ncnn-fortran

Call ncnn from Fortran via mix compiling.

Introduction

How this demo works?

1. Image

   As Fortran is mainly used in scientific computation, there seems to be no workable OpenCV binding for Fortran or other computer vision libraries. The repo ncnn_on_xr806 inspired me for the image was stored in a header file in the format of const unsigned char array and ncnn::Mat::from_pixels is ready for it. In this repo the image is stored in a ASCII-text file with 28 rows and 28 columns, it should be in csv or tsv like format. Here are some so-called images in image/ directory.

2. Main Program

   The main program main.f90 is written in Fortran, it parses command line arguments, loads image, call ncnn binding, and output the most possible label of input image.

3. ncnn Binding

   ncnn Binding inclueds 4 files:

   • ncnn_mnist.f90
   • predict.h
   • predict.cpp
   • mnist-12-bin.h

   ncnn_mnist.f90 is the wrapper of function predict, it interacts with predict.cpp via extern "C" in predict.h. mnist-12-bin.h is the model waits ncnn to load. In predict.cpp, function preduct initializes an ncnn::Mat from a 2d Fortran array, loads the model, and predicts the label. Finally, the predicted label is passed to main program.

Run

Modify the CMakeLists.txt, make sure all paths of ncnn is correct for you.

Then

$ mkdir build
$ cmake ..
$ make
$ cd ..
$ ./build/mnist image/4.dat
 Predict:            4

GNU and Intel compilers have been tested.

References

• ncnn
• ncnn_on_xr806
• MNIST - Handwritten Digit Recognition