This is an unofficial implementation of "Denoising Diffusion Probabilistic Models" in PyTorch(Lightning). For a brief introduction to diffusion models, see my blog post.
To install requirements:
pip install -r requirements.txt
To train a diffusion model, first specify model architecture and hyperparameters in config.json. Once specified, run this command:
python diffusion_lightning.py --train --config config.json --ckpt_dir PATH_TO_CHECKPOINTS --ckpt_freq CHECKPOINT_FREQ --n_gpu NUM_AVAIL_GPUS
To generate samples from a trained diffusion model specified by config.json, run this command:
python diffusion_lightning.py --config config.json --model_dir MODEL_DIRECTORY --sample_dir PATH_TO_SAVE_SAMPLES --n_samples NUM_SAMPLES
Pre-trained diffusion models on CelebA and CIFAR-10 can be found here.
In the paper, the authors perform model selection using the FID score. Here, however, the model is only trained until 1000000 iterations and no model selection is performed due to limited computational resources. This way, we got an FID score of 5.1037.
This repository is built upon the official repository of diffusion models in TensorFlow as well as parts of this unofficial PyTorch implementation.