Implements image classification functions using Axon machine learning models managed by Bumblebee.

Image classification refers to the task of extracting information from an image. In this module, the information extracted is one or more labels desribing the image. Typically something like "sports car" or "Blenheim spaniel". The labels returns depend on the machine learning model used.

Configuration

The machine learning model to be used is configurable. The :model and :featurizer may be any model supported by Bumblebee. Any additional options can be supplied as keyword lists under the :model_options and :featureizer_options keys. The :name is the name given to the classification service process.

The default configuration is:

# runtime.exs
config :image, :classifier,
  model: {:hf, "microsoft/resnet-50"},
  featurizer:  {:hf, "microsoft/resnet-50"},
  model_options: [],
  featurizer_options: [],
  batch_size: 10,
  name: Image.Classification.Server,
  autostart: true

Autostart

If autostart: true is configured (the default) then a process is started under a supervisor to execute the classification requests. If running the process under an application supervision tree is desired, set autostart: false. In that case the function classifer/1 can be used to return a Supervisor.child_spec/0.

Adding a classification server to an application supervision tree

To add image classification to an application supervision tree, use Image.Classification.classifier/1 to return a child spec: For example:

# Application.ex
def start(_type, _args) do
  children = [
    # default classifier configuration
    Image.Classification.classifier()

    # custom classifier configuration
    Image.Classification.classifier(model: {:hf, "google/vit-base-patch16-224"},
      featurizer: {:hf, "google/vit-base-patch16-224"})
  ]

  Supervisor.start_link(
    children,
    strategy: :one_for_one
  )
end

Limitations

This module is provided as a convenience to make it easy and idiomatic to perform simple classification tasks while still managing CPU capacity.

For more complex classification requirements it is recommended to use Bumblebee directly.

Note

This module is only available if the optional dependency Bumblebee is configured in mix.exs.