Code and pre-trained models of ICLR 2024 paper Mind Your Augmentation: The Key to De-coupling Dense Self-supervised Learning.
[paper
] [project page
]
The work presents a thorough investigation of the impact of coupling shortcuts in dense self-supervised learning. We propose a de-coupling strategy to integrate with the existing dense self-supervised learning methods to mitigate the negative impact of the coupling shortcuts, which consistently improves the performance of dense-level evaluation tasks.
Acknowledgement: Our project is built using the iBOT repository.
- June 2024—Release the iBOT-based pre-trained models and code.
- Release the code and pre-trained models for other pre-training methods.
See installation structions for details. You also need to install Kornia for data augmentation:
pip install kornia==0.6.9
We perform pre-training on MS-COCO and dense-level evaluation on MS-COCO and ADE20K. Please download the datasets from their official websites and organize the data as follows:
── your project path/
└── data/
└── ade
│ ├── ADEChallengeData2016/
│ │ ├── annotations/
│ │ ├── images/
└── coco
├── annotations/
├── train2017/
└── val2017/
We use an 8-GPU machine with RTX 3090 GPUs to pre-train the models. To pre-train a ViT-S/16 model on MS-COCO with DDP, you can directly run the script as follows:
bash configs/Pretrain.sh
The parameter --dc
is used to enable the de-coupling strategy. You can remove it to pre-train the model with the vanilla setting.
The pre-trained models are available for download.DC
denotes the de-coupling strategy.
Model | Arch. | DC | COCO Det. | COCO ISeg. | ADE20K Seg. | Checkpoint |
---|---|---|---|---|---|---|
iBOT | ViT-S/16 | × | 42.3 | 37.0 | 39.9 | ckpt |
iBOT | ViT-S/16 | √ | 45.1 | 39.1 | 41.6 | ckpt |
The evaluation protocol is consistent with iBOT. You can refer to Evaluating iBOT on Downstream Tasks for details. Or you can directly run the script to perform dense-level evaluation on MS-COCO and ADE20K.
This repository is released under the MIT license as found in the LICENSE file.
If you find this repository useful, please consider giving a star ⭐ and citation:
@inproceedings{
qiu2024mind,
title={Mind Your Augmentation: The Key to Decoupling Dense Self-Supervised Learning},
author={Congpei Qiu and Tong Zhang and Yanhao Wu and Wei Ke and Mathieu Salzmann and Sabine S{\"u}sstrunk},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=WQYHbr36Fo}
}