Official Pytorch implementation of the paper "MotionCLIP: Exposing Human Motion Generation to CLIP Space".

Please visit our webpage for more details.

If you find this code useful in your research, please cite:

@article{tevet2022motionclip,
title={MotionCLIP: Exposing Human Motion Generation to CLIP Space},
author={Tevet, Guy and Gordon, Brian and Hertz, Amir and Bermano, Amit H and Cohen-Or, Daniel},
journal={arXiv preprint arXiv:2203.08063},
year={2022}
}

31/AUG/22 - Training loop reproduces paper results. (a bug fix)

11/MAY/22 - First release.

conda env create -f environment.yml
conda activate motionclip

The code was tested on Python 3.8 and PyTorch 1.8.1.

Download and unzip the above datasets and place them correspondingly:

• AMASS -> ./data/amass (Download the SMPL+H version for each dataset separately, please note to download ALL the dataset in AMASS website)
• BABEL -> ./data/babel_v1.0_release
• Rendered AMASS images -> ./data/render

bash prepare/download_smpl_files.sh

This will download the SMPL neutral model from this github repo and additionnal files.

In addition, download the Extended SMPL+H model (used in AMASS project) from MANO, and place it in ./models/smplh.

Process the three datasets into a unified dataset with (text, image, motion) triplets.

To parse acording to the AMASS split (for all applications except action recognition), run:

python -m src.datasets.amass_parser --dataset_name amass

Only if you intend to use Action Recognition, run also:

python -m src.datasets.amass_parser --dataset_name babel

First, download the model and place it at ./exps/paper-model

To reproduce paper results, run:

 python -m src.visualize.text2motion ./exps/paper-model/checkpoint_0100.pth.tar --input_file assets/paper_texts.txt

To run MotionCLIP with your own texts, create a text file, with each line depicts a different text input (see paper_texts.txt as a reference) and point to it with --input_file instead.

To reproduce paper results, run:

 python -m src.visualize.motion_editing ./exps/paper-model/checkpoint_0100.pth.tar --input_file assets/paper_edits.csv

To gain the input motions, we support two modes:

• data - Retrieve motions from train/validation sets, according to their textual label. On it first run, src.visualize.motion_editing generates a file containing a list of all textual labels. You can look it up and choose motions for your own editing.
• text - The inputs are free texts, instead of motions. We use CLIP text encoder to get CLIP representations, perform vector editing, then use MotionCLIP decoder to output the edited motion.

To run MotionCLIP on your own editing, create a csv file, with each line depicts a different edit (see paper_edits.csv as a reference) and point to it with --input_file instead.

To reproduce paper results, run:

 python -m src.visualize.motion_interpolation ./exps/paper-model/checkpoint_0100.pth.tar --input_file assets/paper_interps.csv

To gain the input motions, we use the data mode described earlier.

To run MotionCLIP on your own interpolations, create a csv file, with each line depicts a different interpolation (see paper_interps.csv as a reference) and point to it with --input_file instead.

For action recognition, we use a model trained on text class names. Download and place it at ./exps/classes-model.

python -m src.utils.action_classifier ./exps/classes-model/checkpoint_0200.pth.tar

NOTE (11/MAY/22): The paper model is not perfectly reproduced using this code. We are working to resolve this issue. The trained model checkpoint we provide does reproduce results. (Resolved 31/AUG/22)

To reproduce paper-model run:

python -m src.train.train --clip_text_losses cosine --clip_image_losses cosine --pose_rep rot6d \
--lambda_vel 100 --lambda_rc 100 --lambda_rcxyz 100 \
--jointstype vertices --batch_size 20 --num_frames 60 --num_layers 8 \
--lr 0.0001 --glob --translation --no-vertstrans --latent_dim 512 --num_epochs 100 --snapshot 10 \
--device <GPU DEVICE ID> \
--dataset amass \
--datapath ./data/amass_db/amass_30fps_db.pt \
--folder ./exps/my-paper-model

To reproduce classes-model run:

python -m src.train.train --clip_text_losses cosine --clip_image_losses cosine --pose_rep rot6d \
--lambda_vel 95 --lambda_rc 95 --lambda_rcxyz 95 \
--jointstype vertices --batch_size 20 --num_frames 60 --num_layers 8 \
--lr 0.0001 --glob --translation --no-vertstrans --latent_dim 512 --num_epochs 200 --snapshot 10 \
--device <GPU DEVICE ID> \
--dataset babel \
--datapath ./data/amass_db/babel_30fps_db.pt \
--folder ./exps/my-classes-model

The code of the transformer model and the dataloader are based on ACTOR repository.

This code is distributed under an MIT LICENSE.

Note that our code depends on other libraries, including CLIP, SMPL, SMPL-X, PyTorch3D, and uses datasets which each have their own respective licenses that must also be followed.