The Repository contains an XAI toolkit for analyzing ink detection and explaining predictions using various information bottleneck techniques from XAI literature.
The toolkit is designed to work with any type of models as the bottleneck is placed outside of the network (i.e in input space).
This repo builds on the findings of the First Letters and Grand prize of 2023.
I recommend using a docker image like pytorch/pytorch:2.0.1-cuda11.7-cudnn8-devel for your development environment. Kaggle/Colab images should work fine as well.
To install this project, run:
pip install -r requirements.txt
You can download this checkpoint (which is an old I3D I had laying around): link, or you can use any checkpoint you want (refer to externals.models for more details).
A resnet50 checkpoints for the fragments can also be found here
The repo implements different variation of IBA and diffmask
| Method Name | Description |
|---|---|
animate_history |
animates the volume layers importance throughout the optimization run |
attribute_z_diffmask |
Generates depth attribution using Diffmask approach |
attribute_z_with_diffmask_pooling |
Similar to Diffmask pooling with additional pooling to smooth out the depth probabilities. |
attribute_z_lowres |
A downsampled version for IBA for smoother attribution |
attribute_z |
The base IBA method for generating attributions along the z axis |
attribute_3d |
IBA based 3d attribution, performs attribution along z first, then attributes windows in the x-y. |