/dockerfiles

Deep Learning Dockerfiles

Primary LanguagePythonApache License 2.0Apache-2.0

Dockerfiles

Collection of Dockerfiles useful for NLP and Deep Learning. To download the docker images visit: floydhub's Docker Hub.

How to update framework

  1. Dockerfiles are generated using two inputs: matrix.yml and jinja template file inside ./dl/FRAMEWORK directory. matrix.yml provides variable values for jinja template files.
  • The $render list in matrix.yml controls what version of the framework to render.
  • For each version config in matrix.yml, any key starts with _ are global keys, which will get automatically injected into each variant config for that version.

  1. Most of the cases, you only need to update ./dl/FRAMEWORK/matrix.yml to generate a set of dockerfiles for a new version of a framework. If not, you will need to update the jinja file to account for build step changes.

  2. Install floydker: cd floydker && pipenv shell && pipenv install.

  3. Render dockerfiles: cd .. && floydker render ..

  4. Commit new docker images to git and push: git commit -a.

Naming conventions

Dockerfiles are organized into the following directory structure:

CATEGORY/PROJECT_NAME/VERSION/Dockerfile-ENV
CATEGORY/PROJECT_NAME/VERSION/Dockerfile-ENV.gpu

Automated build scripts will generate the following tags for images based on the above dockerfile paths:

floydhub/PROJECT_NAME:VERSION-ENV
floydhub/PROJECT_NAME:VERSION-ENV-gpu

Contains docker images for popular deep learning frameworks including: Tensorflow, PyTorch and Torch.