Official implementation of FSDDIM

arxiv: https://arxiv.org/abs/2312.01742

1. Install requirements

   pip install -r requirements.txt

2. Login to Weights & Biases

   Please follow the official instruction of Weights & Biases for more details.

   wandb login

3. Initialize fid stats

   python calc_clean_fid_stats.py -c <config_file> -d 0 -o <output_directory>

   • args
     • -c, --config
       • Path to the config file.
       • Sample config files can be found in configs/.
     • -d, --gpu-id
       • GPU id.
     • -o, --output-dir
       • Directory in which dataset images will be saved.

We use Hugging Face Accelerate. Please follow the official instruction of Accelerate for more details.

accelerate launch --multi_gpu --num_processes=4 --gpu_ids=0,1,2,3 --mixed_precision fp16 main.py -c <config_file> -n <experiment_name>

• args
  • -c, --config
    • Path to the config file.
    • Sample config files can be found in configs/.
  • -n, --name
    • Experiment name.
    • Please specify a unique name because generated images will be saved in output/<experiment_name>.

The scores of evaluation metrics are approximately as follows.

Please cite our paper if you use this code in your own work:

@article{FSDDIM,
  title={Fully Spiking Denoising Diffusion Implicit Models},
  author={Watanabe, Ryo and Mukuta, Yusuke and Harada, Tatsuya},
  journal={arXiv preprint arXiv:2312.01742},
  year={2023}
}