code release for Inducing Neural Collapse to a Fixed Hierarchy-Aware Frame for Reducing Mistake Severity (ICCV2023)
Our code is running under Python 3.8.8
pip install -r requirements.txt
For iNaturalist2019 and tieredImageNet-H:
Refer to Repository: Making Better Mistakes
For CIFAR-100, we use torchvision.datasets.CIFAR100 API in:
util/data_loader.py
For FGVC-Aircraft, the dataset preparation code is at:
_fgvc_aircraft/fgvc_dataset_preprocessing.py
modify the /data_paths.yml files, e.g.:
fgvc-aircraft: 'C:/DATASET/fgvc_aircraft/'
cifar-100: 'C:/DATASET/CIFAR-100/'
inaturalist19-224: 'C:/DATASET/iNat2019/'
tiered-imagenet-224: 'C:/DATASET/tiered_ImageNet_H/'
For iNaturalist2019 and tieredImageNet-H:
Refer to Repository: Making Better Mistakes.
For CIFAR-100, its hierarchy is provided by Fine-Grained-or-Not.
For FGVC-Aircraft, its hierarchy comes with the dataset's annotations.
We also provide copies of the hierarchies used in this work:
data/fgvc_aircraft_tree.pkl
data/cifar-l5/original/cifar_100_tree.pkl
data/inaturalist19_tree.pkl
data/tiered_imagenet_tree.pkl
The associated hierarchical distances files are located at:
data/fgvc_aircraft_original_distances.pkl.xz
data/cifar-l5/original/cifar_100_ilsvrc_distances.pkl.xz
data/inaturalist19_ilsvrc_distances.pkl.xz
data/tiered_imagenet_ilsvrc_distances.pkl.xz
Training, testing, and neural collapse visualization scripts for dataset are included in:
experiments
e.g., training and evaluation scripts of the proposed HAFrame approach on tieredImageNet-H:
bash experiments/tiered-imagenet-224/type-II-haframe.sh
e.g., generating HAFrame vs ETF neural collapse for iNaturalist2019:
bash experiments/inaturalist19-224/neural-collapse-viz.sh
All the bash scripts need to be executed at the same directory with main.py
All of our experiments are conducted on machines with one Nvidia A100 GPU.
This codebase is largely refactored from the following GitHub repos: