forestflow/forestflow

ForestFlow is a policy-driven Machine Learning Model Server. It is an LF AI Foundation incubation project.

forestflow's Issues

Add ONNX flavor support

Currently ForestFlow only supports H2O flavors. This is extensible and we'd like to add support for ONNX-based models to cover a wider variety of frameworks that know how to export to ONNX.

This SHOULD look into runtime/platform considerations and provide switches for GPU support for example.

Add automatic inference from a Kafka topic

ForestFlow's core doesn't care what the source of inference requests is. It's built this way specifically to enable support for additional interfaces. As it is today, there's support for a basic HTTP REST API in addition to Graphpipe clients.
We'd like to add support for records coming from Kafka topics but lay this out in a way that we can easily extend support for other queuing systems like Faktory or Pulsar, or even an S3 or file-system watcher etc... This ticket will focus on Kafka first since it's already something ForestFlow integrates with for inference logging.

Add S3 SourceStorageProtocol support

ForestFlow already has stubs for supporting different protocols. See here: https://github.com/ForestFlow/ForestFlow/blob/master/core/src/main/scala/ai/forestflow/utils/SourceStorageProtocols.scala

and here for S3: https://github.com/ForestFlow/ForestFlow/blob/master/core/src/main/scala/ai/forestflow/utils/SourceStorageProtocols.scala#L87

It currently has implementations for Git and Local file systems. This ticket is to finish the implementation for S3. Sticking to the AKKA system, preferably using Alpakka for S3.

Model Explainability: Obtain SHAP values from MOJO model

In an effort to provide explanations for each prediction, in Forestflow, I would need to add .setWithDetailedPredictionCol(true), outlined here:

https://s3.amazonaws.com/h2o-release/sparkling-water/spark-2.3/3.26.5-2.3/doc/tutorials/shap_values.html#get-contributions-from-raw-mojo

"The call setWithDetailedPredictionCol(true) tells the service to create an additional prediction column with additional prediction details, such as the contributions. The name of this column is by default “detailed_prediction” and can be modified via setDetailedPredictionCol setter."

The addition would need to go around here in Forestflow:

ForestFlow/core/src/main/scala/ai/forestflow/serving/impl/H2OServable.scala

Line 41 in 3a46f85

.setConvertInvalidNumbersToNa(true)