Lime tutorial

[aobo-y]

Add a tutorial to demonstrate Lime. Made of 2 section:

• one for image classification. Use high-level Lime class
• one for text classification. Manually craft almost every customizable argument with LimeBade class

Github does not render the last text visualization part made of Ipython Html, which should look like below in actual notebook

[facebook-github-bot]

@aobo-y has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[vivekmig]

Looks great, thanks for the awesome work on this :) ! I really like that it covers 2 different use-cases and using both the generic parent class and specific child class.

A few small nits and minor wording suggestions:

• When explaining this line "Lime creates the pertubed samples in the form of interpretable representation, i.e., a binary vector indicating the “presence” or “absence” of features.", would it make sense to make it more concrete with examples of vector length 3 and visualizing the corresponding image? While this is not necessary for usage of Lime, it may be helpful for users to better understand what occurs internally
• Do we want to save the parameters trained for the second model to avoid a user needing to train? How long did the training take?
• For the text visualization issue, the BERT tutorial also includes a png of the results (https://captum.ai/tutorials/Bert_SQUAD_Interpret), could consider that here as well, particularly if the visualization issue still occurs if downloading the notebook or through the website.

Wording notes:

• nit: What does POV stand for in "POV's segmentation masks"? Is this supposed to be VOC?
• nit: "Everytime" -> "Every time"
• nit: "check the document for details" -> "check the documentation for details"
• nit: "machine still has capaciity" -> "capacity". Also can consider rewording this line a bit, since parallelize may be confusing to users, can instead say something like batching multiple samples in one forward pass evaluation, so readers understand this relates to max batch size they may typically use
• nit: return_input_shape defaults to True for consistency with other methods, this doesn't need to be passed in cell 12
• nit: "Lasso regularization can effectively help us filter them" -> can we add something like "by adding regularization to the interpretable model"
• nit: "build-in sklearn" -> "built-in sklearn"
• "pertubed binary interpretable" -> "perturbed binary interpretable"
• "if the word of each position presents or not" -> "if the word of each position is present or not"
• "Setting them to any baselines will surely pollute the calculation and end with imperfect result." -> Can we reword this a little to say that in this example we would like to remove characters rather than replacing with a baseline token? In this example, it makes sense why removing would be preferable considering the embedding bag structure, but I'm concerned this may lead users to think that replacement should not be used for text generally, but in general, just removing arbitrary tokens is also generating out of distribution inputs which may not always be better than replacement.
• "forward_func, the target to attribute" -> target to attribute shouldn't be here?

[facebook-github-bot]

@aobo-y has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[aobo-y]

thanks for the feedback, @vivekmig. I have updated all the wording as suggested.

Just want to further bring up the following items:

"Setting them to any baselines will surely pollute the calculation and end with imperfect result." -> Can we reword this a little to say that in this example we would like to remove characters rather than replacing with a baseline token? In this example, it makes sense why removing would be preferable considering the embedding bag structure, but I'm concerned this may lead users to think that replacement should not be used for text generally, but in general, just removing arbitrary tokens is also generating out of distribution inputs which may not always be better than replacement.

Yes, I agree. I believe setting to baseline can also be used in text. So I tried to soften my tone for our choice in the tutorial. Let me know if you have further suggestions.

When explaining this line "Lime creates the pertubed samples in the form of interpretable representation, i.e., a binary vector indicating the “presence” or “absence” of features.", would it make sense to make it more concrete with examples of vector length 3 and visualizing the corresponding image? While this is not necessary for usage of Lime, it may be helpful for users to better understand what occurs internally

This definitely helps users to understand Lime, but I cannot find any not awkward ways to inject the illustration into the quoted point, because it will interrupt the explanations of how to call attribute which is the most important goal for this tutorial.

So I plan to append a dedicated section for that, "1.4 Understand the sampling process", where I can show what a perturbed interpretable sample looks like, what the converted image looks like, what the model's prediction of it, and the similarity is calculated in feature space. I will ship the current tutorial and make another PR for the new section once I finish so it will be easier to review.

Do we want to save the parameters trained for the second model to avoid a user needing to train? How long did the training take?

As embedding-bag is both shallow & non-sequential, the training is finished around 6 mins on my CPU machine. I feel it is acceptable but I will add a checkpoint in the next PR.

For the text visualization issue, the BERT tutorial also includes a png of the results (https://captum.ai/tutorials/Bert_SQUAD_Interpret), could consider that here as well, particularly if the visualization issue still occurs if downloading the notebook or through the website.

I was curious why those tutorials have duplicate outputs... I believe the visualization issue is just Github's escape of HTML scripts on the web. All the HTML tables in previous tutorials are also collapsed. I have also included one duplicate screenshot to beautify its Github view.

[facebook-github-bot]

@aobo-y has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[facebook-github-bot]

@aobo-y merged this pull request in cf9bf08.