Code for the Paper "Restricting the Flow: Information Bottlenecks for Attribution"

This is the source code for the paper "Restricting the Flow: Information Bottlenecks for Attribution" - Oral at ICLR2020.

Note: This implementation might not be up-to-date. The reference implementation is in another repository

Iterations of the Per-Sample Bottleneck

Setup

1. Clone this repository:

   $ git clone https://github.com/attribution-bottleneck/attribution-bottleneck-pytorch.git && cd attribution-bottleneck-pytorch
   

2. Create a conda environment with all packages:

   $ conda create -n new environment --file requirements.txt
   

3. Using your new conda environment, install this repository with pip:

   $ pip install .
   

4. Download the model weights from the release page and unpack them in the repository root directory:

   $ tar -xvf bottleneck_for_attribution_weights.tar.gz
   

Optional:

5. If you want to retrain the Readout Bottleneck, place the imagenet dataset under data/imagenet. You might just create a link with ln -s [image dir] data/imagenet.

6. Test it with:

   $ python ./scripts/eval_degradation.py resnet50 8 Saliency test
   

Usage

We provide some jupyter notebooks to demonstrate the usage of both per-sample and readout bottleneck.

• example_per-sample.ipynb : Usage of the Per-Sample Bottleneck on an example image
• example_readout.ipynb : Usage of the Readout Bottleneck on an example image
• compare_methods.ipynb : Visually compare different attribution methods on an example image

Scripts

The scripts to reproduce our evaluation can be found in the scripts directory. Following attributions are implemented:

For the bounding box task, replace the model with either vgg16 or resnet50.

$eval_bounding_boxes.py [model] [attribution]

For the degradation task, you also have specify the tile size. In the paper, we used 8 and 14.

$ eval_degradation.py [model] [tile size] [attribution]

The results on sensitivity-n can be calculated with:

eval_sensitivity_n.py [model] [tile size] [attribution]