This repository contains the code used in the reproduction, it builds on top of some existing functionality provided by Geirhos et al. in this repository.
The code is split into two chunks, evaluate.py and analyse.py. The code requires the probabilities_to_decision.py and the scripts in the helper folder from the repository of Geirhos et al.
This script takes as input some cue-conflict images with the names having a specific format and evaluates several pretrained models on these images. The result is then stored with indications of predicted category, shape category, and texture category.
The format for the name of the images is as follows {x}-{y}.JPEG, the extension does not matter, the important part is that {x} is the shape category and {y} is the texture category of the image. Both {x} and {y} can have a number appended to it to distinguish images with the same shape-texture combination.
This script takes csv files resulting from evaluate.py and creates a figure similar to Figure 4 of the paper in Geirhos et al.