即使已经安装了tensorflow-gpu，带有keras软件包的conda bug也会安装tensorflow

Question

I wanted share a problem I encountered and the solution I found that worked on my system.

For reference, this is my environment: Ubuntu 16.04 64-bit, nVidia GPU, conda 4.5.2

I created a conda virtual environment and installed tensorflow-gpu (which resulted in 1.7.0 being installed which is not the latest, but is fine for my purposes).

I then did $ conda install keras

keras was installed correctly, BUT it also installed tensorflow non-gpu version, which I cannot use.

This seems like a dependency bug in which keras package does not recognize tensorflow-gpu package as installed.

I was forced to remove tensorflow (non-gpu package) and to use

$ python3.6 -m pip install keras

This worked, but I would rather have consistent conda managed packages.

I hope that Anaconda fixes this, but in the meantime, you can try my solution.

BTW, don't even get me started on why I couldn't just use pip or pip3 instead of "python3.6 -m pip". Python 3.6 and 3.5 on Ubuntu at least are at war with each other causing all kinds of "module not found" problems. But that's for another post.

Answer 1

You can try keras-gpu

conda install -c anaconda keras-gpu

Where -c anaconda indicates a channel where to look for the package. (You may want to try the default conda install keras-gpu or if you know about other channels, use them).

---

About python versions conflicting, it's a good idea to create specific environments, so you can also try different versions of tensorflow + keras without conflicts.

On my machine, using the intel version of python was a great deal for processing speed (although you're going to use GPU, not my case)

Create a new environment:

#anaconda version
conda create -n yourenvname python=x.x anaconda

#intel version
conda create -n yourenvname intelpython3_core python=3
    #can also be specific, like python=3.5

Then you activate your environment:

source activate yourenvname

And install and use everything with this environment activated.

If using intel Python, it's interesting to follow the rest of the process here to use the optimized intel distributions of numpy, scipy and others: https://software.intel.com/en-us/articles/using-intel-distribution-for-python-with-anaconda

More info on environments: https://uoa-eresearch.github.io/eresearch-cookbook/recipe/2014/11/20/conda/