Captum: AssertionError: Cannot choose target column with output shape torch.Size([100])

[mainulquraishi]

I don't understand the error. Is not the output means the model's output. If so, my predictor's output shape is (batch_size,1).
Here is the complete error trace,

----> 2 interpret_sentence(model, doc)
      3 # for tv_batch in source_test_dataloader:
      4 #     xtv_batch=tv_batch[0]
      5 #     ytv_batch=tv_batch[1]

<ipython-input-46-789ee45a76d0> in interpret_sentence(model, doc)
     42     reference_indices=reference_indices.to(device)
     43     print(reference_indices.shape)
---> 44     attributions_ig, delta = lig.attribute(b_input_ids, reference_indices,target=2 , \
     45                                            n_steps=100, return_convergence_delta=True)
     46     print(attributions_ig.shape)

/opt/conda/lib/python3.8/site-packages/captum/attr/_core/layer/layer_integrated_gradients.py in attribute(self, inputs, baselines, target, additional_forward_args, n_steps, method, internal_batch_size, return_convergence_delta, attribute_to_layer_input)
    350             else inps
    351         )
--> 352         attributions = self.ig.attribute(
    353             inputs_layer,
    354             baselines=baselines_layer,

/opt/conda/lib/python3.8/site-packages/captum/attr/_core/integrated_gradients.py in attribute(self, inputs, baselines, target, additional_forward_args, n_steps, method, internal_batch_size, return_convergence_delta)
    276 
    277         # grads: dim -> (bsz * #steps x inputs[0].shape[1:], ...)
--> 278         grads = _batched_operator(
    279             self.gradient_func,
    280             scaled_features_tpl,

/opt/conda/lib/python3.8/site-packages/captum/attr/_utils/batching.py in _batched_operator(operator, inputs, additional_forward_args, target_ind, internal_batch_size, **kwargs)
    154     of the results of each batch.
    155     """
--> 156     all_outputs = [
    157         operator(
    158             inputs=input,

/opt/conda/lib/python3.8/site-packages/captum/attr/_utils/batching.py in <listcomp>(.0)
    155     """
    156     all_outputs = [
--> 157         operator(
    158             inputs=input,
    159             additional_forward_args=additional,

/opt/conda/lib/python3.8/site-packages/captum/attr/_core/layer/layer_integrated_gradients.py in gradient_func(forward_fn, inputs, target_ind, additional_forward_args)
    331                     hook = self.layer.register_forward_hook(layer_forward_hook)
    332 
--> 333                 output = _run_forward(
    334                     self.forward_func, tuple(), target_ind, additional_forward_args
    335                 )

/opt/conda/lib/python3.8/site-packages/captum/attr/_utils/common.py in _run_forward(forward_func, inputs, target, additional_forward_args)
    503         else inputs
    504     )
--> 505     return _select_targets(output, target)
    506 
    507 

/opt/conda/lib/python3.8/site-packages/captum/attr/_utils/common.py in _select_targets(output, target)
    452     dims = len(output.shape)
    453     if isinstance(target, (int, tuple)):
--> 454         return _verify_select_column(output, target)
    455     elif isinstance(target, torch.Tensor):
    456         if torch.numel(target) == 1 and isinstance(target.item(), int):

/opt/conda/lib/python3.8/site-packages/captum/attr/_utils/common.py in _verify_select_column(output, target)
    439 ) -> Tensor:
    440     target = cast(Tuple[int, ...], (target,) if isinstance(target, int) else target)
--> 441     assert (
    442         len(target) <= len(output.shape) - 1
    443     ), "Cannot choose target column with output shape %r." % (output.shape,)

AssertionError: Cannot choose target column with output shape torch.Size([100]). 

Edit: I just noticed this 100 is coming from "n_step". If change n_step=200, the error shows 200. Other than this, I have no idea why this error happens and how to solve this. Can anybody help?

[NarineK]

@mainulquraishi , the output of the model should be 2 dimensional for an integer-valued target. Currently, you have one dimensional output. We always assume that the first dimension is the number of examples. You want to make sure that the forward function returns an output in a shape [#example x #classes]

[mainulquraishi]

I am doing regression. For regression, should not be the output shape [#example] ?

[vivekmig]

Hi @mainulquraishi , the output shape is fine, but if the output shape is [#examples], you shouldn't provide the target argument when calling LayerIntegratedGradients. This argument should only be provided when there are multiple outputs per example and one needs to be selected, such as in a classification case. More information about this can be found in the documentation for the target parameter here. Hope this helps!

[NarineK]

@mainulquraishi, do you still see the issue ?