Tensorflow for Android and Windows

[bhack]

Do you think that the new google tensorflow will be presettable?

[saudet]

Sure, no reason why not :)

[bhack]

Some more info on the c++ interfaces

[saudet]

It should be possible to make a build for Android with the cross-compiling functionality of Bazel explained here: https://github.com/bazelbuild/examples/tree/master/android-ndk

As for Windows, we'll have to wait until Bazel itself starts supporting builds on it...

[andreas-eberle]

I had a look at the Android build. The most important thing is to patch the WORKSPACE file (see https://github.com/tensorflow/tensorflow/tree/master/tensorflow/examples/android#to-buildinstallrun).

The thing is, that we'll need to specify the path to the local NDK in there. From what I saw, it is not possible to use environment variables in the WORKSPACE file. @saudet: How would you set the path in there dynamically with the path provided by the user to the cppbuild.sh script?

[saudet]

Create a patch, like we have to do with FFmpeg, for example:
https://github.com/bytedeco/javacpp-presets/tree/master/ffmpeg

[saudet]

@andreas-eberle Could you update your patch for TensorFlow 0.11.0rc1?
It doesn't seem to work anymore: c9cf2a8
Send a PR against the tensorflow branch if you could. Thanks!

[andreas-eberle]

I will see if I can fix it.

[andreas-eberle]

@saudet: I have a fixed version. However, it depends on two fixes in tensorflow being merged into the r0.11 branch (see issue tensorflow/tensorflow#5374 and PR tensorflow/tensorflow#5135). I'll keep you updated.

[saudet]

Great thanks! They should be in the release anyway right?

[andreas-eberle]

@saudet: Unfortunately the TensorFlow guys decided that they won't include the improved/fixed python_config.sh file into the r0.11 release. That's why I created a patch to use the fixed version from master. I sent you a PR for your tensorflow branch.

[andreas-eberle]

@saudet: Tensorflow seems to support Windows now. I will try to get it working.

[saudet]

I've investigated on how we could proceed for Windows, and here's what I found:

• TensorFlow doesn't support GCC on Windows, so we need to use MSVC:
  https://groups.google.com/a/tensorflow.org/d/msg/discuss/u6RHBEM5Uwc/BWILiei2AgAJ
  I've tried GCC, and there's just too many things broken, so this means that to export the C++ API we can't use GCC's --enable-runtime-pseudo-reloc .
• TensorFlow doesn't support exporting the C++ API via DLLs with MSVC either:
  Packaged TensorFlow C++ library for bazel-independent use tensorflow/tensorflow#2412 (comment)
• So we could use a static library instead, but Bazel doesn't even support that:
  Build a static library tensorflow/tensorflow#5511
• That leaves the one possibility: Creating a static library with CMake, but the CMake build won't work with MSYS2 tooling and requires instead Anaconda, further complicating the build process:
  https://github.com/tensorflow/tensorflow/tree/master/tensorflow/contrib/cmake

[saudet]

It turns out that we don't need Anaconda when not building the bindings for Python. An artifact for Windows is now available, but until the next full release, since tensorflow-platform doesn't include the windows-x86_64 platform, we need to add it separately:

        <dependency>
            <groupId>org.bytedeco.javacpp-presets</groupId>
            <artifactId>tensorflow</artifactId>
            <version>1.7.0-1.4.1</version>
            <classifier>windows-x86_64</classifier>
        </dependency>

But that's it! Enjoy and thanks for your patience.