CODEWITHSUNDEEP
rtensorflowgpuh2o
filius_arator
filius_arator

Reputation: 139

Building TensorFlow in h2o without CUDA

I have built model with TensorFlow backend on machine with GPU. Now i want to use this model on production machine which doesn't have GPU. I cannot use this model without installed TensorFlow backend but I also cannot install TensorFlow backend on machine without GPU.

Is there any option to built TensorFlow backend for H2O without GPU support?

Upvotes: 1

Views: 194

Answers (1)

Mateusz Dymczyk
Mateusz Dymczyk

Reputation: 15141

but I also cannot install TensorFlow backend on machine without GPU

Do you mean you tried and got errors?

Is there any option to built TensorFlow backend for H2O without GPU support?

Yes but depending on your machine it might be easy or a bit hacky. If you're running on a MacOS system then it should work out of the box, just follow our instruction.

If you are on a Linux based system then you'd have to first change one our scripts, to be exact this one. You'd have to do two things:

  • change export TF_NEED_CUDA=1 to export TF_NEED_CUDA=0
  • change export BUILDFLAGS="--config=cuda --copt=-m64 --linkopt=-m64 --copt=-march=native" to export BUILDFLAGS="--copt=-m64 --linkopt=-m64 --copt=-march=native"

You might also remove --config=cuda under echo "Build pip package".

After this you follow the instructions linked above as normal.

This should work, but I haven't tested it myself so should you get errors please raise an issue on DeepWater's Github issue tracker.

Upvotes: 3

Related Questions