DL2PA: Hyperspherical Classification with Dynamic Label-to-Prototype Assignment

Official repository for the paper DL2PA: Hyperspherical Classification with Dynamic Label-to-Prototype Assignment (CVPR 2024). Arxive

Abstract

Aiming to enhance the utilization of metric space by the parametric softmax classifier, recent studies suggest replacing it with a non-parametric alternative. Although a non-parametric classifier may provide better metric space utilization, it introduces the challenge of capturing inter-class relationships. A shared characteristic among prior non-parametric classifiers is the static assignment of labels to prototypes during the training, i.e., each prototype consistently represents a class throughout the training course. Orthogonal to previous works, we present a simple yet effective method to optimize the category assigned to each prototype (label-to-prototype assignment) during the training. To this aim, we formalize the problem as a two-step optimization objective over network parameters and label-to-prototype assignment mapping. We solve this optimization using a sequential combination of gradient descent and Bipartide matching. We demonstrate the benefits of the proposed approach by conducting experiments on balanced and long-tail classification problems using different backbone network architectures. In particular, our method outperforms its competitors by 1.22% accuracy on CIFAR-100, and 2.15% on ImageNet-200 using a metric space dimension half of the size of its competitors.

Comparison of the proposed method with the conventional PSC and the previous fixed classifier setup, using a toy example with three classes. Each color denotes a distinct class. a) Label-to-prototype assignment remains static during training. In PSC, optimization focuses on the network, consisting of the backbone and prototypes W . In the case of a fixed classifier, only the backbone is optimized, and prototypes remain fixed. b) In the proposed method, prototypes within the hypersphere are fixed, and optimization targets the backbone and the label that each prototype represents. c) Toy example showing changes in label-to-prototype assignment during training.

Table of Contents

• Usage
• Balanced Classification Results
• Long-tailed Classification Results
• Citation
• Acknowledgments

Usage

Prototype Estimation

One can generate equidistributed prototypes with desired dimension:

python Prototype_Estimation.py --seed 100 --num_centroids 100 --batch_size 100 --space_dim 50 --num_epoch 1000

Also, you can find the estimated prototype in link

Training classifier

The configs can be found in ./config/Blanced or LongTail/FILENAME.yaml.

python train.py --cfg {path to config}

Balanced Classification Results

Method	CIFAR-10				ImageNet-200
	d=10	d=25	d=50	d=100	d=25	d=50	d=100	d=200
PSC	25.67	60.0	60.6	62.1	60.0	60.6	62.1	33.1
Word2Vec	29.0	44.5	54.3	57.6	44.5	54.3	57.6	30.0
HPN	51.1	63.0	64.7	65.0	63.0	64.7	65.0	44.7
Ours	57.21	64.63	66.22	62.85	64.63	66.22	62.85	37.28

ImageNet-1K Classification Accuracy (%) when $d=512$:

Method	Venue	Backbone	Optimizer	Accuracy (%)
PSC	CVPR 2016	ResNet-50	SGD	76.51
DNC	ICLR 2022	ResNet-50	SGD	76.49
Goto et al.	WACV 2024	ResNet-50	SGD	77.19
Kasarla et al.	NeurIPS 2022	ResNet-50	SGD	74.80
Ours	CVPR 2024	ResNet-50	SGD	77.47
DNC	ICLR 2022	ResNet-101	SGD	77.80
Goto et al.	WACV 2024	ResNet-101	SGD	78.27
Kasarla et al.	NeurIPS 2022	ResNet-152	SGD	78.50
Ours	CVPR 2024	ResNet-101	SGD	79.63
PSC	CVPR 2016	Swin-T	AdamW	76.91
Ours	CVPR 2024	Swin-T	AdamW	77.26

Long-tailed Classification Results

Method	CIFAR-10 LT (d=64)			SVHN LT (d=64)			STL-10 LT (d=64)
	0.005	0.01	0.02	0.005	0.01	0.02	0.005	0.01	0.02
PSC	67.3	72.8	78.6	40.5	40.9	49.3	33.1	37.9	38.8
ETF	71.9	76.5	81.0	42.8	45.7	49.8	33.5	37.2	37.9
Ours	71.5	76.9	81.4	40.9	47.0	49.7	35.7	35.6	38.0

CIFAR-100 LT Classification Accuracy (%):

Method	d	0.005	0.01	0.02
PSC	128	38.7	43.0	48.1
ETF	128	40.9	45.3	50.4
Ours	128	41.3	44.9	50.7

Citation

@misc{saadabadi2024hyperspherical,
    title={Hyperspherical Classification with Dynamic Label-to-Prototype Assignment},
    author={Mohammad Saeed Ebrahimi Saadabadi and Ali Dabouei and Sahar Rahimi Malakshan and Nasser M. Nasrabad},
    year={2024},
    eprint={2403.16937},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

Acknowledgments

Here are some great resources we benefit from:

• MUNIT for the Long-Tail classification.
• HPN and EBV for the Balanced classification.
• Understanding Contrastive Representation Learning through Alignment and Uniformity on the Hypersphere for Prototype Estimation.

Contact

If there is a question regarding any part of the code, or it needs further clarification, please create an issue or send me an email: [email protected].