Installation with both CPU/GPU tensorflow modules

It is said in the document that keras will automatically run on GPU with tensorflow backend.

However, I found Keras depends on the installation order of tensorflow modules. It seems that keras always use the module installed last. For example,

pip install tensorflow-gpu tensorflow

Keras will use CPU under this circumenstance.

[austinmw]

Is there any way to get the above solution to work from a tensorflow-gpu container? TF fails to import using this container on a non-gpu machine. It’d be nice to be able to use the same docker image for both machine types.

[ziky90]

I have experimented with this a bit on my machine as well and it seems to me that the default behaviour working directly with tensorflow==1.1.0 is following:

If I perform the installation in the following order:

sudo pip install tensorflow
sudo pip install tensorflow-gpu

then GPU version is being used But if I install it in the opposite order:

sudo pip install tensorflow-gpu
sudo pip install tensorflow

then CPU version is being used.

I have performed the experiment not using keras, just importing tensorflow directly (the same way keras does it in the tensorflow backend module). And I have installed it directly to the root python 2.7. on Ubuntu 16.04.2.