MotionGPT: Finetuned LLMs are General-Purpose Motion Generators

The official PyTorch implementation of the paper "MotionGPT: Finetuned LLMs are General-Purpose Motion Generators".

Please visit our Project Page for more details.

If you find MotionGPT useful for your work please cite:

@article{zhang2023motiongpt,
  title={MotionGPT: Finetuned LLMs are General-Purpose Motion Generators},
  author={Zhang, Yaqi and Huang, Di and Liu, Bin and Tang, Shixiang and Lu, Yan and Chen, Lu and Bai, Lei and Chu, Qi and Yu, Nenghai and Ouyang, Wanli},
  journal={arXiv preprint arXiv:2306.10900},
  year={2023}
}

Table of Content

• Installation
• Demo
• Train
• Evaluation
• Visualization
• Acknowledgement

Installation

1. Environment

conda env create -f environment.yml
conda activate motiongpt

2. Dependencies

For text to motion evaluation

bash prepare/download_evaluators.sh
bash prepare/download_glove.sh

For SMPL mesh rendering

bash prepare/download_smpl.sh

For using the LLaMa model weight, follow pyllama to download the original LLaMA model, and then follow Lit-LLaMA to convert the weights to the Lit-LLaMA format. After this process, please move the lit-llama/ directory under the checkpoints/ directory.

Once downloaded, you should have a folder like this:

MotionGPT
├── checkpoints
│   ├── kit
│   │   ├── Comp_v6_KLD005
│   │   ├── Decomp_SP001_SM001_H512
│   │   ├── length_est_bigru
│   │   ├── text_mot_match
│   │   └── VQVAEV3_CB1024_CMT_H1024_NRES3
│   ├── lit-llama
│   │   ├── 7B
│   │   │   └── lit-llama.pth
│   │   ├── 13B
│   │   └── tokenizer.model
│   └── t2m
│       ├── Comp_v6_KLD005
│       ├── M2T_EL4_DL4_NH8_PS
│       ├── T2M_Seq2Seq_NML1_Ear_SME0_N
│       ├── text_mot_match
│       └── VQVAEV3_CB1024_CMT_H1024_NRES3
├── body_models
│   └── smpl
│       ├── J_regressor_extra.npy
│       ├── kintree_table.pkl
│       ├── smplfaces.npy
│       └── SMPL_NEUTRAL.pkl
└── glove
    ├── our_vab_data.npy
    ├── our_vab_idx.pkl
    └── our_vab_words.pkl

3. Pretrained Models

For pretrained VQ-VAE models

bash prepare/download_vqvae.sh

For finetuned LLaMA model

bash prepare/download_lora.sh

Once downloaded, you should have a folder like this:

MotionGPT/checkpoints
├── pretrained_vqvae
│   ├── kit.pth
│   └── t2m.pth
└── pretrained_lora
    └── pretrained.pth

4. Dataset

Please follow HumanML3D to download HumanML3D and KIT-ML datasets and put them under the directory dataset like:

MotionGPT/dataset
├── HumanML3D
└── KIT-ML

To prepare the dataset used for finetuning LLaMA, please follow the instructions below (take HumanML3D as an example)

# Encode the motions to tokens by pretrianed VQ-VAE and save the token sequence results under `./dataset/HumanML3D/VQVAE/`
# For pretrained VQ-VAE, you can use the model provided or train the model by yourself following the training instruction.
python scripts/prepare_data.py --dataname t2m

# Generate the dataset on train split and validation split in the format of {instruction, input, output}
# Results saved as `./data/train.json` and `./data/val.json`
python scripts/generate_dataset.py --dataname t2m

# Generate corresponding instruction tuning dataset
# Results saved as `./data/train.pt` and `./data/val.pt`
python scripts/prepare_motion.py --dataname t2m

Demo

Give task description (--prompt) and conditions (--input) to generate corresponding motion. The motion in npy format (demo.npy) and skeleton visualization result (demo.gif) will be saved under {output_dir}.

Please set --render if you want to render SMPL mesh.

# text-to-motion
python generate_motion.py --prompt "Generate a sequence of motion tokens matching the following human motion description." --input "a person walks forward." --lora_path ./checkpoints/pretrained_lora/pretrained.pth --out_dir {output_dir} --render

# (text, init pose)-to-motion
python generate_motion.py --prompt "Generate a sequence of motion tokens matching the following human motion description given the initial token." --input "a person walks forward.<Motion Token>315</Motion Token>" --lora_path ./checkpoints/pretrained_lora/pretrained.pth --out_dir {output_dir} --render

# (text, last pose)-to-motion
python generate_motion.py --prompt "Generate a sequence of motion tokens matching the following human motion description given the last token." --input "a person walks forward.<Motion Token>406</Motion Token>" --lora_path ./checkpoints/pretrained_lora/pretrained.pth --out_dir {output_dir} --render

# (text, key poses)-to-motion
python generate_motion.py --prompt "Generate a sequence of motion tokens matching the following human motion description given several key tokens." --input "a person walks forward.<Motion Token>315,91,406</Motion Token>" --lora_path ./checkpoints/pretrained_lora/pretrained.pth --out_dir {output_dir} --render

Train

For VQ-VAE training

python train_vqvae.py --out_dir {output_dir} --dataname t2m

For finetuning LLaMA with LoRA

python finetune_motion.py --out_dir {output_dir} --dataname t2m

Evaluation

For VQ-VAE

python eval_vqvae.py --out_dir {output_dir} --resume_pth {vqvae_model_path} --dataname t2m

For LLaMA

python eval.py --vqvae_pth {vqvae_model_path} --lora_path {fintuned_model_path} --out_dir {output_dir} --dataname t2m

Visualization

The generated poses are all saved in npy format with the shape of [seq_len, joint_num, 3]

The output results are saved under the same directory with the corresponding filename in gif format

For visualization in skeleton format

# To visualize all the poses saved in {saved_pose_dir}
python visualization/plot_3d_global.py --dir {saved_pose_dir}

# To visualize selected poses in {saved_pose_dir}
python visualization/plot_3d_global.py --dir {saved_pose_dir} --motion-list {fname1} {fname2} ...

For SMPL mesh rendering

# To visualize all the poses saved in {saved_pose_dir}
python visualization/render.py --dir {saved_pose_dir}

# To visualize selected poses in {saved_pose_dir}
python visualization/render.py --dir {saved_pose_dir} --motion-list {fname1} {fname2} ...

Acknowledgement

Thanks to HumanML3D, T2M-GPT and Lit-LLaMA, our code is partially borrowing from them.