Metrics error due to inplace operation, "computation has been modified by an inplace operation".

[sykrn]

Hey, @williamFalcon, I got a new error since I upgraded the library today.
I used the accuracy metric, but got an error.

Code sample:

# in lightning module
def training_step(self, batch, batch_idx):
        x, y = batch
        y_hat = self(x)
        loss = F.cross_entropy(y_hat, y)
        acc = accuracy(y_hat, y)      # from the functional metric classification
        tensorboard_logs = {'train_loss': loss}
        return {'loss': loss, 'log': tensorboard_logs}

Error msg:

RuntimeError: one of the variables needed for gradient computation has been modified by an inplace operation: [torch.cuda.LongTensor [32]] is at version 2; expected version 1 instead. Hint: enable anomaly detection to find the operation that failed to compute its gradient, with torch.autograd.set_detect_anomaly(True).

Can be solved using .clone() method.

However, when I clone the y before feeding to the accuracy function, no error was shown.

acc = accuracy(y_hat, y.clone())

But, it's inconvenience if user has to do it manually, isn't it?
Actually, I can use the code above without clone before I upgrade to the latest. So, it might due to the latest update/rebase causing this error.

The same error shown for f1_score metric.

[github-actions]

Hi! thanks for your contribution!, great first issue!

[williamFalcon]

cc @justusschock

[justusschock]

cc @Diuven , I think you introduced in-place ops for speed-up, right?
Does the speedup only come from inplace methods or can we simply replace them by out-of-place methods ?

[Diuven]

cc @Diuven , I think you introduced in-place ops for speed-up, right?
Does the speedup only come from inplace methods or can we simply replace them by out-of-place methods ?

Yeah, you're right. I think this is because of the clamp_max_ I used in stat_scores_mulitple_classes. This is the part.

This is a quick PR fixing this issue. If you may, please check the code still gives the issue.

Sorry for the inconvenience!

[sykrn]

Great, It runs without any error, now. Thanks for the hotfix. 👍