README.md

Chapter 3: Neural Rendering Toolkit

Start here for the Slang, SlangPy, and PyTorch examples: automatic differentiation, fitting an image, training a digit classifier, and optimizing a sparse voxel scene. No code or output from earlier chapters is needed.

Setup

The chapter uses Python 3.12, an NVIDIA GPU, and CUDA Toolkit 12.9 with a compatible driver. Run the following from Chapter03, preferably in a virtual environment:

python -m pip install torch==2.8.0 torchvision==0.23.0 --index-url https://download.pytorch.org/whl/cu129
python -m pip install slangpy==0.42.0 numpy matplotlib pillow
python -c "import torch; print('CUDA available:', torch.cuda.is_available())"

SlangPy includes the Slang compiler used by the Python examples. Keep the .slang files beside the scripts. The MNIST example explicitly uses the CUDA backend.

Examples

File	What it does
01_autodiff_basics_forward.py	Computes a derivative with forward-mode autodiff.
01_autodiff_basics_backward.py	Computes parameter gradients with reverse-mode autodiff.
02_match_target_image.py	Fits a circle to a generated target image and saves results in results/.
03_train_mnist_slang.py	Trains an MLP in Slang; downloads MNIST to data/ on the first run.
04_spc_neural_optimization_tutorial.ipynb	Uses Kaolin to optimize voxel colors and densities against a target image.

Run a script directly, for example:

python 01_autodiff_basics_forward.py
python 02_match_target_image.py

Voxel notebook

The notebook also needs Kaolin, JupyterLab, and plotting packages. These commands use the same PyTorch/CUDA pairing as above:

python -m pip install kaolin -f https://nvidia-kaolin.s3.us-east-2.amazonaws.com/torch-2.8.0_cu129.html
python -m pip install jupyterlab ipykernel==6.29.5 jupyter-client==7.4.9 plotly tqdm ipywidgets
python -m jupyterlab 04_spc_neural_optimization_tutorial.ipynb

Select the notebook kernel for this environment and run the cells in order. Kaolin's sparse voxel rendering operations require CUDA.