Bender

Why we included this project

Bender is a Swift layer over Apple's MetalPerformanceShaders for running pretrained neural networks on iOS. Instead of hand-wiring raw MPSCNN kernels, you assemble a network from standard building blocks like convolution, pooling, fully connected, and normalization layers, or point it at an exported TensorFlow graph and have it translated into Bender layers automatically. That import path avoids bundling TensorFlow's static library and gives you GPU access on iOS, which the TensorFlow build for that platform didn't support. The API chains layer operators in a readable way, and the docs list which TensorFlow ops are supported and how they map to Bender layers, so you can check your model before committing. Small teams doing on-device image recognition or other inference without a heavyweight runtime will find it practical, as long as their models stick to the supported layer set.