This directory contains a simple utility to convert TensorFlow models
to C++ or Numpy code.

See tfconvert.py for the main goal/motivation.

----------------------------------------------------------------------
One-time setup:
 - Install Anaconda python
 - conda install keras tensorflow theano

If you're compiling C++ models:
 - tar xf eigen-3.3-beta2.tar.bz2

----------------------------------------------------------------------
Tests - follow the instructions after running the command(s) below
to complete all testing.

 KERAS_BACKEND=tensorflow python test1.py
 KERAS_BACKEND=tensorflow python test2.py

Models trained in Theano can also be converted to TensorFlow:
  https://github.com/fchollet/keras/wiki/Converting-convolution-kernels-from-Theano-to-TensorFlow-and-vice-versa
Theano can be trained on Windows.

----------------------------------------------------------------------
Converting a saved Keras model (from model.save()):
 KERAS_BACKEND=tensorflow python convert_keras.py nnet_bds BdsMpModel model.h5
  # where: "nnet_bds" is the generated output file prefix
  #        "BdsMpModel" is the generated C++ class name

----------------------------------------------------------------------
To install a trained model on a receiver you just need:
 - eigen-3.3-beta2.tar.gz
 - gen/some_file_out_gen.cpp
 - gen/some_file_out_gen.cpp.h
 - tfconvert_ops.h

----------------------------------------------------------------------
Things to investigate:
 - don't store all intermediate values in the class (although this is
   useful for debugging problems).  Perhaps the compiler could do
   more optimizations?
 - investigate using Eigen Matrix instead of Tensor? (but what about
   3D recurrent neural network inputs?)
 - split operations into multiple functions for better optimization?