An example of how to serve a trained doddle-model in a pipeline implemented with Apache Beam and Scio. It includes batching of individual examples for faster vectorized predictions. Code is available here.
Run the pipeline with:
sbt run --runner=DirectRunnerAn example output:
processing a batch of 72 examples
processing a batch of 72 examples
processing a batch of 71 examples
processing a batch of 71 examples
predicted probability: 0.9991 --- label: 1.0
predicted probability: 0.9901 --- label: 1.0
predicted probability: 0.0000 --- label: 0.0
predicted probability: 0.9998 --- label: 1.0
predicted probability: 0.9235 --- label: 1.0
predicted probability: 0.9990 --- label: 1.0
predicted probability: 0.9951 --- label: 1.0
predicted probability: 0.0070 --- label: 0.0
predicted probability: 0.0012 --- label: 0.0
predicted probability: 0.9997 --- label: 1.0
...
To run the examples locally you will need to publish a local snapshot version of the repository:
git clone https://github.com/picnicml/doddle-model.git
cd doddle-model
sbt publishLocal
Ensure the published version matches the version contained within the project/Dependencies.scala file.
The breast cancer dataset is from UCI Machine Learning Repository. Irvine, CA: University of California, School of Information and Computer Science, Dua, D. and Karra Taniskidou, E.