BobDu/dockerfiles

Docker

This folder holds Dockerfiles for our machine learning images. It has the following subfolders:

• **base** holds the base images that provide particular versions of Python, common libraries such as numpy, and GPU drivers. They are named as Python version + CPU/GPU, e.g. tinymind/base:py27-cpu.
  • Regular base images contain conda-installed Python. The -pyenv variant comes with pyenv-installed Python. Unless absolutely necessary, use the regular version. The pyenv version does not contain all dependencies (e.g. pygpu, mkl). In particular, all framework images are built using the conda base.
  • The reason to have a pyenv version is that conda doesn't play nicely with setuptools. See more details here.
• **example** holds example environments for testing purposes.
• **frameworks** holds images for specific frameworks.
  • Use the scripts like tensorflow.py to generate Dockerfiles. You need to be in the frameworks folder when running the scripts.
  • There are two types of framework images - "base" images are for running executions, and "notebook" images inherit from "base" ones and provide additional setup for notebooks.
  • Notebook images assume that the user running the image has uid=1001, gid=1002.

Base Images

To build a new version, do the following:

cd docker/base
# py36cpu is the name of the image. Image name should be the same as the
# suffix of the Dockerfile.
docker build . -t py36cpu -f Dockerfile.py36cpu
# Tag the image with tinymind/NAME.
docker tag py36cpu tinymind/base:py36-cpu
# Push to Docker Hub (need to log in as tinymind).
docker push tinymind/base:py36-cpu

The base folder contains a rebuild-all.sh script that builds all base images.

Frameworks

To build a new version, do the following:

cd docker/frameworks
# By default Dockerfiles for all versions of a framework are generated. You
# can use the flags to selectively generate Dockerfiles.
# --versions: list of framework versions (1.3).
# --langs: list of python versions (py27).
# --archs: list of cpu/gpu.
# --nobase: if specified, don't build "base" images.
# --nonb: if specified, don't build notebook images.
python keras.py --nonb --langs py27 py36

cd /tmp/tmbuild/
sh build.sh

Pre install python package

RUN pip --no-cache-dir install --upgrade \
        Pillow \
        h5py \
        jupyter \
        keras_applications \
        keras_preprocessing \
        matplotlib \
        numpy \
        scipy \
        scikit-learn \
        pandas \
        mkl \
        pyyaml \
        Cython \
        opencv-python \
        tinyenv