Is there a kotlin native dataflow or pipeline engine available, which i can also run on embedded dev...

# datascience

[Yingding Wang]

Is there a kotlin native dataflow or pipeline engine available, which i can also run on embedded device such as android? Unfortunately, spark will not be an option for me.

[Nikita Klimenko [JB]]

https://github.com/Kotlin/dataframe targets JVM and should work on Android too

[altavir]

@Nikita Klimenko [JB] It is not a dataflow pipline.

[altavir]

@Yingding Wang Could you please clarify, what exactly do you need? DataForge is a pipline engine and it could work on Native, but it is a research library, so it should be refitted for each project.

[Yingding Wang]

@altavir I want to do transfer learning with tflite on android. I am building a small ml framework for kotlin to abstract away the tflite specific codes, something very similar to tfx (tensorflow extended) in python. Currently, I started to write some code to prototype. Thought if there is already a light-weighted dataflow or pipeline engine, then i can just reuse and integrate it into my android kotlin mlflow framework at first. Most case a dataflow or pipeline engine has already some orchestration mechanism that i can reuse and write some task container to trigger data processing or training task.

This kind of pipeline engine mustn’t contain its own DSL, a small in-memory scheduler can be just fine. Maybe also with some annotation mechanism so that i can extend tasks for android. Otherwise I will need to write all the stuff on my own. 🙂

[Yingding Wang]

@altavir Can you give me some more details to the DataForge pipeline engine? maybe an example that i can try it out?

[Yingding Wang]

I rephrased my question to “kotlin” instead of “kotlin native”, i think i can leave out the NNAPI on android or ios first 🙂 to make my life somewhat easier at the beginning

[altavir]

For TF-light you can just use TensorFlow API for Java. I think it will be enough.

As for DataForge, it is here: https://github.com/SciProgCentre/dataforge-core. It was developed for much more complicated workflows (particularly, data analysis in particle physics.

Here is an example of task definition: https://github.com/SciProgCentre/dataforge-core/blob/5406a6a64c872159d4cb03a5ee757[…]otlin/space/kscience/dataforge/workspace/SimpleWorkspaceTest.kt.

The problem is that it is a research library, so it is not fully "production ready" It needs to be adjusted for each case. You are welcome to collaborate on the development though.

[altavir]

On the bright side, it is much better suited to work in non-claster or even heterogeneous/polyglot environment than Spark.

[Yingding Wang]

thanks