This repository contains the code (in PyTorch) for the model introduced in the following paper:
Debiased Learning from Naturally Imbalanced Pseudo-Labels
Xudong Wang, Zhirong Wu, Long Lian, and Stella X. Yu
UC Berkeley and Microsoft Research
CVPR 2022
Project Page | Paper | Preprint | Citation
If you find our work inspiring or use our codebase in your research, please consider giving a star ⭐ and a citation.
@inproceedings{wang2022debiased,
title={Debiased Learning from Naturally Imbalanced Pseudo-Labels},
author={Wang, Xudong and Wu, Zhirong and Lian, Long and Yu, Stella X},
booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages={14647--14657},
year={2022}
}
[06/2022] Support DebiasPL w/ CLIP for more label-efficient learning. DebiasPL (ResNet50) achieves 69.6% (71.3%) top-1 accuray on ImageNet only using 0.2% (1%) labels!
[04/2022] Initial Commit. Support zero-shot learning and semi-supervised learning on ImageNet.
- Python >= 3.7, < 3.9
- PyTorch >= 1.6
- torchaudio==0.7.2
- tensorboard >= 1.14 (for visualization)
- tqdm
- faiss-gpu
- pandas
- apex (optional, unless using mixed precision training)
8 GPUs with >= 11G GPU RAM or 4 GPUs with >= 16G GPU RAM are recommended.
Please download pre-trained MoCo-EMAN model, make a new folder called pretrained and place checkpoints under it. Please download the ImageNet dataset from this link. Then, move and extract the training and validation images to labeled subfolders, using the following shell script. The indexes for semi-supervised learning experiments can be found at here. The setting with 1% labeled data is the same as FixMatch. A new list of indexes is made for the setting with 0.2% labeled data by randomly selecting 0.2% of instances from each class. Please put all CSV files in the same location as below:
dataset
└── imagenet
├── indexes
│ ├── train_1p_index.csv
│ ├── train_99p_index.csv
| └── ....
├── train
│ ├── n01440764
│ │ └── *.jpeg
| └── ....
└── val
├── n01440764
│ └── *.jpeg
└── ....
0.2% labeled data (50 epochs):
bash scripts/0.2perc-ssl/train_DebiasPL.sh
1% labeled data (50 epochs):
bash scripts/1perc-ssl/train_DebiasPL.sh
1% labeled data (DebiasPL w/ CLIP, 100 epochs):
bash scripts/1perc-ssl/train_DebiasPL_w_CLIP.sh
| Method | Backbone | epochs | 0.2% labels | 1% labels |
|---|---|---|---|---|
| FixMatch w/ EMAN | RN50 | 50 | 43.6% | 60.9% |
| DebiasPL (reported) | RN50 | 50 | 51.6% | 65.3% |
| DebiasPL (reproduced) | RN50 | 50 | 52.0% [ckpt | log] | 65.6% [ckpt | log] |
| DebiasPL w/ CLIP (reproduced) | RN50 | 50 | 69.6% [ckpt | log] | - |
| DebiasPL w/ CLIP (reproduced) | RN50 | 100 | 70.4% [ckpt | log] | 71.3% [ckpt | log] |
The results reproduced by this codebase are often slightly higher than what was reported in the paper (52.0 vs 51.6; 65.6 vs. 65.3). We find it beneficial to apply cross-level instance-group discrimination loss CLD to unlabeled instances to leverage their information fully.
Please download zero-shot predictions with a pre-trained CLIP (backbone: RN50) model and put them under imagenet/indexes/. Then run experiments on ImageNet-1k with:
bash scripts/zsl/train_DebiasPL.sh
| Method | Backbone | epochs | top-1 acc |
|---|---|---|---|
| CLIP | RN50 | - | 59.6% |
| CLIP | ViT-Base/32 | - | 63.2% |
| DebiasPL (reported) | RN50 | 100 | 68.3% |
| DebiasPL (reproduced) | RN50 | 50 | 68.7% [ckpt | log] |
If you have any general questions, feel free to email us at xdwang at eecs.berkeley.edu. If you have code or implementation-related questions, please feel free to send emails to us or open an issue in this codebase (We recommend that you open an issue in this codebase, because your questions may help others).
This project is licensed under the MIT License. See LICENSE for more details. The parts described below follow their original license.
Part of the code is based on EMAN, FixMatch, CLIP, CLD, and LA.