Diffusion-SR

This repository contains the implementation of a diffusion model for super-resolution. The project aims to enhance image resolution using diffusion techniques.

Table of Contents

• Installation
• Usage
  • Running Training
  • Changing Hyperparameters
• Inference

Installation

1. Clone the repository:

   git clone https://github.com/AngelUrq/diffusion-sr.git
   cd diffusion-sr

2. Download the DSR data and place it in the root folder of the repository.

Usage

Files Description

• ddpm.ipynb: Jupyter notebook for Denoising Diffusion Probabilistic Models (DDPM) implementation.
• diffusion.py: Implementation of the diffusion scheduling.
• dsr.py, dsr_height.py: Pytorch datasets for DSR.
• eda.ipynb: Exploratory Data Analysis notebook.
• eval.ipynb: Notebook for model evaluation.
• launch_tensorboard.sh: Script to launch TensorBoard for monitoring training.
• launch_training.sh: Script to launch the training process.
• model.py: U-net implementation, not used in the final version. We adopted Diffusers implementation.
• samplers.py: Contains sampling methods for the diffusion process.
• srd.py: Pytorch dataset for flowers.
• super_resolution.ipynb: Notebook demonstrating the super resolution process.
• train.py, train_height.py: Training scripts.

Running Training

To run the training process, execute the following script:

bash launch_training.sh

This script will initiate the training using the default settings specified in train.py.

Changing Hyperparameters

To change hyperparameters, modify the train.py file. Here are some key hyperparameters you might want to adjust:

• learning_rate: Adjust the learning rate for the optimizer.
• batch_size: Change the number of samples per batch.
• num_epochs: Set the number of training epochs.
• data_path: Specify the path to your training dataset.

Example:

# train.py
learning_rate = 0.001
batch_size = 32
num_epochs = 100
data_path = 'path/to/your/dataset'

Inference

Pre-trained models are available in the models folder. To perform inference using these models, run the following notebook:

jupyter notebook super_resolution.ipynb

This notebook contains code to load the pre-trained models and apply them to your images for super-resolution. Note that this method works for models trained with artificial images.

For inference with models trained on real data, we use the Image-Super-Resolution-via-Iterative-Refinement GitHub repository. We chose this approach to validate our model by testing it with different codebases. In this case, you need to modify config/sr_sr3_16_128.json. We have already adjusted this config to load our latest checkpoint and the model files are in the experiments folder. Instructions for running the model can be found in the README of that repository.