HS-PAC

Preparation

Configure the conf file properly

Dataset

dataset_freihand: FreiHand loader ported from mobrecon

Files

train.py is the training script without acceleration trainer.py is a refactored trainer with improved extensibility src/utils/data_keys.py manages all data loading and access keys in one place to avoid confusion

New Features

Progress bar now adapts to window size changes, preventing screen flooding

Environment Setup

First install PyTorch. CUDAGraph acceleration requires pytorch>=2.0

conda create -n yourname python=3.8
conda activate yourname
pip install torch==2.4.1 torchvision==0.19.1 --index-url https://download.pytorch.org/whl/cu121

Reference versions: torch 2.4.1 torchvideotransforms 0.1.2 torchvision 0.19.1

pip install git+https://github.com/hassony2/torch_videovision
pip install torch-geometric -f https://data.pyg.org/whl/torch-2.4.1+cu121.html

pip install attrs brotlipy certifi chumpy cycler fonttools fvcore h5py imageio iniconfig iopath Jinja2 joblib kiwisolver MarkupSafe matplotlib mkl-fft mkl-service networkx numpy opencv-python openmesh packaging pandas Pillow pluggy portalocker protobuf pycocotools PyOpenGL pyparsing pytest python-dateutil pytz PyWavelets PyYAML pyzmq scikit-image scikit-learn scipy tabulate tensorboardX termcolor threadpoolctl tifffile tomli tqdm transforms3d trimesh vctoolkit vctools yacs open3d
pip install torch-cluster torch-geometric torch-scatter torch-sparse torch-spline-conv # These install very slowly as they need compilation
# If installation fails, try the following with specified torch version
pip install torch-scatter torch-sparse torch-cluster torch-spline-conv -f https://data.pyg.org/whl/torch-2.4.1+cu121.html

#ln -s data/model/MANO_RIGHT.pkl template/MANO_RIGHT.pkl

Install MPI-IS Mesh from source Have AI generate Linux-specific commands for installation, as the configuration steps in the official repository are somewhat confusing. You can first try the commands below to see if they work successfully

conda install -c conda-forge boost pyopengl
git clone https://github.com/MPI-IS/mesh.git # Don't forget to change directory
cd mesh

sudo apt-get update
sudo apt-get install -y build-essential cmake libboost-all-dev libeigen3-dev

The following commands need to be copied and executed at once

export BOOST_INCLUDE_DIRS=$CONDA_PREFIX/include
export CPLUS_INCLUDE_PATH=$CONDA_PREFIX/include:$CPLUS_INCLUDE_PATH
pip install --verbose --no-deps --no-cache-dir .

Verify successful installation

python -c "import psbody.mesh; print('Installation successful!')"

Recommended: Create a dedicated environment for viewing tensorboard files

conda create -n tb python=3.8
conda activate tb  # Replace 'tb' with your actual environment name
pip install tensorboard
tensorboard --version
CUDA_VISIBLE_DEVICES="" tensorboard --logdir=/

Environment configuration has been verified

For additional setup, refer to:

https://github.com/lixiny/manotorch pip install git+https://github.com/mattloper/chumpy pip install pyrender