I’m using the overruling dataset is a small (n=2400, evenly split) binary classification task. I’m also using a tiny BERT model (4.4M params).

I’m also assuming (following this FAQ that I don’t need to explicitly provide a target.

My call to attribute() looks like this:

attributions_ig, delta = lig.attribute(in_tensor, reference_indices, 
                    additional_forward_args (ttype_tensor,attn_tensor), 
                                        n_steps=500, return_convergence_delta=True)

• It makes it to LayerIntegratedGradients.gradient_func L#480 and then
  fails on this assertion:

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

• Indeed, output[0].numel() = 1000; output[0].shape = torch.Size([500, 2])

• Looking back up the stack, _attribute() is getting the right
  inputs argument and inputs[0].shape is [1,128,128] but the
  scaled_features_tpl passed on as inputs to gradient_func as
  inputs has shape [500, 128,128] ?! I also notice that the inputs
  have ALSO been prepended (redundantly?) onto additional_forward_arguments?

(Also posted as captum issue )