DiffusionModel

An easy and straightforward implementation of Diffusion Model for Image Generation

Dataset

For this experiment I am using CIFAR-10 dataset, which gets downloaded in the dataset.py script

Run commands

1. mkdir model
2. cd model
3. Make directory by name of your run say mkdir nk_diffusion
4. cd ../..
5. mkdir results
6. cd results
7. Make directory by name of your run say mkdir nk_diffusion
8. cd ../..

Model training

To train and sample from the model, run the command(with or without specific parameters)

python3 main.py

Parameters List

1. diffusion_parser.add_argument("--name", type=str, default="nk_diffusion",
                                  help="Give name to this run of Diffusion Model")
2. diffusion_parser.add_argument("--dataset_path", type=str, default='./data')
3. diffusion_parser.add_argument("--epochs", type=int, default=10)
4. diffusion_parser.add_argument("--batch_size", type=int, default=64)
5. diffusion_parser.add_argument("--img_size", type=int, default=32)
6. diffusion_parser.add_argument("--num_classes", type=int, default=10)
7. diffusion_parser.add_argument("--device",
                                  default=torch.device(
                                      "cuda" if torch.cuda.is_available() else "cpu"))
8. diffusion_parser.add_argument("--lr", type=float, default=1e-5)
9. diffusion_parser.add_argument("--guidance_scale", type=int, default=3)

You might need to make some changes, if you are not running this code on CUDA enabled device in model.py file. The parameters are not tuned to generate best out of the dataset.

Presentation

This repo also contains a presentation delivered during the Technology Upskilling Series hosted by The Department of Computer Science at University of Toronto with the Vector Institute for AI on 25th March 2023.

Follow ups

Visit my blogging channel to read more around Diffusion and other advancements in the space of Deep Learning.