neuralnet_rs
- 1 Devlogs
- 16 Total hours
A Neural Network library written in Rust from scratch
MNIST digit classification network working with around 97.5% accuracy with 2 hidden layers (128x LeakyRelu + 64x Tanh). Training (10 epochs, batch size=64 on full training set) takes around 8.5s on single core (parallel training is not yet supported) on my Ryzen 7 2700 thanks to allocation avoidance, matrix architecture and Rust optimizations.
Saving and loading networks to custom MessagePack .nnrs format is supported as well as export to ONNX, which allows running trained models on production runtimes.
The project has been converted to library and unit tests were added - currently everything is successfully passing.
TODO:
- optimizers
- dropout
- multithreaded training
- possibly further optimizations
