Loading Pytorch Model in C++

Implementation of DNN face landmark detector from python by https://github.com/1adrianb/face-alignment

Please Refer to here for more details.

Prerequisite

• Visual studio 2015/17
• Libtorch 1.0.0
• dlib 19.16
• Opencv 2/3 --

./face-alignment/ folder contains the base python for the DNN implementation
./cpp_app/ folder contains example of loading the torch script in C++

Tracing Pytorch model for C++ API

1. To convert the pytorch network model for C++ use, the model must be traced. There is two ways to convert the model into torch script.

2. Tracing

   • Fast, may not be able to handle complex control flow

3. Annotation

   • Slow (can be very slow), able to handle complex control flow Please refer to Pytorch C++ documentation for more detail explanation

4. Serialize the model with it's weight

   • Save the model in file

5. Load it in C++ application

   • For the landmark detector, some pre-processing is done using dlib and pytorch.
   • In the C++ implementation, these are re-implemented using dlib for C++ and libtorch. (The result might not be 100% same with the python version)

Including Libtorch (Pytorch C++) in visual studio

Using prebuild library for vs2015/17 x64 release only. Please build from source for different build. Include Directories

• $(ProjectDir)include\dlib
• $(ProjectDir)include\libtorch
• $(ProjectDir)include\libtorch\torch\csrc\api\include
• $(ProjectDir)include\opencv

Library Directories

• $(ProjectDir)lib\dlib
• $(ProjectDir)lib\libtorch
• $(ProjectDir)lib\opencv

Linker Input

• Add to vs project 'linker->input->additional dependencies'
• Copy the following

dlib19.16.0_release_64bit_msvc1900.lib
torch.lib
caffe2.lib
libprotobuf.lib
c10.lib

Adding dlib source to visual studio

From vs solution explorer->properties->add existing item Add '/include/dlib/dlib/all/source.cpp'

Including all the required DLL files

Copy dll from x64->Release

torch.dll
caffe2.dll
libiomp5md.dll
c10.dll
(opencv).dll