Template for Implicit Neural Representations (INR) projects, mostly with the Signed Distance Function (SDF) in mind.
- 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?
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
The scripts can be launched from the main directory with:
python scripts/script.py [--option [VALUE]]