mapnn is designed to combine the strengths of different high-performance neutral network, such as ncnn, MNN or Tengine. With this framework, one can easily test the performance of each kernel from different inference frameworks. Of cause, this framework provide a simple method that map some operator(defined by training) to kernel(defined by inference). So that user given the ability to choose the best kernel for inference task throught the map.

• caffe: alexnet, googlenet inceptionv3/v4 mobilenetv1/v2, resnet, vgg16
• onnx : alexnet, googlenet inceptionv1/v2 mobilenetv2, resnet, vgg16, shufflenet1.1, yolov2

System	armv7	armv8	x86	amd64
Ubuntu(GCC)	—	—	—
Ubuntu(Clang)	—	—	—
Linux			—	—
Windows(MSVC)	—	—
Android
MacOS(Clang)	—	—	—

• Build for Linux-x86-64
• Build for Linux-armv7l
• Build for Linux-armv7hl
• Build for android-v7a
• Build for android-v8a

According to the step 4, Compiling and install this library with your toolchain. And copy(if need) the layer_test to your platform. Then run ther layer_test to get perf.txt.

./layer_test conv -h
# Usage: layer_test conv -k 3 -s 3 -c 1 -g 2
# Option:
#       -k kernel INT
#       -s stride INT
#       -c cycle  INT
#       -g gap    INT
#       -o output INT
#       -h this help

The following show some kernels benchmark(up picture) and corresponding accuracy(down picture).

• raspebery pi3B+ (right for conv3x3s1, left for conv3x3s2)

• SDM855 arm32 (right for conv3x3s1, left for conv3x3s2)

• SDM855 arm64 (right for conv3x3s1, left for conv3x3s2)

• intel 10700k x86-64 (right for conv3x3s1, left for conv3x3s2)

node: This benchmark not means benchmark of different CNN frameworks.

int ret; 
mapnn::Net* net = new mapnn::Net(); // new object
ret = net->load("model_path.onnx"); // load model
//ret = net->load("model_path.proto", "model_path.caffemodel");
ret = net->prepare(3, 224, 224);    // prepare net
ret = net->inference(float_data, 3, 224, 224); // inference
Tensor& output = net->getTensor("output_name");
delete net;

• Add new operation from other training frameworks.
• Add new kernel from other inference frameworks.
• Improve frameworks and do good pull Request.
• Fix and Report the issue on the Github issues page.
• Star or fork this project.

• notice A: if you contribute a good code, please append(or apply for new code) the follow boilerplate notice to your code: Copyright [yyyy] [name of copyright owner]

• notice B: mapnn include some third party libraries as following:

  • 3rdparty/flatbuffers
  • 3rdparty/MNN
  • 3rdparty/ncnn
  • 3rdparty/protobuf
  • 3rdparty/stb
  • 3rdparty/Tengine