This post is an udate of the last post on how to make a standalone C++ project with tensorflow without using Bazel. This post aims at upgrading tensorflow to the latest stable version 1.8.0 and some tweaks that help tensorflow work nicely with opencv.

I was quite happy with the setup that I had from the last post. However, one thing that frustrates me is that I cannot use some of opencv’s functions like imread or imshow when including tensorflow headers in my codes. The reasons are likely to be the conflict of symbols between tensorflow and opencv around the protobuf library. For more details you can look at issues 14267, 13278 and 1924. The solution for this problem will be showed in the next part.

Update tensorflow to version 1.8.0:

git pull origin master
git fetch --tags
git checkout v1.8.0

Now you can built your libtensorflow_cc.so in adding the flag --config=monolithic which will condense everything together into one shared object (no libtensorflow_framework dependence) and seal off non-TensorFlow symbols. This prevents the symbol conflict with opencv. I add also some flags that will help your tensorflow run faster:

bazel build -c opt --copt=-mavx --copt=-mavx2 --copt=-mfma --copt=-mfpmath=both --copt=-msse4.1 --copt=-msse4.2 --config=monolithic //tensorflow:libtensorflow_cc.so

You will get a libtensorflow_cc.so as the output of the command above. Use it in your project (you don’t need libtensorflow_framework.so anymore) with opencv.

If you have any question, feel free to contact me to: phan at tuanphuc dot com.