git clone https://github.com/loctxmoreh/improved-diffusion
cd improved-diffusion

This example will use the cifar10 dataset, which has download script in ./datasets.

On A100 machine, use a100env.yml to create a conda environment with torch==1.7.1:

conda env create -f a100env.yml
conda activate iddpm

And a100env2.yml to create an environment with torch==1.12.1:

conda env create -f a100env2.yml
conda activate iddpm2

Run the training script train-baseline.sh to train a baseline iDDPM model. Checkpoints are saved in OPENAI_LOGDIR (set to ./log in the script).

./train-baseline.sh

Set the DIFFUSION_TRAINING_TEST environment variable to train for only a limited time. Or else, the training will go on forever.

To sample from a iDDPM checkpoint, run the sample script sample-baseline.sh. Checkpoints are saved in OPENAI_LOGDIR. It is recommended to sample from the ema_0.9999_*.pt checkpoints.

./sample-baseline.sh <path-to-checkpoint>

# For example:
./sample-baseline.sh log/ema_0.9999_001000.pt