contribute.md

Contribute

Requirements for contributing

We use black as our code formatter, with line length 88 (default of black). Please install black and format your files accordingly by running python -m black .. The easiest way is to adapt the settings.json in VSCode to format automatically, with the following settings.json:

{ 
    "editor.formatOnSave": true,
    "[python]": {
        "editor.defaultFormatter": "ms-python.black-formatter",
        "editor.formatOnSave": true,
        "editor.rulers": [
          88
        ],
      },
      "black-formatter.args": [
        "--line-length",
        "88"
    ]
}

How to add a new 3D representation

• Implement dataset class following BaseDataset: The dataset class takes the object of a specific dataset as input, e.g. a ShapenetPreprocessedIterator, which stores the paths to the meshes (and some preprocessed data such as SDFs, pointcloud etc). The dataset class needs to implement the __getitem__ method where the desired files are loaded (e.g. sdf_grid.npz), transformed to the representation, and returned as tensor
• Add representation/dataset config in the configs/data/ folder, which provides the configuration for initializing the dataset. (See for example sdf)
• Add EncoderDecoder class following BaseEncoderDecoder: Most representations are combined with a specific autoencoder (or similar model, e.g. VAE) to encode the representation in a latent vector. The EncoderDecoder class implements the encode and decode methods to transform to the latent and back. It should be able to load a pretrained model when providing the argument ckpt_path.
• Add encoder decoder config This requires two configs: The model specifics (e.g. n_layers) should be defined in a file in configs/net_encode. In addition, you need to add a wrapper file in configs/model: This config file allows hydra to initialize an EncoderDecoderTrainer with your EncoderDecoder model as the net argument.
• Add training config in configs/experiment: Follow the structure in configs/experiment/acc_dualoctree_ae.yaml to specify optimizer, learning rate etc for training the encoder-decoder model on your representation