Dlib to tensorflow/keras converter

Dlib face recognition network is a neural network trained by Davis King (https://github.com/davisking/dlib) using his C++ library/toolkit.

Please see more details about this network here at this link - https://github.com/davisking/dlib-models

My main driver for this work was to better understand the architecture of this network as well as the dlib library.

Objective

This repository provides set of scripts to convert dlib's face recognition network into other formats such as:

• Keras hd5
• Tensorflow's saved model and frozen graph
• ONNX [TODO]

Usage and how it works

Step 1

The weights for the model provided by dlib are in stored in the binary fomat (.dat file).

Here is the location from where you can download it - http://dlib.net/files/dlib_face_recognition_resnet_model_v1.dat.bz2

Step 2

Dlib toolkit provides a method to take the serialized weights file and convert to an XML file. This is done using C++ so I am providing a tool called xml_generator. All it really does is that it defines the network in C++ (following the example from dlib), loads the weights and then serialize it.

In order to use this tool you should have following installed on your machine -

• dlib (and required dependencies)
• cmake

cd xml_generator
mkdir build
cd build
cmake -DCMAKE_BUILD_TYPE=Release ..
cmake --build . --config Release

If you are able to compile this tool xml_generator then you would have build/bin directory with the executable.

# now simply do
# [Make sure you have dlib_face_recognition_resnet_model_v1.dat in build directory]
./bin/xml_generator

You should now have dlib_face_recognition_resnet_model_v1.xml file in the build directory.

Step 3

Time for the final convertion.

# install necessary pip packages
pip install tensorflow
pip install numpy
pip install xmltodict

# It's as simple as 
# [Make sure you are at the root of this repository]
python main.py --xml-weights xml_generator/build/dlib_face_recognition_resnet_model_v1.xml

At the root of the repository you should have -

• Keras format - dlib_face_recognition_resnet_model_v1.h5

• A folder named exported where saved model is saved !

• Frozen graph - dlib_face_recognition_resnet_model_v1.pb

Analysis

There is also a juypter notebook dlib_analysis.ipynb that provides :

• Analysis of the network

• Description of the keras layers/summary/weight etc

• Performs inference using the converted keras model

Acknowledgements

The main person to thank here is Davis King, the creator of dlib [https://github.com/davisking/dlib] and trainer of this recognition network. I discovered his repository last year and am extemely impressed by his work. Not only the quality of what he produces is at the highest level but the regular support he provides to the users of his library is commendable. Thanks Davis!

Another person I want to thank is Vincent Muhler (awesomedudethathacks) whose work on face-api.js showed me that it was possible to extract weights and reuse them. In the end it was just the grunt work but doable and the fact that someone had done it already made sure that it was possible. Thanks Vincent for leading the way !