The repository of the paper "Stochastic Covariance Regularization for Imbalanced Datasets", accepted into ICANN 2025.
pip install scor
The "scor" package is organized as a library that could easily be imported and used in applications. It includes all utility functions, measurement functions, model classes and data loaders used in the paper, organized into modules.
This module includes data utilities and data loader functions for datasets that were used in the paper. Data loading functions for MNIST, CIFAR-10, Flower102, Oxford IIIT Pets, DTD, STL-10 and LFW People are given. For LFW People, you need to download it then extract it first, because the link that PyTorch provides within its own data loader is broken.
This module includes the SCoR loss function alongside other that were used in the paper.
This module has utility functions that calculate ranks and singular values of weight matrices of models, that were interpreted as indicators of matrix perturbations in the paper.
This module hosts model classes used in tests. It does not include ResNet32 and ResNet50 classes, because respective tests in the paper were adopted from other sources. Several options for ConvNeXt is supported, but in the paper only ConvNeXt-Tiny is used.
This module hosts several training functions for several tests in the paper. These functions cover first phase of experiments in the paper, where MNIST and CIFAR-10 is artificially imbalanced by imbalancing a random class, or 9 random classes multiplicatively by a reduction factor.
Example notebooks that use the "scor" package for training on given datasets are provided under the "examples" directory.
This file is for training models with various loss functions on the ImageNet dataset. Code from here is adapted to be used with loss functions defined in this repository.
A "loss" option is included to the adapted code. Run python imagenet.py --help
for a proper description.