itkTensorFlow

ITK pipeline using provided tensorflow .pd model.
Heavily based on Neargye/hello_tf_c_api Feel free to contact me if you need new itkTensorFlow features, for example multiple inputs/outputs.

System	Badge
OSX/Linux build - Travis
Windows build - AppVeyor
Downloads

How to use the executable

You can download the library and the executable from here. Options available for the executable execution:

>> itkTensorFlowExe -h
Use: itkTensorFlowExe
	 -r inputRgbFilePath
	 -g inputGrayFilePaths, can be used multiple times
	 -m inputGraphPath tensor flow model/graph (in .pb format)
	 -o outputDirPath
	 -s scale (multiply output values by this value)
	 -h help

Example 1: one rgb image as input

If you want to use one rgb image (Windows command line):

itkTensorFlowExe.exe -r "image2.jpg" -m "path/to/model.pb" -o "output"

OSX/Linux command line:

./itkTensorFlowExe -r "image2.jpg" -m "path/to/model.pb" -o "output" 

Example 2: several grayscale images as input

If you want to use one or more grayscale images (Windows command line):

itkTensorFlowExe.exe ^
                    -g "image0.dcm" ^
                    -g "image1.dcm" ^
                    -g "image2.dcm" ^
                    -m "path/to/model.pb" ^
                    -o "output"

OSX/Linux command line:

./itkTensorFlowExe \
                    -g "image0.dcm" \
                    -g "image1.dcm" \
                    -g "image2.dcm" \
                    -m "path/to/model.pb"  \
                    -o "output"

You can find examples of using executable here: https://github.com/MRKonrad/itkTensorFlow/tree/master/scripts

How to compile

OSX/Linux

git clone https://github.com/MRKonrad/itkTensorFlow.git
cd itkTensorflow
git submodule update --init --recursive
cmake . -Bbin -DITK_DIR_HINTS="../ITK_install"
cmake --build bin

Windows

git clone https://github.com/MRKonrad/itkTensorFlow.git
cd itkTensorflow
git submodule update --init --recursive
cmake . -Bbin -G"Visual Studio 15 2017 Win64" -DCMAKE_INSTALL_PREFIX=install -DITK_DIR_HINTS="../ITK_install"
cmake --build bin --config RELSEASE --target INSTALL

Models/Graphs tested

The sample model used in tests can be found in a nice jupyter notebook: https://colab.research.google.com/github/tensorflow/models/blob/master/research/deeplab/deeplab_demo.ipynb
In this notebook, to get the image indexing right, I was playing with rotated image: http://mrkonrad.github.io/MRKonrad/files/cnnRotatedImage/image2_90.jpg
http://mrkonrad.github.io/MRKonrad/files/cnnRotatedImage/image2_180.jpg
http://mrkonrad.github.io/MRKonrad/files/cnnRotatedImage/image2_270.jpg

Troubleshooting

• missing msvcp140.dll Download https://www.microsoft.com/en-us/download/details.aspx?id=48145 as suggested in https://stackoverflow.com/questions/32998902/msvcp140-dll-missing
• how to generate .pb (tensorflow) file from .h5 (keras)
  https://stackoverflow.com/questions/45466020/how-to-export-keras-h5-to-tensorflow-pb
  https://github.com/amir-abdi/keras_to_tensorflow

python keras_to_tensorflow.py 
       --input_model="path/to/keras/model.h5" 
       --output_model="path/to/save/model.pb"

• How to see model in a tensorboard

python ~/Code/tensorflow/tensorflow/python/tools/import_pb_to_tensorboard.py --model_dir tests/testData/model.pb --log_dir tests/testData/log
tensorboard --logdir tests/testData/log

TODO:

• remove dependency to hello_tf_c_api
• make it work with 32bit windows. I have unsuccessfully tried to:
  • compile tensorflow myself: https://github.com/MRKonrad/tensorflowBuild
  • use playertwo compiled tensorflow 1.6: https://github.com/MRKonrad/build-tensorflow
  • possibly I could start with playertwo solution from scratch and see where it goes

Thoughts

Reason for using 3d itk input images

These are the ways to provide n 2d input images:

• itkImage with third dimension as the n-th image
  Pro: can be defined in the runtime
  Con: breaches the logic behind itk image
• itkImage with n-dim vector type
  Pro: makes sense with itk logic
  Con: have to be defined at the compile time

PixelType = itk::Vector< float, n >; // n has to be static, cannot be defined in the runtime
ImageType = itk::Image< PixelType, 2 >;

• n inputs to a itkImageFilter class
  Pro: makes sense with itk logic
  Con: number of itkImageFilter have to be defined in a constructor. itkImageFilter constructor does not accept arguments. Hence the n has to be defined in the compile time
  Partial solution: allow big number of inputs

In this project I went with the first option.

Git and conventional commits

I am trying to follow the convetional commits guide:

• https://www.conventionalcommits.org/en/v1.0.0-beta.2/
• https://www.cheatography.com/albelop/cheat-sheets/conventional-commits/

Other

• colormap in rgb: https://itk.org/Doxygen/html/classitk_1_1ScalarToRGBColormapImageFilter.html