TensorFlow training in C++

Create a DNN model for MNIST using Keras and train it using TensorFlow in C++.

Requirement

• Python 3.6
• keras 2.2.4
• tensorflow 1.10.0
• Visual Studio 2015

Setup

• Download MNIST Data from here and extract them into "MNIST_data" folder.

[MNIST_data]
  +- train-images.idx3-ubyte
  +- train-labels.idx1-ubyte
  +- t10k-images.idx3-ubyte
  +- t10k-labels.idx1-ubyte

• Download tensorflow 1.10.0 for C++ from here and unzip it into a directory.
• Set the path to the environment variable "TF_ROOT".

set TF_ROOT={PATH_TO_TENSOR_FLOW_DIR}\libtensorflow-cpu-windows-x86_64-1.10.0-avx2

• Copy "tensorflow.dll" to this directory.

Create graph

• Install TensorFlow 1.10.0 and Keras.
• Execute the following command.

python create_graph.py

Training

Build C++ module

Open "Visual Studio Developper Command Prompt" and execute the following command.

cl /EHsc /I %TF_ROOT%/include \
  /I %TF_ROOT%/include/tensorflow/contrib/cmake/build \
  /I %TF_ROOT%/include/tensorflow/contrib/cmake/build/protobuf/src/protobuf/src \
  /I %TF_ROOT%/include/tensorflow/contrib/cmake/build/eigen/src/eigen \
  /I %TF_ROOT%/include/tensorflow/contrib/cmake/build/external/nsync/public \
  /D "COMPILER_MSVC" /D "WIN32" /D "NOMINMAX" /GS /GL /source-charset:utf-8 \
  train_and_predict.cpp common.cpp %TF_ROOT%/lib/tensorflow.lib

Run

train_and_predict.exe

Freeze Graph

python freeze_graph.py

Inference

Build C++ module

Open "Visual Studio Developper Command Prompt" and execute the following command.

cl /EHsc /I %TF_ROOT%/include \
  /I %TF_ROOT%/include/tensorflow/contrib/cmake/build \
  /I %TF_ROOT%/include/tensorflow/contrib/cmake/build/protobuf/src/protobuf/src \
  /I %TF_ROOT%/include/tensorflow/contrib/cmake/build/eigen/src/eigen \
  /I %TF_ROOT%/include/tensorflow/contrib/cmake/build/external/nsync/public \
  /D "COMPILER_MSVC" /D "WIN32" /D "NOMINMAX" /GS /GL /source-charset:utf-8 \
  predict.cpp common.cpp %TF_ROOT%/lib/tensorflow.lib

Run

predict.exe