Python 3.6 in tensorflow gpu docker images

Question

How can I have python3.6 in tensorflow docker images.

All the images I tried (latest, nighty) are using python3.5 and I don't want to modify all my scripts.

Answer 1

The Tensorflow images are based on Ubuntu 16.04, as you can see from the Dockerfile. This release ships with Python 3.5 as standard.

So you'll have to re-build the image, and the Dockerfile will need editing, even though you need to do the actual build with the parameterized_docker_build.sh script.

This answer on ask Ubuntu covers how to get Python 3.6 on Ubuntu 16.04

The simplest way would probably be just to change the From line in the Dockerfile to FROM ubuntu:16.10, and python to python3.6 in the initial apt-get install line

Of course, this may break some other Ubuntu version-specific thing, so an alternative would be to keep Ubuntu 16.04 and install one of the alternative ppa's also listed in the linked answer:

RUN add-apt-repository ppa:deadsnakes/ppa &&
    apt-get update &&
    apt-get install -y python3.6

• Note that you'll need this after the initial apt-get install, because that installs software-properties-common, which you need to add the ppa.
• Note also, as in the comments to the linked answer, that you will need to symlink to Python 3.6.
• Finally, note that I haven't tried any of this. The may be gotchas, and you may need to make another change to ensure that the correct version of Python is used by the running container.

Answer 2

You can use stable images which are supplied by third parties, like ufoym/deepo. One that fits TensorFlow, python3.6 and cuda10 can be found here or you can pull it directly using the command docker pull ufoym/deepo:py36-cu100

I use their images all the time, never had problems

Answer 3

With this anwer, I just wanted to specify how I solved this problem (the previous answer of SiHa helped me a lot but I had to add a few steps so that it worked completly).

Context:

I'm using a package (segmentation model for unet++) that requires tensorflow==1.4.0 and keras==2.2.2. I tried to use the docker image for tensorflow 1.4.0, however, the default version of python of this image is 3.5 which is not compatible with my package.

I managed to install python3.6 on the docker images thanks to the following files:

My Dockerfile contains the following lines:

Dockerfile:

FROM tensorflow/tensorflow:1.4.0-gpu-py3

RUN mkdir /AI_PLATFORM
WORKDIR /AI_PLATFORM

COPY ./install.sh ./install.sh
COPY ./requirements.txt ./requirements.txt
COPY ./computer_vision ./computer_vision
COPY ./config.ini ./config.ini

RUN bash install.sh

Install.sh:

#!/urs/bin/env bash

pip install --upgrade pip
apt-get update
apt-get install -y python3-pip
add-apt-repository ppa:deadsnakes/ppa &&
apt-get update &&
apt-get install python3.6 --assume-yes
apt-get install libpython3.6
python3.6 -m pip install --upgrade pip
python3.6 -m pip install -r requirements.txt

Three things are important:

• use python3.6 -m pip instead of pip, else the packages are installed on python 3.5 default version of Ubuntu 16.04
• use docker run python3.6 <command> to run your containers with python==3.6
• in the requirements.txt file, I had to specify the following things:

h5py==2.10.0
tensorflow-gpu==1.4.1
keras==2.2.2
keras-applications==1.0.4
keras-preprocessing==1.0.2

I hope that this answer will be useful

Answer 4

Maybe the image I created will help you. It is based on the cuda-10.0-devel image and has tensorflow 2.0a-gpu installed.

You can use it as base image for your own implementation. The image itself doesn't do anything. I put the image on dockerhub https://cloud.docker.com/repository/docker/patientzero/tensorflow2.0a-gpu-py3.6

The github repo is located here: https://github.com/patientzero/tensorflow2.0-python3.6-Docker

Pulling it won't do much, but for completeness:

$ docker pull patientzero/tensorflow2.0-gpu-py3.6

edit: changed to general tensorflow 2.0x image. Also as mentioned here, the official image for the beta 2.0 release now comes with python 3.6 support