Explain text predictions of Keras classifiers

[teabolt]

Based on #315 and #329, we use Grad-CAM to highlight parts of text that contribute to a prediction of a Keras classifier.

For example, we can use a call similar to this:

eli5.show_prediction(model, doc, 
        tokens=doc_tokens,
        pad_x=word_index['<PAD>'],
        padding='post',
        relu=False,
)

To roughly highlight positive and negative parts of a text (explaining the score given in sentiment analysis):

This PR also makes refactorings (including changes to the public API) to the code at the linked PR's.

WIP items:

• Tutorial.
• Pass CI and coverage.
• Tests (manual and automated).
• Docs.
• Mypy.

[teabolt]

@lopuhin Do you think there is anything major left to do in this PR? (besides making the CI pass)

[lopuhin]

@teabolt I don't recall anything significant, and from a quick glance it looks like all review feedback is addressed so I think it's almost ready.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

❌ Patch coverage is 98.26990% with 5 lines in your changes missing coverage. Please review.
✅ Project coverage is 97.34%. Comparing base (017c738) to head (347ed35).
⚠️ Report is 15 commits behind head on master.

Files with missing lines	Patch %	Lines
eli5/nn/text.py	96.00%	2 Missing and 2 partials ⚠️
eli5/formatters/image.py	95.83%	0 Missing and 1 partial ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #325      +/-   ##
==========================================
+ Coverage   97.32%   97.34%   +0.02%     
==========================================
  Files          49       52       +3     
  Lines        3142     3320     +178     
  Branches      585      623      +38     
==========================================
+ Hits         3058     3232     +174     
- Misses         44       46       +2     
- Partials       40       42       +2     

Files with missing lines	Coverage Δ
eli5/base.py	100.00% <100.00%> (ø)
eli5/ipython.py	100.00% <ø> (ø)
eli5/keras/__init__.py	100.00% <100.00%> (ø)
eli5/keras/explain_prediction.py	98.37% <100.00%> (+1.98%)	⬆️
eli5/keras/gradcam.py	100.00% <100.00%> (ø)
eli5/nn/__init__.py	100.00% <100.00%> (ø)
eli5/nn/gradcam.py	100.00% <100.00%> (ø)
eli5/formatters/image.py	98.71% <95.83%> (-1.29%)	⬇️
eli5/nn/text.py	96.00% <96.00%> (ø)

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[lopuhin]

thanks @teabolt I hope I can finish review in a week, looks really solid, left a few minor comments

[Hellisotherpeople]

Is this ever going to be merged? I'm super interested in seeing this get added to ELI5 ASAP!!!

[teabolt]

Is this ever going to be merged? I'm super interested in seeing this get added to ELI5 ASAP!!!

Hey @Hellisotherpeople . Sorry for the late reply. If you are still interested you can try installing the PR branch directly with pip:

1. If you are using virtualenv or a similar environment manager, make a clone of your environment.
2. Install the PR branch with pip install -U git+https://github.com/teabolt/eli5.git@keras-gradcam-text

To test that it worked run:

import eli5.keras
eli5.keras.explain_prediction_keras_text()

The output should be TypeError: explain_prediction_keras_text() missing 2 required positional arguments: 'model' and 'doc'

Tested this with Python 3.7 and Ubuntu 20.04.