This repository allow the creation of a dataset and a model adapted to this dataset.
The dataset is a classification dataset of images, the images have four parts and each part have a shape. One of the parts is highlighted (selected), the selected shape is the one we want to predict.
The model was built specifically for this dataset and every weights were set manually. The objective is to have a neural network that we understand to evaluate explainability methods. We will compare the explanations with the functionment we know and take a step back on what explanation methods really provide.
The model is built with four blocks:
- the first one that separates the different part of the image,
- the second that find which is the part to look at,
- the third in parallel of the second that find shapes scores for each part,
- the fourth one using both previous part to selected the shapes scores.
