BUG: Backward hook with shap.DeepExplainer on simple models with PyTorch about shap HOT 4 OPEN

Puuhh, this seems pretty tough to fix. I tried to set every ReLU activation to inplace=False but that still fails at the line

out += identity

of the resnet module with the same error. Strangely this works if I replace it with

out = out + identity

but then the shapes mismatch.

from shap.

Thanks for the reproducible example. I can confirm this on master.

I will have a look at this in the upcoming weeks.

from shap.

I am attempting to apply DeepExplainer for resnet50 and densenet121 and I am getting BackwardHookFunctionBackward error msg. I tried to modify the ReLU activation to disable the inplace operation by setting it to False. I am using the following approach to update the inplace:

class ModifiedDenseNet121(nn.Module):
    def __init__(self, pretrained=True, input_channels=3, num_classes=1000):
        super(ModifiedDenseNet121, self).__init__()
        self.densenet = models.densenet121(weights=DenseNet121_Weights.DEFAULT if pretrained else None)
        if input_channels != 3:
            self.densenet.features[0] = nn.Conv2d(input_channels, 64, kernel_size=7, stride=2, padding=3, bias=False)
        self.densenet.classifier = nn.Linear(self.densenet.classifier.in_features, num_classes)
        # Fix the usage of ReLU activation with inplace=True
        for module in self.densenet.modules():
            if isinstance(module, nn.ReLU):
                module.inplace = False
    def forward(self, x):
        return self.densenet(x)

However, I'm still encountering a runtime error: RuntimeError: Output 0 of BackwardHookFunctionBackward is a view and is being modified inplace. Any guidance or updates on how to resolve this error would be greatly appreciated. Thank you!

from shap.

If you really need this maybe have a look at captum. Would be great if you could report back your findings, so that we can borrow a couple of ideas from them

from shap.