[openreview]

• Python >= 3.8
• CUDA 11.1
• Pytorch >= 1.9
• Pytorch3D
• trimesh
• tqdm
• scipy
• tensorboard
• PyMCubes

Or you can setup the environment using conda:

conda env create -f environment.yaml
conda activate hiert2s

cd metrics/pvd_metrics/ChamferDistancePytorch/chamfer3D
python setup.py install

• PyTorchEMD

cd metrics/pvd_metrics/PyTorchEMD
python setup.py install
cp build/**/emd_cuda.cpython-38-x86_64-linux-gnu.so .

• pointnet2-ops

pip install git+https://gitee.com/Fukexue/Pointnet2_PyTorch.git@acda965224f35854bc331cd5fe140393216b0a71#subdirectory=pointnet2_ops_lib

We follow the instructions from AutoSDF to preprocess data.

Every script will produce a logger log.txt to record your hyperparameter settings and metrics during evaluating.

# To evaluating models for text-driven generation.
python evaluation_scripts/eval_textdf.py --ckpt {MODEL_PATH}

The whole training has two stages, which means the P-VQ-VAE and discrete diffusion generator are trained seperately.

1. First train the P-VQ-VAE:

./launchers/train_pvqvae_snet.sh

After training, copy the trained P-VQ-VAE checkpoint (pretrained-vqvae-snet.ckpt) to the ./saved_ckpt folder. Based on this quantized feature representation of shape, diffusion models on various tasks can be trained.

2. Then extract the code:

./launchers/extract_pvqvae_snet.sh

3. Train the diffusion models:

# Tarin the discrete diffusion generator for text-driven generation.
./launchers/train_textdf_shapeglot.sh