Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning

Yaxin Hou, Bo Han, Yuheng Jia, Hui Liu, Junhui Hou, Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning, Advances in Neural Information Processing Systems, 2nd-7th December, San Diego, 2025.

This is an official PyTorch implementation for Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic Long-Tailed Semi-Supervised Learning.

Introduction

This code is based on the public and widely-used codebase USB.

What I've done is just adding our CPG algorithm in semilearn/imb_algorithms/cpg.

Also, I've made corresponding modifications to semilearn/nets/ and several __init__.py.

How to run

For example, on CIFAR-10-LT with long-tailed labeled data ($\gamma_l=100$) and arbitrary unlabeled data($\gamma_u=100$)

CUDA_VISIBLE_DEVICES=0 python train.py --c "/config/config-1/cpg/203-fixmatch_cpg_cifar10_lb400_100_ulb4600_100_random_0.0_1.yaml"

(Note: I know that USB supports multi-GPUs, but I still recommend you to run on single GPU, as some weird problems may occur.)

The model will be automatically evaluated every 1024 iterations during training. After training, the last two lines in /saved_models/cpg/203-fixmatch_cpg_cifar10_lb400_100_ulb4600_100_random_0.0_1/log.txt will tell you the best accuracy.

For example,

[2025-04-20 13:35:17,784 INFO] model saved: ./saved_models/cpg/203-fixmatch_cpg_cifar10_lb400_100_ulb4600_100_random_0.0_1/latest_model.pth
[2025-04-20 13:35:17,815 INFO] Model result - eval/best_acc : 0.8233
[2025-04-20 13:35:17,816 INFO] Model result - eval/best_it : 244735

Results

The reported accuracies in Table 1, 2, and 3 in our paper are the average over three different runs (random seeds are 0/1/2).

Citation

If you find our method useful, please consider citing our paper:

@inproceedings{CPG,
author       = {Yaxin Hou and
                Bo Han and
                Yuheng Jia and
                Hui Liu and
                Junhui Hou},
title        = {Keep It on a Leash: Controllable Pseudo-label Generation Towards Realistic
                Long-Tailed Semi-Supervised Learning},
booktitle    = {Advances in Neural Information Processing Systems (NeurIPS)},
volume       = {38},
pages        = {118519--118549},
year         = {2025}
}