Graph SDF

This project implement a model to learn SDF from mesh-based geometries using graph neural networks. In particular, we use the encode-process-decode introduced in MeshGraphNet. The encode-process-decode was implemented separately in another repo called graph_networks

Dependencies

• torch >= 1.8 (We use Lazy layers introduced in pytorch since 1.8 version)
• torch-geometric
• tensorflow (no cuda is necessary)
• scipy, matplotlib, numpy,
• json, joblib, tqdm
• trimesh, meshio

as well as graph_networks. By default, the implementation assumes graph_networks folder exists in the parent folder of graph_sdf. Otherwise, you need to pass the address to graph_network parent folder using --extra-path command argument

The example folder should have three json script files

• data_configs.json: containing parameters for converting mesh data to graph data.
• network_configs.json: containing parameters of encode-process-decode network.
• training_configs.json: containing training parameters

The folder script contains a set of such config files. Notice that configs in the script folder assume the code is running on four GPUs. If only a single GPU is available or the course must be run on CPU,

• train_configs["train"]["device"]="cuda" or "cpu" and
• data_configs["train_data"]["dataloader_params"]["device"]=false.

Usage

• train: python run.py -e D:/experiments/graph_sdf_experiments/e1
• train consistency network: python main.py -e D:/experiments/graph_sdf_experiments/e1