-
Notifications
You must be signed in to change notification settings - Fork 101
TensorFlow module
To enable this module set FAST_MODULE_TensorFlow in CMake, reconfigure and compile. A couple of new CMake options will appear (FAST_BUILD_TensorFlow_*), one for each version of TensorFlow, by default only the TensorFlow CPU version is built. You can build the CUDA and ROCm versions by selecting these in CMake. Please make sure you have the requirements installed before you proceed to build TensorFlow (see below). With these options Tensorflow will be downloaded from github, configured and built automatically when you build FAST. Note that Tensorflow is a big library and compiling it takes a long time, especially on Windows.
Current TensorFlow version used by FAST is 1.14.
TensorFlow will from FAST version 3.2 be using TensorFlow 2.3.0.
- TensorFlow is built using a Google's build system called bazel, Google's build system.
- Note that bazel requires java to be installed.
- For FAST <= 3.1 make sure you download version 0.24.1 from here: https://github.com/bazelbuild/bazel/releases/tag/0.24.1
- For FAST >= 3.2 use bazel 3.1.0: https://github.com/bazelbuild/bazel/releases/tag/3.1.0
- If you install bazel on Windows remember to put the bazel.exe in system path.
- For Windows you also need msys2 and some packages for bazel, see intructions here: https://docs.bazel.build/versions/master/install-windows.html#installing-compilers-and-language-runtimes
- Python 3 with numpy installed, use version 1.19.0 of numpy as latest versions have some issues on windows: pip install numpy==1.19.0
- For the CUDA version install both CUDA toolkit and cuDNN.
- For the ROCm version install the ROCm compute stack
If this wiki page lacks some information or is incorrect please let us know! You can edit this wiki page yourself, send an email to [email protected] or