VS loss

Implementation of VS loss for deep-net experiments on imbalanced classification reported in the NeurIPS 2021 paper:

Ganesh Ramachandra Kini, Orestis Paraskevas, Samet Oymak, Christos Thrampoulidis

Label-Imbalanced and Group-Sensitive Classification under Overparameterization

Abstract

The goal in label-imbalanced and group-sensitive classification is to optimize relevant metrics such as balanced error and equal opportunity. Classical methods, such as weighted cross-entropy, fail when training deep nets to the terminal phase of training (TPT), that is training beyond zero training error. This observation has motivated recent flurry of activity in developing heuristic alternatives following the intuitive mechanism of promoting larger margin for minorities. In contrast to previous heuristics, we follow a principled analysis explaining how different loss adjustments affect margins. First, we prove that for all linear classifiers trained in TPT, it is necessary to introduce multiplicative, rather than additive, logit adjustments so that the interclass margins change appropriately. To show this, we discover a connection of the multiplicative CE modification to the cost-sensitive support-vector machines. Perhaps counterintuitively, we also find that, at the start of training, the same multiplicative weights can actually harm the minority classes. Thus, while additive adjustments are ineffective in the TPT, we show that they can speed up convergence by countering the initial negative effect of the multiplicative weights. Motivated by these findings, we formulate the vector-scaling (VS) loss, that captures existing techniques as special cases. Moreover, we introduce a natural extension of the VS-loss to group-sensitive classification, thus treating the two common types of imbalances (label/group) in a unifying way. Importantly, our experiments on state-of-the-art datasets are fully consistent with our theoretical insights and confirm the superior performance of our algorithms. Finally, for imbalanced Gaussian-mixtures data, we perform a generalization analysis, revealing tradeoffs between balanced / standard error and equal opportunity.

Dependencies

• python 3.8.10
• matplotlib 3.0.3
• numpy 1.17.4
• pandas 0.25.3
• Pillow 8.4.0
• pytorch_transformers 1.2.0
• scikit-learn 1.0.1
• torch 1.10.0
• torchvision 0.10.0
• TensorboardX 2.1
• tqdm 4.62.3

Class Imbalance

We built on the implementation of LDAM-DRW, to which we add the implementation of VS loss. LDAM loss was featured on the paper:

Kaidi Cao, Colin Wei, Adrien Gaidon, Nikos Arechiga, Tengyu Ma

Learning Imbalanced Datasets with Label-Distribution-Aware Margin Loss

Datasets

• This repository contains our experiments on the CIFAR10 and CIFAR100 datasets. You can load these datasets with the STEP and LT imbalance profiles and with different imbalance ratios using generate_cifar.py.
• For the experiments on synthetic data, MNIST and ImageNet please contact the authors.

Sample commands

To train a network on CIFAR with one of the available losses, you will need to run the train.py file from the class_imbalance directory. Find some example commands below:

• To train on CIFAR100 with the STEP-10 imbalance profile using the wCE loss run:

python class_imbalance/train.py --dataset cifar100 --loss_type CE --imb_type step --imb_factor 0.1 --train_rule Reweight

• To train on CIFAR10 with the LT-100 profile using the VS loss with parameters gamma=0.2 and tau=1.2 run:

python class_imbalance/train.py --dataset cifar10 --loss_type VS --imb_type exp --imb_factor 0.01 --gamma 0.2 --tau 1.2

The LA and CDT losses are also implemented and can be used as options. All losses can be used with standard ERM, resampling or the DRW training rule.

Group Imbalance

We built on the implementation of group DRO, to which we add the implementation of group-VS loss. Group DRO was featured on the paper:

Shiori Sagawa*, Pang Wei Koh*, Tatsunori Hashimoto, and Percy Liang

Distributionally Robust Neural Networks for Group Shifts: On the Importance of Regularization for Worst-Case Generalization

Dataset

For the group-sensitive classification experiments we use the Waterbirds dataset, which consists of 2 classes (waterbirds and landbirds) and 4 groups (waterbirds and landbirds on land and water backgrounds). For more details about the waterbirds dataset please refer to the group DRO repository and paper.

Before running any group-sensitive experiments the dataset must be downloaded and placed in the group_imbalance/data folder. You can do this by running the following commands:

wget -P group_imbalance/data https://nlp.stanford.edu/data/dro/waterbird_complete95_forest2water2.tar.gz
tar -xf group_imbalance/data/waterbird_complete95_forest2water2.tar.gz -C group_imbalance/data

Sample commands

To perform group-sensitive classification on the Waterbirds dataset with VS loss, you will need to run the run_expt.py file from the group_imbalance directory. You can also use DRO in conjunction with any of the losses.

• To train on Waterbirds with VS-loss and DRO:

python group_imbalance/run_expt.py -s confounder -d CUB -t waterbird_complete95 -c forest2water2 --lr 0.001 --batch_size 64 --weight_decay 0.0001 --model resnet50 --n_epochs 300  --gamma 0.1 --generalization_adjustment 0 --loss vs --vs_alpha 0.3 --dont_set_seed 1 --robust

(Note: the hyperparameter vs_alpha is same as the hyperparameter gamma=0.3 used in the paper)

Reference

If you find our paper or this repository helpful for your research, please consider citing it as:

@article{kini2021label,
  title={Label-Imbalanced and Group-Sensitive Classification under Overparameterization},
  author={Kini, Ganesh Ramachandra and Paraskevas, Orestis and Oymak, Samet and Thrampoulidis, Christos},
  journal={arXiv preprint arXiv:2103.01550},
  year={2021}
}