This is a personal project to re-implement part of Lucid in TensorFlow 2. Lucid is a package for research in neural network interpretability, and is built on TensorFlow 1.
Specifically, this repository covers the content of this article on feature visualization. The key idea is to visualize features learned in a component of a model by maximizing its activation in a generated image.
This approach of feature visualization consists of the following ingredients:
- Objective function. A function to be maximized in an image, representing what feature to visualize.
- Image parametrization. A choice of how to represent the degrees of freedom in an image, which can impact the optimization result.
- Transformations. Small perturbations of the image applied throughout optimization, in order to arrive at a more robust result.
[Note. The collage at the top is a sample of visualizations of Inception V3 (pre-trained on ImageNet) created using this repository. Each image in the k-th column comes from a channel in layer 'mixedk'.]
The following notebooks illustrates some applications of this repository:
This project only aims to cover the functionalities of Lucid discussed in this article on feature visualization. Within this scope, the main difference is our use of TensorFlow 2 instead of 1. In particular:
- The code is written and structured in eager mode.
- Optimization relies on a gradient tape.
- In view of the abundance of publicly available pre-trained models (e.g. via Keras Applications), no model zoo is included.
There are also a few minor, technical differences:
- A subclass of objectives is created for those depending only on a layer of a model. Combining multiple objectives associated to the same layer can be done without repeatedly referring to that layer or the underlying model, and invokes only a single evaluation of the layer.
- Only parametrizations in frequency (Fourier) space are considered. The standard parametrization in pixel space should be equivalent to one in frequency space with frequency decay rate set to zero.
- The translation from frequency space parameters to pixels is slightly modified so that the entire calculation is reversible (or, in math lingo, bijective).
- A tensor-native rotation function is implemented 'from scratch', as the one used in Lucid has been deprecated in TensorFlow 2 and the only off-the-shelf option seems to be in TensorFlow Addons, whose dependency we want to avoid.