Skip to content

Install TensorFlow#

We use TensorFlow as a Keras backend. Official page: link

Also see install-tensorflow_cc for an alternative that is more prepared for use with CMake.

Dependencies for building TensorFlow#

To build TensorFlow from source (e.g. for GPU support), you'll need:

  • Bazel
  • The python enum module, which on Ubuntu can be installed via: sudo apt install python-enum34

Install TensorFlow with GPU (Ubuntu 16.04)#

This page is to track working setups and common pitfalls. It is not replacement for the official documentation, above. Before attempting GPU installation, check the official requirements on GPU (micro-architecture, etc), and go through all the official steps during the process.

Some very important steps are those of the Post-installation Actions (CUDA). The end of your .profile will probably end up looking something like:

# cuDNN
Note: better than LD_LIBRARY_PATH, put in correct place and run ldconfig.

Working setups#

RTX 2080 Ti#

  • CUDA 10, cuDNN 7.5.
  • As general practice, better .run than .deb: e.g. gives option to install CUDA without modifying your installed NVIDIA drivers)
  • As in ended up using bazel 0.18.0 via .run (the .deb had a java package issue), and tensorflow r1.11. Otherwise ended up with Bazel not finding tensorflow configuration.
  • keras via pip, 2.2.3 or similar with no issues.

GM200 GeForce GTX TITAN X rev a1#

  • CUDA 9.0 (uninstalls any NVIDIA driver, installs 384.130 driver, so you may not need to Install NVIDIA drivers). We go to legacy and get deb (local) (cuda-repo-ubuntu1604-9-0-local_9.0.176-1_amd64.deb and patches from here). Official documentation is currently at CUDA 9.0, and here they say 9.1 will be skipped. + cuDNN v7.1.3 (April 17, 2018) for CUDA 9.0.
    • tensorflow 1.8-rc0 at e1e5f305e5359fd50340ea76ea2f737f6e87a0d7 (tried 1.7 but was broken for GPU). From source, with CUDA, said Yes to cuDNN 7.0 even with 7.1.3, without TensorRT (Ubuntu 16.04 local deb v3 was installed, but said default No).
    • tensorflow 1.5 (directly using tensorflow-gpu binary), without TensorRT.
  • Not tested: CUDA 8.0 + tensorflow 1.4 (directly using tensorflow-gpu binary)

Non-working setups#

TensorFlow for GPU at says:

  • CUDA 9.0:
    • Which recommends Drivers 390, with no GeForce 200 Series support (min GeForce 400 Series), but should support GeForce GTX 675M.
  • CUDA micro-arch 3.0 Kepler from source, or 3.5 Kepler for bin: GTX 260 is 1.3 Tesla, and GTX 675M is 2.1 Fermi.