PyTorch Recipes > Model Interpretability using Captum throws exception

[SebastianRiechert]

📚 Documentation

This jupyter-notebook-example-code from the docs:

https://pytorch.org/tutorials/recipes/recipes/Captum_Recipe.html

throws the following error in the 'Visualizing the Results'-Cell:

---------------------------------------------------------------------------
AssertionError                            Traceback (most recent call last)
<ipython-input-4-43fe1d15438b> in <module>
     15                                       vis_signs,
     16                                       ["attribution for dog", "image"],
---> 17                                       show_colorbar = True
     18                                      )
     19 

~\Anaconda3\lib\site-packages\captum\attr\_utils\visualization.py in visualize_image_attr_multiple(attr, original_image, methods, signs, titles, fig_size, use_pyplot, **kwargs)
    387             >>>                     ["all", "positive"], attribution, orig_image)
    388     """
--> 389     assert len(methods) == len(signs), "Methods and signs array lengths must match."
    390     if titles is not None:
    391         assert len(methods) == len(titles), (

AssertionError: Methods and signs array lengths must match.

This is due to the trailing comma in vis_sings:

import numpy as np
from captum.attr import visualization as viz

# Convert the compute attribution tensor into an image-like numpy array
attribution_dog = np.transpose(attribution_dog.squeeze().cpu().detach().numpy(), (1,2,0))

vis_types = ["heat_map", "original_image"]
vis_signs = ["all", "all"], # "positive", "negative", or "all" to show both
# positive attribution indicates that the presence of the area increases the prediction score
# negative attribution indicates distractor areas whose absence increases the score

_ = viz.visualize_image_attr_multiple(attribution_dog,
                                      center_crop(img),
                                      vis_types,
                                      vis_signs,
                                      ["attribution for dog", "image"],
                                      show_colorbar = True
                                     )


attribution_cat = np.transpose(attribution_cat.squeeze().cpu().detach().numpy(), (1,2,0))

_ = viz.visualize_image_attr_multiple(attribution_cat,
                                      center_crop(img),
                                      ["heat_map", "original_image"],
                                      ["all", "all"], # positive/negative attribution or all
                                      ["attribution for cat", "image"],
                                      show_colorbar = True
                                     )

Relevant packages in my env:

torch==1.6.0
torchvision==0.7.0
captum==0.2.0

Python 3.7.7

[bilalsal]

Thank you for reporting this issue @SebastianRiechert!
We have created a PR in the PyTorch recipes branch to fix the issue.
I will close the issue once the PR is merged and the notebook is updated.

[NarineK]

@bilalsal, it looks like pytorch/tutorials#1271 got approved. Can we merge it ?

[bilalsal]

@bilalsal, it looks like #553 got approved. Can we merge it ?

PR merged. Captum_Recipe.html and the associated Jupyter notebook should be updated as soon as the tutorials are recompiled from pytorch/tutorials/master

[bilalsal]

Tutorial is recompiled https://pytorch.org/tutorials/recipes/recipes/Captum_Recipe.html