            print(loss.item()) 
            return loss 
        optimizer_qnn.step(closure)   

Your immediate problem is likely here. You are calling loss.backward()
with retain_graph = True. First, you should think carefully about whether
you need retain_graph = True, and if so, why.

In the last iteration of your enumerate (train_loader) loop, you build a
computation graph that connects output to the parameters of model4. This
graph is preserved. In general, optimizer.step() modifies the parameters
of your model inplace. When closure() is then executed, it calls backward()
on loss_fun (output, target), which will backpropagate through the graph
that connects output to the parameters of model4. But those parameters
have been modified inplace, causing the error.

  File "C:\Users\Srushti\AppData\Local\Temp\ipykernel_68744\1791620977.py", line 24, in forward
    x = self.fc2(x).clone()
  File "C:\Users\Srushti\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.8_qbz5n2kfra8p0\LocalCache\local-packages\Python38\site-packages\torch\nn\modules\module.py", line 1501, in _call_impl
    return forward_call(*args, **kwargs)
  File "C:\Users\Srushti\AppData\Local\Packages\PythonSoftwareFoundation.Python.3.8_qbz5n2kfra8p0\LocalCache\local-packages\Python38\site-packages\torch\nn\modules\linear.py", line 114, in forward
    return F.linear(input, self.weight, self.bias)
 (Triggered internally at C:\actions-runner\_work\pytorch\pytorch\builder\windows\pytorch\torch\csrc\autograd\python_anomaly_mode.cpp:119.)

The forward-call traceback generated by set_detect_anomaly (True) is
complaining about the call to fc2 (x) in your model4. This agrees with the
above analysis in that that fc2.weight has been modified inplace by the call
to optimizer.step().

To fix this, you will need to think through the logic of your use case. Does
optimizer_qnn.step (closure) need gradients of output with respect
to the “regular” parameters of model4, such as fc2? If so, would it be practical
to rebuild the model4–output graph after calling optimizer.step(), perhaps
inside of closure()?

For some examples that show how to debug and fix inplace-modification errors,
see this post:

"RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor [64, 1]], which is output 0 of AsStridedBackward0, is at version 3; expected version 2 instead. Hint: the backtrace further a

Hi Fahmyadan and Sangyoon! Here are some suggestions about how to track down (and maybe fix) inplace-modification errors. Note that an inplace modification in the forward pass is not necessarily* an error – it depends on whether and how the tensor that was modified is used in the backward pass. Note that inplace operations can be useful for saving memory – if you replace an innocent inplace operation with an out-of-place equivalent, your training will use more memory (and, to a minor e…

Good luck!

K. Frank