Denoising Heat-inspired Diffusion

Training and visualization of the model from Denoising Heat-inspired Diffusion with Insulators for Collision Free Motion Planning (Neurips 2023 Workshop on Diffusion Models).

Installation

conda create -n dnheat python=3.9
conda activate dnheat
pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu121
pip install tensorboard
pip install -r requirements.txt

Using pretrained model

Please download the 'runs' file from this link and incorporate it into the repository.

Inference

Evaluate the model in a random or pre-set map, as specified in inference.ipynb, applying both the pre-trained weights and the hyperparameters.

Training from scratch

1. Train your own model using train.ipynb

   • Default hyperparameters are located in ./denoisingheat/configs/heat_diffusion.yaml.
   • You can modify the hyperparameters in the heat_diffusion.yaml file.

2. Evaluate your trained model using inference.ipynb

   • Update the 'config_dir' path in the first block and the '.pt' file path in the second block to reflect the directory where you saved your new model.

3. Monitor the training progress and analyze logs via TensorBoard:

tensorboard --logdir=./runs

Reference

@article{chang2023denoising,
  title={Denoising Heat-inspired Diffusion with Insulators for 
Collision Free Motion Planning},
  author={Chang, Junwoo and Ryu, Hyunwoo and Kim, Jiwoo and Yoo, Soochul and Seo, Joohwan and Prakash, Nikhil and Choi, Jongeun and Horowitz, Roberto},
  journal={arXiv preprint arXiv:2310.12609},
  year={2023}
}

Acknowledgement

Our diffusion model is implemented based on Phil Wang's 'denoising-diffusion-pytorch' GitHub repository.