jisuchoi01 / connx Goto Github PK

View Code? Open in Web Editor NEW

This project forked from tsnlab/connx

C implementation of Open Neural Network Exchange Runtime

License: GNU General Public License v3.0

Python 6.59% Makefile 0.33% Shell 4.04% C 87.72% CMake 1.32%

connx's Introduction

CONNX

C implementation of Open Neural Network Exchange Runtime

Requirements

python3 # to run examples and test cases
cmake >= 3.1
ninja-build

Quick start

Compile in release mode

ports/linux$ mkdir build              # Make build directory
ports/linux/build$ cmake .. -G Ninja  # Generate build files with "Release" mode
ports/linux/build$ ninja              # Compile

You can find 'connx' executable in ports/linux/build directory.

Compile with sub-opset (optional)

If you want to compile CONNX with subset of operators, in case of inferencing MNIST only, just make ports/linux/opset.txt file as below. And just follow the compile process.

Add Conv MatMul MaxPool Relu Reshape

Run examples

Run MNIST example. (Mobilenet and YOLO will be coming soon)

ports/linux/build$ ninja mnist

Notice: If you want to run on Raspberry Pi 3, please compile with Release mode(CMAKE_BUILD_TYPE=Release) for sanitizer makes some problem.

ONNX compatibility test

Run the test cases in 'test' directory which contains converted test cases from ONNX. python3 with Numpy is required.

ports/linux/build$ ninja onnx

Performance profile report

If you want to profile performance, you need to compile CONNX in debugging mode first.

ports/linux/build$ cmake .. -G Ninja -DCMAKE_BUILD_TYPE=Debug  # Generate build files
ports/linux/build$ ninja                                       # Compile
ports/linux/build$ ninja mnist                                 # Run an any example
ports/linux/build$ ninja prof                                  # Print performance profile report