Skip to content

Latest commit

 

History

History
83 lines (55 loc) · 3.85 KB

README.md

File metadata and controls

83 lines (55 loc) · 3.85 KB
Raw

acLSTM_motion

This folder contains an implementation of acRNN for the CMU motion database written in Pytorch, extended with euler, 6D, and quaternion representations of the frame data.

See the following links for more background:

Auto-Conditioned Recurrent Networks for Extended Complex Human Motion Synthesis

CMU Motion Capture Database

You can find here some pretrained models, and some visualization of the results: Drive Folder

Prequisite

You can find the neccesary libraries in the enviorment.yml and requirments.txt files: If you use conda:

conda env create -f environment.yml

if you have python 3.6

pip install -r requirements.txt

VS-CODE (Racomended)

This project has a reaady to use lauch.json config for VS-CODE, as such you will have all the configurations ready to run on the debug tab of the editor.

Data Preparation

To begin, you need to download the motion data form the CMU motion database in the form of bvh files. Or use the bvh file already present inside the ./train_data_bvh folder (I have already put some sample bvh files including "salsa", "martial" and "indian" in the "train_data_bvh" folder).

Then to transform the bvh files into training data, go to the folder "code" and run generate_training_data.py, it will convert the data and keep 30% of sequencies for testing.

Example:

python3 train_data_bvh/martial/ train_data_xyz/positional --representation=positional

Accpeted representatins are (case sensitive):

  • positional
  • euler
  • 6d
  • quaternions

NOTE: Be sure to run the different script from th parent folder of ./code/ or the program will crash.

Training

After generating the training data, you can start to train the network by running the pytorch_train_aclstm.py.

Example:

python3 ./code/pytorch_train_aclstm.py --representation=positional --dances_folder=train_data_xyz/positional/ --write_weight_folder=run/postitional/weigths/ --write_bvh_motion_folder=runs/positional/bvh/ --dance_frame_rate=120 --batch_size=32 --epochs=50000--representation=positional

--read_weight_path= can be used to resume train from a checkpoint.

Eval

When the training is done, you can use pytorch_test_prediction.py to evaluate the model prediction on x frame.

Example:

python3 ./code/pytorch_test_prediction.py   --dances_folder=train_data_xyz/positional/test/ --read_weight_path=runs/positional/weigths/0025000.weight --out_folder=eval/positional/ --dance_frame_rate=120 --dances_test_size=10 --representation=positional

Note: to keep x low or the script wil crash by beign uable to aquire enough reference frames.

Syntesis

When the training is done, you can use pytorch_test_synthesize_motion.py to synthesize motions.

Example:

python3 ./code/pytorch_test_synthesize_motion.py --dances_folder=train_data_xyz/positional/test/ --read_weight_path=runs/positional/weigths/0025000.weight --write_bvh_motion_folder=eval/positional/bvh/ --dance_frame_rate=120 --dances_test_size=10 -representation=positional

The output motions from the network usually have artifacts of sliding feet and sometimes underneath-ground feet. If you are not satisfied with these details, you can use fix_feet.py to solve it. The algorithm in this code is very simple and you are welcome to write a more complex version that can preserve the kinematics of the human body and share it to us.

For rendering the bvh motion, you can use softwares like MotionBuilder, Maya, 3D max or most easily, use an online BVH renderer for example: http://lo-th.github.io/olympe/BVH_player.html

Enjoy!