Dropout Diffusion

This codebase is forked from openai/improved-diffusion.

This repo implements code for Dropout Diffusion, an variant of DDPM focusing on boosting up sampling speed.

Quick start

Install requirements

pip install -e .

pip install mpi4py

Training

The are several configuration arguments for training, sampling and lipschitz calculation. The easiest way to manage this is export is as environment variables.

Example:

export DIFFUSION_FLAGS="--diffusion_steps 200 --noise_schedule exponential --diffusion_dropout 0.002 --num_sample 100"
export MODEL_FLAGS="--image_size 32 --num_channels 128 --num_res_blocks 3 --dropout 0.1"
export TRAIN_FLAGS="--lr 1e-4 --batch_size 64 --save_interval 10000"

python scripts/image_train.py --data_dir {path_to_data_dir} $DIFFUSION_FLAGS $MODEL_FLAGS $TRAIN_FLAGS

For more detail, read CONFIGURATION.md.

Sampling

Sampling is similar to training, the only diffirence is replacement of SAMPLE_FLAGS for TRAIN_FLAGS.

Lipschitz calculation