Anyone interested in Tensorflow packaging? Makefile available now.
Hello d-Science and d-Python team,
Long time ago, TensorFlow packaging was blocked by bazel,
a google's java-based building system which is hard to harness.
At that time, cmake build was available but it's written for
windows.
However, by chance I discovered that tensorflow now ships
a set of makefiles[1], which is explicitly saying "support Ubuntu".
That means packaging TensorFlow is now viable, as long as we patch
the build system to prevent it from downloading tarballs.
If you are looking for things to do, or interested in machine learning
or deep learning applications, you might want to have a look at it.
However, I'm not working on it because:
1. I'm a bit overload these days, working on the Julia package.
2. unfortunately I don't like TensorFlow. I like it's main
competitor -- PyTorch.
But if someone is going to do something about it, I'd be happy to help.
Let's have a look at the dependencies:
* eigen [OK]
* gemmlowp [OK, maintained by me]
* googletest [OK]
* nsync [missing]
* protobuf [OK]
* re2 [OK]
* fft2d [missing]
* abseil [missing]
* cub [missing, NVIDIA's BSD-3-clause software that depends on CUDA]
We don't need this pile of code when compiling CPU-only tensorflow.
If someone is going to hack the makefiles, I'd suggest that,
try building a CPU-only version first.
Best,
[1] https://salsa.debian.org/science-team/tensorflow/blob/master/tensorflow/contrib/makefile/build_all_linux.sh
[2] https://salsa.debian.org/science-team/tensorflow/blob/master/tensorflow/contrib/makefile/download_dependencies.sh
Reply to: