HuwCampbell/grenade

Grenade uses dependent types to encode network architecture and tensor shapes at compile-time, catching shape mismatches before runtime. It provides composable layers (convolution, pooling, LSTM, fully connected, etc.) backed by BLAS/LAPACK for performance, enabling both feedforward and recurrent networks with automatic differentiation for backpropagation. The purely functional design supports complex topologies like residual networks and GANs through type-level composition, where trained networks themselves become reusable layers.

1,453 stars. No commits in the last 6 months.