Building Tensorflow from Source
I figure out a way that may get multiple GPU support with the latest CUDA. If I built a tensorflow 1.15 configured for CUDA 10.2 from source, I'll be able to build the PIP package .whl. My questions is once I have the PIP package, how to I point the DSS code env to import it as the tensorflow package to use?
Answers
-
Hi,
As stated in my previous answer, in the case of the Visual Deep Learning in DSS, the latest versions we support are:
- tensorflow 1.15
- keras 2.3
- CUDA 10.0
- cuDNN 7.6
If you use these versions specifically, you will be able to build your code-env directly from pip. This way you'll be able to create a managed code-env in DSS.
I would not recommend building tensorflow from source (unless there is a specific feature of CUDA 10.2 which is not available in CUDA 10.0).
Cheers,
Alex
