Transition to Python and Serverless

[drewbo]

For fast release, this library was originally written in Node. In order to extend it's functionality, we will be gradually re-writing in Python:

• Better integration with the machine learning community
• Better support for non-image inputs (numpy arrays, GeoTIFF, sentinelhub-py to start)
• Replace kes.js with serverless to reduce our institutional maintenance burden.

We will also add easier configurability for the primary lambda function (download-and-predict) at the same time, though this functionality is largely language agnostic.

[moloned]

I'd recommend integrating fast.ai and Google colab for smaller entities and students working on EO data without a big budget for AWS

[drewbo]

Thanks @moloned, we will eventually move things more in that direction and currently do our trainings using those tools:

• This will currently support hosting models using fastai via fastai-serving
• Google colab imposes a connection time and RAM limit which can hinder running large-scale inference
• This should provide a relatively cost effective model. For classification models, we estimate that it costs about $10 per every million tiles/chips predicted