COMP6248-Reproducibility-Challenge-2022

Trying to reproduce the paper "ImageNet-trained CNNs are biased towards texture; increasing shape bias improves accuracy and robustness " https://openreview.net/forum?id=Bygh9j09KX

Generating Stylized Tiny Image Net

Clone: https://github.com/bethgelab/stylize-datasets and follow instructions.

https://github.com/naoto0804/pytorch-AdaIN can also be used to generate style transferred images.

finetune_resnet_ontin.py

This script is used to finetune the ResNet50 images on Tiny-ImageNet. Each category of images should be in a separate folder so that they can be loaded by ImageFolder. tin_data_folders_format.py and stin_data_folders_format.py is used to rearrange the folder structure on linux.

test_models_general.py

pretrained models from original paper are loaded using the code in this repository in models/load_pretrained_models.py: https://github.com/rgeirhos/texture-vs-shape

tin_data_folders_format.py and stin_data_folders_format.py

These scripts are used to format the validation folders for the downloaded tiny-imagenet and stylized tiny-imagenet into a format where each image is in the folder corresponding to its category. This means the data can be loaded easily using torchvision.datasets ImageFolder.