Skip to content

Latest commit

 

History

History
36 lines (32 loc) · 1.97 KB

install.md

File metadata and controls

36 lines (32 loc) · 1.97 KB
Raw

Environment setup

  1. We have tested our code with this docker environment pytorch/pytorch:1.13.0-cuda11.6-cudnn8-devel and NVIDIA V100 GPUs.
  2. Install PyTorch: pytorch==1.13.0 torchvision==0.14.0 torchaudio==0.13.0
  3. Install other dependencies:
# install libraries
apt-get install -y \
    libglfw3-dev \
    libgles2-mesa-dev \
    libglib2.0-0 \
    libosmesa6-dev \
# install requirements
pip install -r requirements.txt
# install kaolin
pip install kaolin==0.13.0 -f https://nvidia-kaolin.s3.us-east-2.amazonaws.com/torch-${YOUR_TORCH_VERSION}_${YOUR_CUDA_VERSION}.html
  1. Git clone both TeCH and thirdparties
git clone --recurse-submodules git@github.com:huangyangyi/TeCH.git
  1. Build modules
cd core/lib/freqencoder
python setup.py install
cd ../gridencoder
python setup.py install
cd ../../
  1. Download necessary data for body models: bash scripts/download_body_data.sh. If Username/Password Authentication Failed, you need firstly register at ICON and choose "Yes" for all the projects listed in "Register for other projects".
  2. Download pretrained models of MODNet: bash scripits/download_modnets.sh
  3. Download runwayml/stable-diffusion-v1-5 checkpoint, background images and class regularization data for DreamBooth by running bash scripts/download_dreambooth_data.sh, you can also try using another version of SD model, or use other images of man and woman for regularization (We simply generates these data with the SD model).

If you still struggle with the package conflict, YuliangXiu/TeCH shows a cleaner version to setup TeCH at Ubuntu 22.04.3 LTS, NVIDIA A100 (80G), CUDA=11.7 with anaconda. This forked repository has removed almost all the package version requirements, see requirements.txt.