A repository for generating salient co-speech gesture. This is the repository for paper Salient Co-Speech Gesture Synthesizing with Discrete Motion Representation.
conda create -f environment.yml python=3.8
- Download dataset to
data/baijia_all
- Run
train_vae.py -c data/motion_vqvae.yml
to train Motion Auto-Encoder. - Run
train_transformer.py -c data/motion_vqvae.yml
to train Speech-Gesture Mapping network. - Run
evaluate_transformer.py
to evaluate FGD and MVD. - Run
inference_transformer.py
to evaluate SMS.
We leverage lmdb
to store the dataset. The data can be downloaed at link, and then run cat xa* > data.mdb
.
TODO