没找到require.txt

暂时没有openai api

This repository contains the code and data for the paper "LGTM: Local-to-Global Text-Driven Human Motion Diffusion Model" (SIGGRAPH 2024).

python -m pip install --user virtualenv # install virtual environment manager
virtualenv --prompt lgtm .env           # create a virtual environment named lgtm
source .env/bin/activate                # activate it

Then install all dependency

pip install -r ./requirements.txt

cd third_package/HumanML3D

follow the third_package/HumanML3D/README.md to prepare original HumanML3D motion data.

Prepare pre-trained LGTM model, TMR encoders and augmented part-level annotations.

bash prepare_data_models.sh

You can generate part-level motion description by yourself. Firstly, you are required to provide OpenAI api key in ./third_packages/TMR/prepare/body_part_annotation_augmentation.py

BASE_URL = "https://api.openai.com/v1",
API_KEY = "",

Here, the API_KEY can get from your OpenAI account followed this document Then run the following commands:

cd third_packages/TMR
python prepare/body_part_annotation_augmentation.py

This script will use original full-body motion annotation for third_packages/TMR/datasets/annotations/humanml3d/annotations.json as input, and generate part-level motion description for each body part. The output will be saved in

• third_packages/TMR/datasets/annotations/humanml3d/body_part_annotations.json
• third_packages/TMR/datasets/annotations/body_part. The former output is group all part annotation in a map, the latter output is save each part annotation in a separate folder for train TMR encoders. You can use the body_part_annotations.json for what you want to do.

The playground.ipynb notebook contains how to use LGTM to generate a motion from text. You can follow the steps in the notebook to generate a motion from text.

python -m lgtm.model.motion_diffusion fit --config configs/lgtm.yaml  --trainer.max_epoch=200