Experimenting faster and effecient docker builds for projects including tensorflow ( which is large ).

Using Tensorflow Docker Image with GPU

REF: https://www.tensorflow.org/install/docker

Method 1 (Not recommended)

Manual installation of tensorflow as a python package.

• Not implemented Due to size of the package better have a separate base image that is to be used in an application image

Method 2 (Recommended)

Using docker hub tensorflow docker images as base images.

Example:

FROM tensorflow/tensorflow:latest-gpu

# Add your application here

REF: https://hub.docker.com/r/tensorflow/tensorflow/

Tested tags:

• latest tag
• latest gpu tag
• nightly tag
• nightly gpu tag

Limitations

• The base images are large. So, the final image size will be large.
• The base images are large. So, the builds will take reasonable time.
• The base images are large. So, the docker hub pulls will take reasonable time.
• No application is added to the base images. So, build time may vary depending on the application.

Observations

• Using docker hub tensorflow images as base images is fast ( Not tested Manual installation of tensorflow as a python package. ). Since the base images are already built and available in docker hub. Builds take reasonable time.
  • latest tag
    • 37s
  • latest gpu tag
    • 1m 0s
  • nightly tag
    • 24s
  • nightly gpu tag
    • 1m 7s

Disclaimer

• All runs are done in Github Actions.