Bug in LimeBase example?

[th789]

🐛 Bug

I am trying to run the example provided for LimeBase on https://captum.ai/api/lime.html. However, running the example leads to the following error message: "TypeError: () got an unexpected keyword argument 'kernel_width'" (more info below).

To Reproduce

import torch
import torch.nn as nn
from captum.attr import LimeBase
from captum._utils.models.linear_model import SkLearnLinearModel

class SimpleClassifier(nn.Module):
    def __init__(self):
        super().__init__()
        self.linear = nn.Linear(5, 3)
        self.sigmoid = nn.Sigmoid()
    def forward(self, x):
        out = self.linear(x)
        out = self.sigmoid(out)
        return out

###all code below copied from example provided
net = SimpleClassifier()

def similarity_kernel(original_input, perturbed_input, perturbed_interpretable_input, **kwargs):
      # kernel_width will be provided to attribute as a kwarg
      kernel_width = kwargs["kernel_width"]
      l2_dist = torch.norm(original_input - perturbed_input)
      return torch.exp(- (l2_dist**2) / (kernel_width**2))

def perturb_func(original_input, **kwargs):
      return original_input + torch.randn_like(original_input)

input = torch.randn(2, 5)

lime_attr = LimeBase(net,
                     SkLearnLinearModel("linear_model.Ridge"),
                     similarity_func=similarity_kernel,
                     perturb_func=perturb_func,
                     perturb_interpretable_space=False,
                     from_interp_rep_transform=None,
                     to_interp_rep_transform=lambda x: x)

attr_coefs = lime_attr.attribute(input, target=1, kernel_width=1.1) #error message appears after running this line

Error message:

Expected behavior

There should be no error message and attr_coefs should return the feature attributions.

Environment

Describe the environment used for Captum

 - Captum version: 0.5.0
 - Pytorch version: 1.10.0+cu111
 - OS (e.g., Linux): macOS
 - How you installed Captum (`conda`, `pip`, source): 'conda' and 'pip' --> this error message arises whether I use `conda install captum -c pytorch` or `pip install captum` to install captum
 - Python version: 3.7.12

[vivekmig]

Hi @th789 , thanks for catching this! It seems that the lambda function provided as to_interp_rep_transform should be changed to accept kwargs. We will update the documentation to correct this.

[th789]

Hi @vivekmig, thank you so much for the reply! I had tried that, but it led to another error message (see below). Perhaps I misunderstood what you meant?

#(keeping all lines above the same)
def to_interp_rep_transform_none(x, **kwargs):
    return x

lime_attr = LimeBase(net,
                     SkLearnLinearModel("linear_model.Ridge"),
                     similarity_func=similarity_kernel,
                     perturb_func=perturb_func,
                     perturb_interpretable_space=False,
                     from_interp_rep_transform=None,
                     to_interp_rep_transform=to_interp_rep_transform_none)

attr_coefs = lime_attr.attribute(input, target=1, kernel_width=1.1)

[vivekmig]

Hi @th789 , the signature also needs to include a positional argument for original inputs, based on the expected function signature defined here. Seems like this wasn't updated after a change in the function signature, sorry about that! You can try out the updated sample in #908 .

In addition to the function, you would also need to change to a batch size of 1, since LimeBase supports attribution for one example at a time, unless the forward function returns a batch-wise output or loss.

[th789]

Thank you very much @vivekmig, I really appreciate it!