TetGAN [Project Page]

This repository contains code from the paper TetGAN: A Convolutional Neural Network for Tetrahedral Mesh Generation. The proposed neural network layers learn deep features over each tetrahedron and learn to extract patterns within spatial regions across multiple scales. We illustrate the capabilities of our technique to encode tetrahedral meshes into a semantically meaningful latent-space which can be used for shape editing and synthesis.

Installation

Install all dependencies using anaconda and the env.yml file provided.

conda env create -f env.yml
conda activate tetGAN

Data

To train TetGAN, we need files specifying the meshes in the dataset and an initial tetrahedral grid.

Tetrahedral Grid Files

We generate tetrahedral grids by passing a unit cube (with corners at (0, 0, 0) and (1, 1, 1)) to quartet. We have provided examples of these grids in tetGAN/grids. Each file is named cube_{p}.tet where p specifies a resolution parameter passed to quartet.

Mesh Data

There are two options regarding mesh data:

You may download our preprocessed data (coming soon)
You may provide a directory of mesh files to our preprocessing script.

Preprocessed Data

We provide preprocessed data on various ShapeNet categories. Upload coming soon.

Build Your Own Data

You may build your own dataset through the following:

Create a directory containing all of the training meshes.

 training_data
 |- 000.obj
 |- 001.obj
 |- 002.obj
 |- ...

Pass this directory to the main script in src/data.py, also specifying an initial grid file and a subdivision depth.

IMPORTANT NOTE: The routine we use to compute occupancies requires that input meshes be watertight and manifold. We use the Watertight Manifold repository to preprocess ShapeNet objects. The accuracy of the occupancy labels is not guaranteed for non-manifold shapes.

The processed data comes in the form of a folder containing 3 pickled tensors per mesh file.

    processed_data
    |- 000
    |-|- occ.pt
    |-|- def.pt
    |-|- def_c.pt
    |- ...

Training

To train a model, use src/main.py which accepts a config file as well as command-line arguments. We have provided an example config used to produce our paper results: configs/example_config.yml. Any arguments specified by the config file will be overriden by command-line arguments.

# Train with example config on processed data
python src/main.py --config configs/example_config.yml --dataset_path ./data/processed_data

# Train with example config, but change the number of epochs and batch size
python src/main.py --config configs/example_config.yml --num_epochs 200 --batch_size 15

# Train with example config, but alter the training resolution by changing the initial grid and the subdivision depth
python src/main.py --config configs/example_config.yml --initial_grid ./grids/cube_0.25.tet --subdivision_depth 4

Inference and Mesh Extraction

Through our script

We src/predict.py to sample shapes from a model checkpoint. This script accepts the training config file as input in order to instantiate a correct version of the network. See the following example usage:

# Use the training config as input
# Sample 20 shapes, using GPU, from the checkpoint ../logs/checkpoints/gen_epoch_100.ckpt
# Write the shapes as tetrahedral meshes
python predict.py --training_config ../configs/example_config.yml --gpu \ 
    --checkpoint ../logs/checkpoints/gen_epoch_100.ckpt --num_samples 20 \
    --out_dir ../logs/sampled_meshes --mesh_type tetrahedral --laplacian_smoothing 1

Uploaded checkpoints of trained models from the paper are coming soon (with the processed data).

With your own code

If you wish to write your own inference code, here we provide an example code snippet of these possible inference operations.

# The variable vae is a trained network
# Sample a latent code and decode it

dist = torch.distributions.normal.Normal(0.0, 1.0)
sample = dist.sample(sample_shape=(1, cfg['code_size']))
decoded = vae.decode(sample)

# Encode and decode a shape
# The variable feat are the occupancies and centroid deformations computed for the mesh

encoded = vae.encode(feat.unsqueeze(0))[0]
decoded = vae.decode(encoded)

# You may now use the encoding for latent operations such as arithmetic/interpolation

encoded_1 = vae.encode(feat_1.unsqueeze(0))[0]
decoded = vae.decode((encoded + encoded_1) / 2)

To extract a mesh from network output, refer to the function extract_mesh in src/nets.py. Calling this function with a network and some output generates either a TriangleMesh or a TetMesh object which then may be written to file through their respective I/O functions.

# Extract mesh from network output
extracted_surface = vae.extract_mesh(decoded, 'triangle', dataset.deformation_scalar, smoothing_iterations=2)
extracted_volume = vae.extract_mesh(decoded, 'tetrahedral', dataset.deformation_scakar, smoothing_iterations=2)

Citation

@inproceedings{tetGAN,
    title = {TetGAN: A Convolutional Neural Network for Tetrahedral Mesh Generation},
    author = {Gao, Wang, Metzer, Yeh, Hanocka},
    booktitle = {Proceedings British Machine Vision Conference (BMVC)},
    year = {2022}
}

Questions / Issues

If you have questions or issues regarding this code, please open an issue.

threedle/tetGAN