FlowMDM 调试流程

https://github.com/BarqueroGerman/FlowMDMhttps://github.com/BarqueroGerman/FlowMDM1. 安装 ffmpeg

sudo apt update
sudo apt install ffmpeg

 2.conda 环境

conda env create -f environment.yml
conda activate FlowMDM
python -m spacy download en_core_web_sm  # 报错1
pip install git+https://github.com/openai/CLIP.git  #报错2
pip install git+https://github.com/GuyTevet/smplx.git  #报错3
conda install ffmpeg -y

 报错1、解决：手动下载

1）查看自己 spacy 版本：

pip list

2）进入 https://github.com/explosion/spacy-models/releases

3）找到符合版本的 en_core_web_sm 模型4）下载到本地,安装

pip install en_core_web_sm_PATH

报错2、3：手动下载

1）进入 https://github.com/openai/CLIP ，下载整个项目、解压、进入目录

2)  进入GitHub - GuyTevet/smplx: SMPL-X - an FK only variant，下载整个项目、解压、进入目录

3）使用自己的环境（FlowMDM）两个文件分别执行 下面命令

python setup.py install

3.下载数据集和依赖项及预训练模型

1）数据集：HumanML3D 数据集复制到  ./dataset/HumanML3D 下

HumanML3D 完整数据集下载_(torch_render) roots@roots:~/humanml3d-main$ ./raw-CSDN博客

bable 数据：下载解压安装 ./dataset/babel 下

https://drive.google.com/file/d/1PBlbxawaeFTxtKkKDsoJwQGuDTdp52DD/view
https://drive.google.com/file/d/18a4eRh8mbIFb55FMHlnmI8B8tSTkbp4t/view

 2）依赖：

bash runners/prepare/download_smpl_files.sh
bash runners/prepare/download_glove.sh
bash runners/prepare/download_t2m_evaluators.sh

① download_smpl_files.sh 手动下载，压缩保存至 body_models/ 下 smpl、smplh

".\smplh\smplh\SMPLH_MALE.pkl"  复制到  ".\smplh\SMPLH_MALE.pkl"

https://drive.google.com/uc?id=1INYlGA76ak_cKGzvpOV2Pe6RkYTlXTW2
https://drive.google.com/file/d/1zHTQ1VrVgr-qGl_ahc0UDgHlXgnwx_lM/view

② download_glove.sh 手动下载，压缩 glove/

https://drive.google.com/file/d/1cmXKUT31pqd7_XpJAiWEo1K81TMYHA5n/view?usp=sharing

③ download_t2m_evaluators.sh 手动下载，压缩 t2m/ 、dataset/

https://drive.google.com/file/d/1ZL81tHLaGA3D7ZhLcbc7JKEs40OgzLov/view
https://drive.google.com/file/d/1nNZOSlYxDjyuUHAXzauSWsEFgRi0N5ON/view

④ 模型：下载安装 result/

https://drive.google.com/file/d/1fyx4rd6P_m26Vzb2xEvlkkSw9aDhT8rV/view

4. 运行测试代码：两种数据集 输出结果 result/......... 有 .npy .MP4 结果

python -m runners.generate --model_path ./results/babel/FlowMDM/model001300000.pt --num_repetitions 1 --bpe_denoising_step 60 --guidance_param 1.5 --instructions_file ./runners/jsons/composition_babel.json

python -m runners.generate --model_path ./results/humanml/FlowMDM/model000500000.pt --num_repetitions 1 --bpe_denoising_step 125 --guidance_param 2.5 --instructions_file ./runners/jsons/composition_humanml.json --use_chunked_att

5.运行 smpl 渲染代码，不要改变 第4步文件位置，输出结果在输入地址下 *_smpl_params.npy

python -m runners.render_mesh --input_path /path/to/sample_rep00.mp4