Implementation of DTL, a parameter-efficient paradigm that disentangles trainable parameters from a backbone model via a Compact Side Network (CSN). This repository supports both image classification and object detection tasks.

Minghao Fu, Ke Zhu, Jianxin Wu
DTL: Disentangled Transfer Learning for Visual Recognition, arxiv • AAAI Paper

• Memory-efficient: significantly reduces GPU memory footprint by offloading trainable parameters to CSN.
• Parameter-efficient: only a small fraction of parameters (LoRA-style) is fine-tuned.
• Accurate: achieves state-of-the-art performance on ImageNet, VTAB, and COCO benchmarks.
• Versatile: easy extension to image classification and object detection frameworks.
• Seamless: integrates with PyTorch and MMDetection pipelines.

• Classification: see classification/README.md for training and evaluation steps.
• Detection: see detection/README.md for COCO training and evaluation instructions.

If you use DTL in your work, please cite:

@inproceedings{fu_dtl2024,
author = {Fu, Minghao and Zhu, Ke and Wu, Jianxin},
title = {{DTL}: disentangled transfer learning for visual recognition},
year = {2024},
booktitle = {Proceedings of the Thirty-Eighth AAAI Conference on Artificial Intelligence},
pages={12082 - 12090}
}