Parameter target in CaptumExplainer

[CYQFWang]

I find following codes in torch_geometric\explain\algorithm\captum_explainer.py
Line 157

        if self.model_config.mode == ModelMode.regression:
            target = None
        else:
            target = target[index]

        attributions = self.attribution_method.attribute(
            inputs=inputs,
            target=target,
            additional_forward_args=add_forward_args,
            **self.kwargs,
        )

Since target are "Output indices (for classification cases, this is usually the target class)", I think target ought to be None when self.model_config.mode ==ModelMode.binary_classification?

[rusty1s]

@RBendias Do you know why we set target to None here?
@CYQFWang Do you mean regression rather than binary classification?

[CYQFWang]

@RBendias I think Captum wants a scalar as output for IntegratedGradients and this is why we are asked to provide output indices when our model actually have multi-dimensional output.

        assert outputs[0].numel() == 1, (
            "Target not provided when necessary, cannot"
            " take gradient with respect to multiple outputs."
        )

(https://github.com/pytorch/captum/blob/master/captum/_utils/gradient.py#L113)
Therefore, the target parameter should only depends on the shape of our model's output decided in Line151 as

            captum_model = CaptumModel(model, mask_type, index)

What we should do probaly is make sure that this model's output is a scalar and provide target if the model gives us a n-dimensional vector.

[RBendias]

According to my understanding of the code, the model will only return a scalar in case of regression (here, we do not provide the target). Due to the transformation in CaptumModel, we will always have a multi-dimensional output for classification (both binary and multi-class), here, we provide target.

[CYQFWang]

@RBendias I'm afraid attribution_method in captum.attr doesn't want any multi-dimensional output model base on assert outputs[0].numel() == 1?

[RBendias]

If there is multi-dimensional output captum uses the target to transform it to a scalar here.

[CYQFWang]

Yes, I think it works after the converting you mentioned and the target in attribution_method should be distinguished from the target provided by users in Explainer.