Scripts to create Debian packages with the Tensorflow C/C++ API for Ubuntu x86_64 and Raspberry Pi
The binaries are built with Tensorflow 1.13.1. The generic binaries are now built for minimum Haswell of Intel and Piledriver of AMD (AVX, FMA instructions). The optimized binaries are now built for minimum Skylake of Intel and Excavator of AMD (AVX, AVX2, FMA instructions). CUDA enabled binaries are built with CUDA 10.1 and based on tensorflow_cc project. Credits go to FloopCZ.
CHANGES: The built CUDA packages are restricted to Compute Capabilities 6.1 and 7.5. The older capabilities were removed to speed up the initialization of the runtime. You can check the capability of your card here.
EDIT (2019/05/11): New generic/optimized non-CUDA version to fix two bugs in dynamic/static linking. EDIT (2019/05/20): Adding Raspberry Pi package (Ubuntu Xenial build) EDIT (2019/11/19): Added legacy packages compiled for a minimum of Ivy Bridge processors. EDIT (2020/01/16): Adding Raspberry Pi package (Ubuntu Bionic build) - the filename has "bionic" inside.
The binaries are built with Tensorflow 1.13.1. The generic binaries are now built for minimum Haswell of Intel and Piledriver of AMD (AVX, FMA instructions). The optimized binaries are now built for minimum Skylake of Intel and Excavator of AMD (AVX, AVX2, FMA instructions). CUDA enabled binaries are built with CUDA 10.1 and based on tensorflow_cc project. Credits go to FloopCZ.
The generic, optimized and optimized/CUDA builds for x86_64 were built on Ubuntu Bionic (18.04).
Changes:
The generic and optimized builds for x86_64 were built on Ubuntu Bionic (18.04).
The Raspberry Pi release was built on Ubuntu MATE (Xenial) since MATE has not been updated to Bionic on Raspberry Pi yet.
EDIT1: 2018-08-30: I found a few minor problems in the generated CMake configuration for the generic and optimized builds thus I fixed the issues and updated the Debian packages here. EDIT2: 2018-08-30: I add tar.xz versions of the generic and optimized builds for non-Debian systems.