Applcations of Machine Learning: Diffusion Models on MNIST

Introduction

In this project the objective is to develop a standard denoising diffusion probabilistic model and taking inspiration from the "Cold Diffusion" paper by Bansal et al., (2022) implement a custom diffusion model with a unique degradation strategy.

The repository contains a series of Python scripts that explore both stochastic and deterministic diffusion models. The following section provides instructions on how to run the main.py script.

by Max Talberg

Running the script locally

Setting Up the Project

1. Clone the repository:

   • Clone the repository from GitLab:

     git clone git@gitlab.developers.cam.ac.uk:phy/data-intensive-science-mphil/M2_Assessment/mt942.git

2. Set up the virtual environment:

   • Navigate to the project directory:

       cd mt942

   • Create virtual environment:

     conda env create -f environment.yml

   • Activate virtual environment:

     conda activate m2-env

3. Running the script:

   • Run the main script:

     python src/main.py --model <model_config> <options>

     • Replace <model_config> with the desired model:
       • default_model.yaml (default)
       • model1.yaml
       • model2.yaml
       • blur_model.yaml
     • Additional <options>:
       • --fid_score: Calculates and plots the FID score
       • --quick_test: Performs a quick test run (single bath)
     • Running only python src/main.py will run the default model with no additional options.

Notes

• Running the provided script will produce a sequence of plots in the content directory:
  • Grid plots of the MNIST dataset at each epoch
  • Progress plots over a range of timesteps at each epoch
  • Training loss plots
  • If specified, the FID score

Running the script on Docker

Setting Up the Project

1. Clone the repository:

   • Clone the repository from GitLab:

     git clone git@gitlab.developers.cam.ac.uk:phy/data-intensive-science-mphil/M2_Assessment/mt942.git

2. Build the Docker image:

   • Navigate to the project directory:

       cd mt942

   • Build image:

     docker build -t diffusion-model .

3. Running the script:

   • Run the main script:

     docker run -v $(pwd)/output:/app/output diffusion-model src/main.py --model <model_config> <options>

     • Replace <model_config> with the desired model:
       • default_model.yaml (default)
       • model1.yaml
       • model2.yaml
       • blur_model.yaml
     • Additional <options>:
       • --fid_score: Calculates and plots the FID score
       • --quick_test: Performs a quick test run (single bath)
     • Running only docker run my-diffusion-model src/main.py will run the default model with no additional options.
     • The images are mounted onto the current working directory into an 'output` folder.
     • The diffusion model runs a lot slower in Docker than the local script. As such I would recommend running the script locally or using --quick_test to speed up the process.

License

This project is licensed under the MIT License - see the LICENSE file for details.

Use of generative tools

This project has utilised auto-generative tools in the development of documentation that is compatible with auto-documentation tools, latex formatting and the development of plotting functions.

Example prompts used for this project:

• Generate doc-strings in NumPy format for this function.
• Generate Latex code for an algorithm.
• Generate Latex code for an equation.
• Generate Python code for a 9 by 1 subplot.