Skip to content

code for training the models from the paper "Learning Individual Styles of Conversational Gestures"

Notifications You must be signed in to change notification settings

amirbar/speech2gesture

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

24 Commits
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Learning Individual Styles of Conversational Gestures

alt text

Prerequisites:

  1. python 2.7
  2. cuda 9.0
  3. cuDNN v7.6.2
  4. sudo apt-get install ffmpeg
  5. pip install -r requirments.txt

Data

  1. Download the dataset as described here

Instructions

  1. Extract training/validation data
  2. Train a model
  3. Perform inference using a trained model

Extract training data

Start by extracting training data:

python -m data.train_test_data_extraction.extract_data_for_training --base_dataset_path <base_path> --speaker <speaker_name> -np <number of processes> --speaker <speaker name>`
once done you should see the following directories structure:
(notice train.csv and a train folder within the relevant speaker)

Gestures
├── frames.csv
├── train.csv
├── almaram
    ├── frames
    ├── videos
    ├── keypoints_all
    ├── keypoints_simple
    ├── videos
    └── train
...
└── shelly
    ├── frames
    ├── videos
    ├── keypoints_all
    ├── keypoints_simple
    ├── videos
    └── train

train.csv is a csv file in which every row represents a single training sample. Unlike in frames.csv, here, a sample is few seconds long.
alt text

Columns documentation:

audio_fn - path to audio filename associated with training sample
dataset - train/dev/test
start - start time in the video
end - end time in the video
pose_fn - path to .npz file containing training sample
speaker - name of a speaker in the dataset
video_fn - name of the video file

Training a speaker specific model

Training run command example:

python -m audio_to_multiple_pose_gan.train --gans 1 --name test_run --arch_g audio_to_pose_gans --arch_d pose_D --speaker oliver --output_path /tmp

During training, example outputs are saved in the define output_path

Inference

optionally get a pretrained model here.

Perform inference on a random sample from validation set:

python -m audio_to_multiple_pose_gan.predict_to_videos --train_csv <path/to/train.csv>--seq_len 64 --output_path </tmp/my_output_folder> --checkpoint <model checkpoint path> --speaker <speaker_name> -ag audio_to_pose_gans --gans 1

Perform inference on an audio sample:

python -m audio_to_multiple_pose_gan.predict_audio --audio_path <path_to_file.wav> --output_path </tmp/my_output_folder> --checkpoint <model checkpoint path> --speaker <speaker_name> -ag audio_to_pose_gans --gans 1

Reference

If you found this code useful, please cite the following paper:

@InProceedings{ginosar2019gestures,
  author={S. Ginosar and A. Bar and G. Kohavi and C. Chan and A. Owens and J. Malik},
  title = {Learning Individual Styles of Conversational Gesture},
  booktitle = {Computer Vision and Pattern Recognition (CVPR)}
  publisher = {IEEE},
  year={2019},
  month=jun
}

About

code for training the models from the paper "Learning Individual Styles of Conversational Gestures"

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages