Skip to content

threedle/TEDi

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

3 Commits
 
 
 
 
 
 

Repository files navigation

TEDi: Temporally-Entangled Diffusion for Long-Term Motion Synthesis

Installation

First clone the main branch and create the envrionment by running

conda env create -f environment.yml
conda activate tedi

Usage

Training

To train from scratch on the CMU dataset, first download the data from this Google Drive link. (Note: this is a very densely sampled version standing at 34G uncompressed unfortunately, see custom dataset if you want to use your own set of BVH files). Please extract the downloaded file and place it under a new directory data/processed/

Then, to train with the default set of hyperparameters described in src/util/options.py (you might want to replace the default data_path option with the actual path of your extracted data) run

cd src
python train_progressive_diffusion.py

(Note: Training is long and can easily take more than 24 hours depending on the GPU, but it does support training with accelerate credit to denoising-diffusion-pytorch)

Testing

To generate long sequences conditioning on a primer, you need to download the CMU data as in the previous section (the primers will be randomly selected motion sequences), and then download a pretrained model from this Google Drive link. Please extract the downloaded file and place it under a new directory exps/ under the project directory.

Then, run

cd src
python test_progressive_diffusion.py --exp_names {path to pretrained model directory} --n_samples {number of generated motions} --sample_len {number of frames}

(Note: primers for CMU data is 500 frames, so sample_len should be greater than that).

Custom dataset

Coming soon...

Acknowledgements

The code is based on denoising-diffusion-pytorch.

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages