/tensorflow-c-api

Primary LanguageShell

Documentation for setting up running environment for Tensorflow C++ API without bazel

[Updated 2/2/2018]

Possible problems:

  • Libtensorflow_framework.so not found!

    Locate file at ~/Desktop/tensorflow/bazel-bin/tensorflow. Need to run sudo ldconfig to refresh linker.

  • Nsync.h not found!

    Download at nsync git repos, and place it in the right folder tensorflow/core/platform/default/.

  • Tensorflow version = 1.5.0-rc1/master; Bazel version = 0.9.0; Protocbuff version = 3.4.0

  • Remember to run ./configure in tensorflow main repo before run bazel command bazel build //tensorflow:libtensorflow_cc.so.

  • *.pb.h missing.

Follow this discussion.

  • libcupti.* missing or no extras/CUPTI related problem
  1. Install libcupti8.0 and libcupti-dev from deb packages. See this post.
  2. create etras/CUPTI directory and link it with libcupti libraries. See this post.

Steps:

  1. Install protocbuf, bazel and eigen.

    eigen comes with VENTOS already, so no need to install it again.

    protobuf need to be installed from source code. Follow the instruction in here to install protobuf from source.

    bazel released version can be downloaded from here. Use the following command to install bazel:

    sudo dpkg -i bazel_0.9.0-linux-x86_64.deb

    As for tensorflow version v1.5.0-rc1 and bazel 0.9.0, only protobuf v3.4.0 works. (As Feb 2, 2018, bazel 0.10.0 does not work with tensorflow v1.5.0-rc1. Use bazel 0.9.0 instead.).

  2. Install cuda-8.0 and cudnn-6.1. Follow this instruction.

  3. Start to build tensorflow libraries.

./configure
bazel build  //tensorflow:libtensorflow_cc.so

  1. Merge bazel-genfiles/tensorflow with tensorlow/tensorflow

  2. Then Copy the following include headers and dynamic shared library to /usr/local/lib and /usr/local/include (You are at tensorflow main folder):

cp -r tensorflow /usr/local/include/
cp -r third_party /usr/local/include/
cp -r bazel-bin/tensorflow/libtensorflow_cc.so /usr/local/lib/
cp -r bazel-bin/tensorflow/libtensorflow_framework.so /usr/local/lib/
  1. Lastly, compile using an example:
g++ -std=c++11 -o tTest test.cc -I/usr/local/include/tensorflow/contrib/makefile/downloads/absl -I/usr/local/include/tf -I/usr/local/include/eigen3 -g -Wall -D_DEBUG -Wshadow -Wno-sign-compare -w -L/usr/local/lib/libtensorflow_cc -ltensorflow_cc -L/usr/local/lib/libtensorflow_framework -ltensorflow_framework `pkg-config --cflags --libs protobuf`

How to port to VENTOS

  1. Add -I/usr/local/include/eigen3 to Property->OMNeT++->Makemake->src->Options->Preview.

  2. Add tensorflow_cc and tensorflow_framework to Property->OMNeT++->Makemake->src->Options->Link.

  3. Include tensorflow head files.

Graph building

Note on exporting graph and model in tensorflow:

  1. Simple tensorflow graph file *.pb only contains graph topology information. It doesn't contain any variables like weight or bias information.

  2. Metagraph(checkpoint files) contains graph information in *.meta, and variable values in *.data and *.index files.

  3. One can consolidate *.pb and *.meta/*.data/*.index into a single *.pb file which is known as frozen graph. (Python API provides a frozen_graph.py script to freeze a graph)

  4. However, frozen graph can only be used for inference. One can reload the frozen graph in other applications and use it to do tasks like prediction. But it can't be used for further training. This is because the variables are converted to constants when the graph is frozen.

  5. In order to make the saved graph applicable for continuing training, the following approach can be used (at least for TF version 1.3.0): load checkpoint files. This is so far the only possible way to reuse a pre-trained model for continuing training purpose.

Related discussions:

Written with StackEdit.