Denoising diffusion probabilistic models (DDPM) are deep generative models that generate high quality images achieving state-of-the-art Fréchet inception distance (FID) and inception scores (IS) on benchmark datasets. However, sampling from the models is extremely time consuming compared to VAE, GAN and Flow-based models, making it computationally challenging to gain the full potential of these models on large datasets. This thesis presents an investigation of the balance between sampling time and image quality by analysing the effect of input dimensionality and model size. Based on these insights, this work introduces downsampled denoising diffusion probabilistic models (dDDPM), which adds light-weight downsampling and upsampling networks around the standard unconditional DDPM implementation. The dDDPMs achieve between 29 and 266 times faster sampling on
Start by installing some required packages listed in requirements.txt with
pip install -r requirements.txt
Then install torch for GPU (CUDA 11.1) with:
pip install torch==1.9.0+cu111 torchvision==0.10.0+cu111 torchaudio==0.9.0 -f https://download.pytorch.org/whl/torch_stable.html
Training of a dDDPM is performed by running train.py, where hyper paremeters can be altered in the config dict. To run a dDDPM x3 for 800k training steps on CelebAMask-HQ with a batch size of 32 use the following
python train.py -m ddpm -e 800000 -mute -d celeba_hq -bs 32 -is 256 -downsample 3
In order to run the standard DDPM set the -downsample flag to 0.
Additionally, samples can be generated from a saved checkpoint of the dDDPM by running generate_model_samples.py, where the saved_model variable inside the script is defined to be the name of the checkpoint located at the CHECKPOINT_DIR defined in utils.paths.
python generate_model_samples.py
This GitHub repository uses the following datasets: MNIST, CIFAR-10, CelebA and CelebAMask-HQ. The MNIST and CIFAR-10 datasets are simply downloaded using the download=True argument in their respective torchvision.datasets class. The CelebA dataset is downloaded from Kaggle and then resized to 64 x 64, while the CelebAMask-HQ is downloaded from a GitHub repo which is rescaled 256 x 256. For easy download you should be able to download all the datasets from my Google Drive folder.
I would like to thank Ole Winther and Giorgio Giannone for supervising this project and for all their insightful comments and suggestions throughout the course of this thesis.
If you are running weight and biases (wandb) as done in this project, it might be a good idea to once in a while delete local files that are synced to the website with the following command wandb sync --clean.
- Yann LeCun, Corinna Cortes, Christopher J.C. Burges, The MNIST Database
- Alex Krizhevsky, 2009, Learning Multiple Layers of Features from Tiny Images, Chapter 3
- Ziwei Liu, Ping Luo, Xiaogang Wang, Xiaoou Tang, 2014, Deep Learning Face Attributes in the Wild, arXiv:1411.7766
- Tero Karras, Timo Aila, Samuli Laine, Jaakko Lehtinen, 2017, Progressive Growing of GANs for Improved Quality, Stability, and Variation, arXiv:1710.10196
- Cheng-Han Lee, Ziwei Liu, Lingyun Wu, Ping Luo, 2019, MaskGAN: Towards Diverse and Interactive Facial Image Manipulation, arXiv:1907.11922
- https://github.com/lucidrains/denoising-diffusion-pytorch
- https://github.com/hojonathanho/diffusion
- https://github.com/openai/guided-diffusion
- https://github.com/openai/improved-diffusion