Skip to content

Latest commit

 

History

History
68 lines (58 loc) · 2.68 KB

README.md

File metadata and controls

68 lines (58 loc) · 2.68 KB
Raw

Guided-DIP

Guiding DIP Early Stopping with DDPM-inspired Supervision

The framework

The highlight words are parameters for guided-dip. In this project, I do not conduct the ablation studies on it.

How to start the code

install Dependencies

  • running on your local device
pip install -r requirements.txt
  • running in colab
pip install python-dotenv

create the .env file

First, you should copy the .env.example to .env

cp .env.example .env

Download dataset if you need it

I recommend you run the code with the images in data/test directory without installing any files first.
  • replace the your kaggle token to the following two lines in the .env file
KAGGLE_USERNAME="your user name in kaggle"
KAGGLE_KEY="your api key in kaggle"
  • Then, run the following shell script. It will download the image from kaggle dataset
./download_data.sh

Notebook details

  • test.ipynb is the demonstration of simple DDPM diffusion forward process and DIP training process on the clean image
  • main.ipynb is the notebook for running the Guided-DIP and DIP on noisy image

How to run with different experiments

  • all the settings of dip, ddpm and guided-dip can be altered by rewriting the config.yaml, you should follow the models.model.py configuration parsing for details.

  • Here, I show some changes you can try to rewrite

    • backbone_name: segnet or unet

    • scheduler: cosine or linear

    • device: cuda or cpu

    • For more detailed comparisons, you can change any numerical value to find the best parameter setting (like count, beta_start, learning_rate, etc) in the psnr score

The whole comparison

numerical metrics

  • With my proposed method Guided-DIP, I beat the original DIP in both best PSNR score and final PSNR score by leveraging cosine beta scheduler.
  • Another interesting result is that I don't need to set early-stop point anymore. The final result is good enough and the best time is closed to stop time.

generated results

  • The restored quality is much better when using SegNet

More details

To see more details, please open the report.pdf file. There are pseudo code of the training method, detailed experiment settings, descriptions about every parameters, and snapshots during DIP/Guided-DIP training procedure.

Acknowledgements