ntalabot/base-inr

Template for Implicit Neural Representations projects.

Implicit Neural Representations

Template for Implicit Neural Representations (INR) projects, mostly with the Signed Distance Function (SDF) in mind.

TODO

• Update models to take (latent, xyz) without repeating the latents
• Look into IoU metric and Winding-Number (libigl)
• Speed up reconstruction function
• Add device instead of defaulting to CUDA
• Make clamping part of SDF loss (make sure when it applies)?
• Restructure training (and inference/eval?) as Trainer classes
• Docs
  • Expected data structure
• Make more general to arbitrary input dimension (e.g., to work in 3D and 2D)
• Rework deepsdf network classes into a single main one?

File structure

base-inr/
├── experiments/        <- Experimental config and results (untracked by default)
│   └── template/
│       └── specs.json  <- Template experimental config
├── notebooks/          <- Jupyter notebooks for testing code and visualizing results
├── scripts/            <- Python scripts
└── src/                <- Source code

How to use

The scripts can be launched from the main directory with:

python scripts/script.py [--option [VALUE]]