In this repository, we introduce a cutting-edge technique termed "Frequency Regularization" that remarkably compresses a UNet model from 31 million parameters down to a mere 759 non-zero parameters without compromising on performance. This technique has been meticulously tailored and evaluated, achieving an impressive Dice Score of over 97% on the Carvana Image Masking Challenge dataset. The original UNet model has a size of 366MB, but with our frequency regularization applied, the size is drastically reduced to 4KB, as demonstrated in the unet_fr.tar.xz file included in this demo.
- Ensure that you have the
imageiopackage installed for loading testing images. Follow the installation instructions if you don't have it installed yet.
- Clone this repository to your local machine.
- Navigate to the project directory.
- Execute the following command to run the demo:
bash run.sh
Due to the double-blind policy, comments within the code have been omitted. However, the demonstration is structured to be straightforward and the execution command provided should seamlessly run the demo.
Feel free to reach out for any inquiries or further clarifications regarding the implementation and performance of the frequency regularization technique on the UNet model.
| Milestones | Status |
|---|---|
| Initial U-net Model | ✔️ Completed |
| Additional Model Releases | 🔜 Upcoming |
| Example Training Code | 🔜 Upcoming |
| Pip Repository Creation | 🔜 Upcoming |
As of now, we have successfully developed and validated an initial model utilizing the U-net architecture. This serves as a proof of concept demonstrating the efficacy of our Frequency Regularization technique in significantly compressing the model size while retaining a high level of performance.
Looking ahead, we plan to release a variety of models alongside their corresponding training code to provide a more comprehensive understanding and utilization of our technique. These releases will encapsulate a broad range of architectures and datasets to demonstrate the versatility and robustness of Frequency Regularization.
Furthermore, we are in the process of packaging our compression technique into a pip repository. Once completed, this will facilitate an effortless installation and integration of our Frequency Regularization technique into existing and new projects, thereby advancing the ease of deploying highly efficient and compact models in real-world scenarios.
If you find our Frequency Regularization technique intriguing or utilize it in your research, we kindly encourage you to cite our paper:
@ARTICLE{10266314,
author={Zhao, Chenqiu and Dong, Guanfang and Zhang, Shupei and Tan, Zijie and Basu, Anup},
journal={IEEE Access},
title={Frequency Regularization: Reducing Information Redundancy in Convolutional Neural Networks},
year={2023},
volume={11},
number={},
pages={106793-106802},
doi={10.1109/ACCESS.2023.3320642}
}The paper is published in IEEE Access, and you can access it here.
Title: Frequency Regularization: Reducing Information Redundancy in Convolutional Neural Networks
Your acknowledgment greatly supports our ongoing research and contributes to fostering advancements in network compression techniques.