README

Extension of ECF (Evolutionary Computation Framework) for deep neural networks with TensorFlow r1.0 on MS Windows.

REQUIREMENTS

• MS Visual Studio 2015 Update 2 or later versions
• TensorFlow r1.0

Setting up TensorFlow

More information can be found in TensorFlow readme.

Pre-requisites

• Git
• SWIG (recommended version swigwin-3.0.10)
• Python 3 (preferably Anaconda)
• CMake (recommended version 3.6.3)
• if using GPU:
  • NVidia CUDA Toolkit 8.0
  • NVidia CUDNN 5.1

Installing pre-requisites

SWIG

Download the zip archive and extract it to C:\tools\ and add C:\tools\ to PATH environment variable.

NVidia CUDNN

Download the zip archive after installing CUDA Toolkit and extract it to C:\local\. Add C:\local\cuda to your PATH.

Additional notes

Make sure that cmake and git are installed and in your PATH. Check if there's CUDA_PATH in your system variables.

TensorFlow build

• Setup environment.
  • run cmd.exe and the following command:

C:\temp> "C:\Program Files (x86)\Microsoft Visual Studio 14.0\VC\bin\amd64\vcvars64.bat"

• Clone the TensorFlow repository, checkout r1.0 branch and create a working directory for your build:

C:\temp> git clone -b r1.0 --single-branch https://github.com/tensorflow/tensorflow.git
C:\temp> cd tensorflow\tensorflow\contrib\cmake
C:\temp\tensorflow\tensorflow\contrib\cmake> mkdir build
C:\temp\tensorflow\tensorflow\contrib\cmake> cd build
C:\temp\tensorflow\tensorflow\contrib\cmake\build>

• Invoke CMake to create Visual Studio solution and project files.
  • If you are building without GPU support, do not use the last two lines (and remove the ^ character (line continuation) from the last line.

cmake .. -A x64 -DCMAKE_BUILD_TYPE=Release ^
-DSWIG_EXECUTABLE=C:\tools\swigwin-3.0.10\swig.exe ^
-DPYTHON_EXECUTABLE=C:\Users\%USERNAME%\Anaconda3\python.exe ^
-DPYTHON_LIBRARIES=C:\Users\%USERNAME%\Anaconda3\libs\python35.lib ^
-DPYTHON_INCLUDE_DIR=C:\Users\%USERNAME%\Anaconda3\include ^
-DNUMPY_INCLUDE_DIR=C:\Users\%USERNAME%\Anaconda3\Lib\site-packages\numpy\core\include ^
-Dtensorflow_ENABLE_GPU=ON ^
-DCUDNN_HOME=C:\local\cuda\

• Invoke MSBuild to build TensorFlow.
  • Run the following line:

C:\...\build> MSBuild /p:Configuration=Release tf_tutorials_example_trainer.vcxproj

- Be patient, this may take few hours.
- After the build is finished, test it:

C:\...\build> Release\tf_tutorials_example_trainer.exe

Setting up project files

• all include paths in the DeepModelTraining.vcxproj are set over user-defined macro in Visual Studio, so after loading the solution, do the following:

  • select View > Property manager
  • select Release | x64 and right click on property ending with .user, then select Properties
  • under User Macro in common properties, add TENSORFLOW_ROOT, ECF_ROOT and BOOST_ROOT variables with value of location of their root directories
  • make sure you set option Set this macro as an environment variable in the build environment

• building ECF:

  • in ECF_lib.sln, set build configuration to Release x64
  • in solution properties, under C/C++ > Code Generation, set Runtime Library option to Multi-threaded DLL instead of Multi-threaded
  • optional: enable parameterization of multiple algorithms in ECF (used for hybrid approach which combines Backpropagation with other metaheuristic algorithms):
    • add the following lines to State::parseAlgorithmNode method in State.cpp file, just before final return true statement:

#!c++

for (int i = 1; i < n; ++i)
{
    XMLNode tmpNode = node.getChildNode(i);
    registry_->readEntries(tmpNode, tmpNode.getName());
}

• stop TensorFlow log messages (optional):
  • set environment variable TF_CPP_MIN_LOG_LEVEL with value 2 (this disables warnings and debug log)
  • to do it from Visual Studio, add TF_CPP_MIN_LOG_LEVEL=2 to Configuration properties > Debugging > Environment in DeepModelTraining project properties

Release notes

• release 1.0 includes GPU version (r1.0-gpu branch) and non-GPU version (r1.0-no-gpu branch)
• known limitation: PaddedMaxPool layer does not work on non-gpu version because TensorFlow does not support MaxPoolWithArgmax operation without GPU