Awesome

denoising-diffusion-pytorch

Implementation of Denoising Diffusion Probabilistic Models in PyTorch

Installation

First please install tensorfn

pip install tensorfn

It is simple convenience library for machine learning experiments. Sorry for the inconvenience.

Training

First prepare lmdb dataset:

python prepare_data.py --size [SIZES, e.g. 128,256] --out [LMDB NAME] [DATASET PATH]

Then run training looop!

python train.py --n_gpu [NUMBER OF GPUS FOR TRAINING] --conf config/diffusion.conf 

Also, you can tried to use improved configurations from https://openreview.net/forum?id=-NEXDKk8gZ (Cosine beta schedule, more multi-headed self attention, affine modulation for time conditions, new training hyperparameters.)

python train.py --n_gpu [NUMBER OF GPUS FOR TRAINING] --conf config/improved.conf 

Samples

Samples from FFHQ