Problems training a model once, the interpreter says that a tensor required for the backpropagation was modified but i dont see where

[KFrank]

Hi Ranon!

@ranon-rat

in the first iteration there is no problem but in the second one it gives me the next error:

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.FloatTensor [448, 112]] is at version 2; expected version 1 instead. Hint: the backtrace further above shows the operation that failed to compute its gradient. The variable in question was changed in there or anywhere later. Good luck!

…

train_once(self,indicator,winned,hidden,action,reward,next_indicator,next_winned,done):
    ...
    pred, hidden = self.model(indicator, winned, hidden)
    ...
    loss.backward()
    self.optimizer.step()

…

    output,new_hidden=self.lstm(output,hidden)

I am going to speculate as follows:

The first time through, hidden doesn’t depend on the parameters of lstm.
However, the second time through the (old) hidden depends on the parameters
of lstm before they were updated by optimizer.step().

In general, optimizer.step() modifies inplace the parameters of the model
being optimized.

When you call .backward() the second time, loss, through target, depends
on the old hidden, which depends on the old lstm parameters, so you get
the inplace-modification error.

Would it make sense for your use case (assuming that this is the cause of your
error) to .detach() the old hidden from the computation graph? You would
still pass the old hidden into model so the new pred and hidden would
depend on the old hidden’s values, but those values would be .detach()ed
from their dependency on the old values of lstm’s parameters.

Aside from this speculation, please see the suggestions for locating and fixing
inplace-modification errors given in the following 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