Have you followed this page:
https://www.tensorflow.org/install/install_sources

I agree that ultimate installation instructions are for python pip packages. That may be confusing. But once you have the source directory, you should be able to build anything.

@asimshankar do we have a document to describe libtensorflow_cc.so build and installation?

Yes, I have gone through all of the steps to build and install the python components into site-packages from source. I've also used bazel to build libtensorflow_cc.so. But there are still several other manual steps required to arrange for all of the necessary C++ headers to be installed in such a way that C++ applications can compile independently of the tensorflow repository.

In my search for the best way to do this, the best resource I have found is on stackoverflow:
https://stackoverflow.com/questions/33620794/how-to-build-and-use-google-tensorflow-c-api

That question is two years old(!!) and none of the answers are (IMHO) satisfactory. The answer that comes closest is this one. But the steps as shown there did not work for me, possibly due to changes to tensorflow between the time the answer was written and the current r1.4 release. I have been able modify the steps to make something that does work, but it is really unacceptable for the Tensorflow team to expect every developer to discover on their own how to do this. There really should be something like the standard unix recipe: .configure; make cc-installer; sudo install-cc.sh;. With bazel, the steps might look like this:

1. ./configure # a new question is added: "Do you want to build for C++ development?"
2. bazel build --config=opt --config=cuda //tensorflow/tools/cc_package:build_cc_package
3. sudo install-cc-package.sh

The result from the above steps would be the installation of a directory at the default path /usr/local/include/tf, but the actual path should be an option provided during the ./configure step. The tf directory should contain all of the headers necessary for C++ development. It would a superset of all of the headers currently installed into ../site-packages/tensforflow/include, but following the same directory organization. The installer script would also install libtensorflow_cc.so and libtensorflow_framework.so into /usr/local/lib or some other location chosen during the ./configure step.

Let me be clear: I personally am not blocked, as I now have a working configuration on both my Mac and Linux development machines. But other developers (and me sometime in the future) will need to do this, and it seems bizarre that the Tensorflow team does not have a documented and easy to follow process that all developers who need C++ development can follow, and will continue to work with each new tensorflow release. I could document what I do, possibly on the above stackoverflow thread, but the Tensorflow community is not served well by that hack.

It has been 14 days with no activity and this issue has an assignee.Please update the label and/or status accordingly.

Nagging Assignee: It has been 14 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

I'm curious if there has been any progress on this issue. I'm currently working on making a Docker image to deploy to Google Cloud and I've discovered that I am once again blocked by this issue.

If I pull one of the docker images, I find that libtensorflow_cc.so is not included anywhere in the image. What is worse, when I try to build it (with gpu support) I get errors, similar to this issue: #2143

I tried both nightly-devel-gpu-py3 and 1.5.0-devel-gpu-py3 and obtained the same error.

This error seems to indicate that libtensorflow_cc.so is not even built in your continuous integration. Is that correct?

@asimshankar @allenl to comment.
I think many symbols exporrted by libtensorflow_cc.so are also exported by the .so files distributed by our pip packages. We should wait for confirmation though.
You are correct that in our dockerfiles we do not have anything other than the py libraries. Maybe we should also make C and Java libraries available in these.

The pip package only includes libtensorflow_framework.so, which is not sufficient for C++ development.

pywrap_tensorflow.so is also in the pip package, it may be in a harder to locate place, and may need to be moved around in our pip package. And as far as I know, that .so file should be exporting every single C++ object in TF (which is something we are working to fix).

Nagging Assignee @gunan: It has been 14 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

"you should be able to build anything"

I agree, I should be.
But how do I do it?
How do I even know what it is I want to build?
How do I know that libtensorflow_cc.so is a target I should ask Bazel to build?
What other targets should I build?
Which headers should go into a public include directory, as opposed to are for internal use only?

I can't find any documentation at all on how to build and user TensorFlow as a C and C++ library, or much about what all can be built with the Bazel builder at all. I'm sure some of that is me not being good with Bazel, but that can't be the only thing getting in the way here.

Note that https://www.tensorflow.org/install/install_sources explicitly talks about how to build the pip packages, not how to build for C++.

@asimshankar Do we have documentation for libtensorflow?

Nagging Assignee @gunan: It has been 14 days with no activity and this issue has an assignee. Please update the label and/or status accordingly.

Can @asimshankar please reply to this thread?

The C library package (consisting essentially of a single header file and a few shared libraries) can be built using the process we use for building release binaries. (See README and StackOverflow answer).

For the C++ library, we haven't had the bandwidth to support that ourselves yet and have relied on community support for that (see #2412 (comment) and #2412 (comment) and #2412 (comment) - credit to @FloopCZ and others).

If someone would like to contribute a packaging rule for the C++ library by adding a BUILD target similar to libtensorflow in https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/lib_package/BUILD, that would be awesome.

Please remove the assignee, as this issue is inviting external contributions. Otherwise, remove the contributions welcome label. Thank you.

Propose that this project should be part of the official tensorflow docker images: https://github.com/FloopCZ/tensorflow_cc

Building and using libtensorflow_cc.so from source is non-trivial and burns dev time. Would be good for community if the official images support easy c++ dev and integration. Thanks.

This issue is stale because it has been open for 180 days with no activity. It will be closed if no further activity occurs. Thank you.

This issue was closed because it has been inactive for 1 year.