Skip to content

Latest commit

 

History

History
99 lines (66 loc) · 4.63 KB

README.md

File metadata and controls

99 lines (66 loc) · 4.63 KB
Raw

gsoc19-eli5-gradcam

Google Summer of Code 2019 Work Product Submission for the ELI5 Grad-CAM project under Scrapy, Python Software Foundation.

Hi! Welcome to the final report for the project titled “Grad-CAM Neural Network Explanations for ELI5” (https://summerofcode.withgoogle.com/archive/2019/projects/5066973807902720/).

Summary

All of my work was done as additions to the python library “ELI5” https://github.com/TeamHG-Memex/eli5.

I have worked on adding support to explain image and text classifier predictions in Keras and PyTorch. I have implemented the Grad-CAM technique (https://arxiv.org/abs/1610.02391) for explanations. This resulted in mypy annotated and commented Python source code (eli5.keras, eli5.nn, eli5.pytorch, eli5.formatters.image), pytest unit and integration tests, and sphinx docstrings, library descriptions and tutorials.

I have decided to exclude adding support for Tensorflow and non-classifier networks in this project.

Example of an explanation for why a network classified this image as “dog”

gradcam cat dog image heatmap overlay

Example explanation for why a certain score was given by a sentiment analysis network

gradcam text heatmap over tokens

What is merged

Explaining Keras image classifiers

In this piece of work we explain the predictions done by a Keras classifier that takes images as its input.

Covered by two merged pull requests

See https://eli5.readthedocs.io/en/latest/libraries/keras.html and https://eli5.readthedocs.io/en/latest/tutorials/keras-image-classifiers.html for usage. The user can now call eli5.explain_prediction() top level function on a Keras image classifier.

What is left to do

Explaining Keras text models

We explain the predictions of a Keras classifier that takes text as its input.

Covered by this open pull request: TeamHG-Memex/eli5#325

Left to do

  • Complete some TODO items in source comments.
  • Reviewer feedback.

Explaining PyTorch image and text models

We explain classifiers defined in PyTorch that take either images or text as its input. This is a scaled down version of support we would have for Keras.

Covered by this open pull request: TeamHG-Memex/eli5#327

Left to do

  • Unit and integration tests.
  • Text tutorial.
  • PyTorch library description.
  • Docstrings.
  • Reviewer feedback.

What is working (based on manual testing)

  • Reasonable explanations for image and text classifier predictions.
  • Most of Keras features are supported (targets, layer, relu, counterfactual, and various text arguments).

Other (according to the original project plan)

These are optional. Either I can do them myself or let the users/maintainers do it.

Delivering a release

  • Open a PR to help maintainers make a new ELI5 release (bump version, etc).

Tensorflow

  • Implement support for Keras models defined through Tensorflow (tf.keras).

Non-classifier models

  • Support models for regression, Visual Question Answering, etc.
  • Current code attempts to make this an easy thing to add support for.
  • Regression can be converted into classification.

Challenges

  • Usual coding stuff - working with Keras, numpy, Pillow, PyTorch.
  • Later down the line: making changes that “break” a lot of things, i.e. need to update other code, tests, docs.
  • Prioritizing what to work on. Working up to speed to get "features" done.
  • Estimating how long certain tasks will take.
  • Finding hardware suitable for Machine Learning, i.e. using Kaggle kernels.

Learnings

  • Grad-CAM as a gradient method for explanations.
  • Keras training and inference.
  • PyTorch training and inference.
  • Text and image preprocessing.
  • Numpy, Pillow.
  • Sphinx, reStructuredText.
  • Mypy, pytest, tox, Travis CI, codecov.
  • Git, virtualenv, pip.
  • GitHub, Kaggle, Jupyter Notebook.

Thank you to my mentors Konstatin and Mikhail. I hope this work proves useful!