Official repository for CVPRW 2024 HuMoGen paper "Exploring Text-to-Motion Generation with Human Preference"

Directory structure

assets/: contains the generated motions from MotionGPT, PPO, and DPO models for the demo.

checkpoints/:

• MotionGPT-base/ - finetuned MotionGPT model.
• dpo/ - our finetuned DPO model
• ppo/ - our finetuned RLHF PPO model
• rm - our finetuned reward model

commands/:

• ppo_train.sh - script for training PPO model with shared base model and separate value and policy heads
• ppo_sep_critic_train.sh - script for training PPO model with separate value and policy models
• dpo_train.sh - script for training DPO model
• rm_train - script for training reward model

preference_data/: preference dataset

MotionGPT/: the MotionGPT codebase with the following changes

• We changed the following files

  • MotionGPT/test.py
  • MotionGPT/mGPT/config.py
  • MotionGPT/mGPT/utils/load_checkpoint.py
  • MotionGPT/requirements.txt
  • MotionGPT/mGPT/archs/mgpt_lm.py

• We added the following files

  • MotionGPT/configs/config_eval_during_training.yaml
  • MotionGPT/generate_npy.py
  • MotionGPT/generate_videos.py

src/:

• models/ - scripts for training and evaluation
• scripts/ - scripts for running experiments
• trainer/ - scripts for training and evaluation

Installations

1. Download the preference dataset from Baidu cloud (code 6gcq) or from Tsinghua Cloud and put it at preference_data/
2. Download our checkpoints from Baidu cloud (code 8ig7) and put them at checkpoints/
3. Download the HumanML3D dataset from https://github.com/EricGuo5513/HumanML3D, preprocess it according to their instructions, and put it under MotionGPT/datasets/
4. Set up environment according to MotionGPT setup instructions below:

conda activate mgpt
conda install pytorch==2.0.0 torchvision==0.15.0 torchaudio==2.0.0 pytorch-cuda=11.7 -c pytorch -c nvidia
cd MotionGPT
pip install -r requirements.txt
python -m spacy download en_core_web_sm
bash prepare/download_smpl_model.sh
bash prepare/prepare_t5.sh
bash prepare/download_t2m_evaluators.sh
bash prepare/download_pretrained_models.sh

5. (optional) set up visualization dependencies. Please refer to MotionGPT for set up instructions.

Training

To train the reward model, modify the hyperparameters and paths in src/scripts/rm_train.sh and run the following command:

bash src/scripts/rm_train.sh

To train the PPO model with shared base model and separate value and policy heads, modify the hyperparameters and paths in src/scripts/ppo_train.sh and run the following command (does not support PEFT):

bash src/scripts/ppo_train.sh

To train the PPO model with separate value and policy models, modify the hyperparameters and paths in src/scripts/ppo_sep_critic_train.sh and run the following command (does not support PEFT):

bash src/scripts/ppo_sep_critic_train.sh

To train the DPO model, modify the hyperparameters and paths in src/scripts/dpo_train.sh and run the following command (does not support distributed training):

bash src/scripts/dpo_train.sh

Evaluation

cd into MotionGPT first.

To evaluate without peft:

python test.py --cfg configs/config_h3d_stage3.yaml --task t2m --checkpoint /path/to/trained_model.pt

To evaluate with peft:

python test.py --cfg configs/config_h3d_stage3.yaml --task t2m --checkpoint /path/to/trained_model.pt --peft --r 8 --lora_alpha 16 --lora_dropout 0.05 

Visualization

To generate npy files for visualization:

python generate_npy.py --cfg configs/config_h3d_stage3.yaml --task t2m --checkpoint /path/to/trained_model.pt --peft --r 8 --lora_alpha 16 --lora_dropout 0.05 

To generate videos for visualization:

python generate_videos.py --data_dir /path/to/generated_npys --video_dir /path/to/generated_videos

Demo

We provide a demo of the generated motions from MotionGPT, PPO, and DPO models (with temperature 1.0). The following table shows the generated motions for the given text instructions.

Text Instruction	MotionGPT Generated Motion	RLHF Generated Motion	DPO Generated Motion
"the individual is shaking their head from side to side"	motiongpt_468.mp4	ppo_468.mp4	dpo_468.mp4
"someone leaps off a concrete block"	motiongpt_3751.mp4	ppo_3751.mp4	dpo_3751.mp4
"a person lifts their arms, widens the space between their legs, and joins their hands together"	motiongpt_4710.mp4	ppo_4710.mp4	dpo_4710.mp4
"he moves his feet back and forth while dancing"	motiongpt_5018.mp4	ppo_5018.mp4	dpo_5018.mp4
"move the body vigorously and then plop down on the ground"	motiongpt_6050.mp4	ppo_6050.mp4	dpo_6050.mp4

Citation

If you find this code useful, please consider citing our paper:

@misc{sheng2024exploring,
      title={Exploring Text-to-Motion Generation with Human Preference}, 
      author={Jenny Sheng and Matthieu Lin and Andrew Zhao and Kevin Pruvost and Yu-Hui Wen and Yangguang Li and Gao Huang and Yong-Jin Liu},
      year={2024},
      eprint={2404.09445},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Acknowledgments

Thank you to MotionGPT authors for providing the codebase and the finetuned model. Our code is partially borrowing from them.

License

This code is distributed under an MIT LICENSE.

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