Mint is a single-file header only library for tensor manipulation. It also enables importing and executing some of neural net models. Mint aims to be dependency-free (except for C standard lib) and easily distributed. However, it is possible to integrate with the other libraries such as BLAS if needed.
Some of notable features:
- NumPy style broadcasting
- BLAS backend (optional)
- ARM NEON SIMD acceleration (enable with
#define MT_USE_NEON).- For Apple Silicon devices, you can additionally define
#define MT_USE_APPLE_ACCELERATEto even more performance. Specify Accelerate framework during compilation.
- For Apple Silicon devices, you can additionally define
- Got a
layer kind xxx is not supported yeterror? Try runningonnxsim. It might trim those pesky unsupported layers - If onnxsim still gives you unsupported layers: just open an issue and I'll try to implement it when I can
¯\_(ツ)_/¯
The torchvision models are dumped into ONNX, then converted to Mint model format for inference.
- AlexNet
- VGG-19
- ResNet-18
All models here with opset 8 work well