Rather than provide specific configuration options, we'll provide a host of utility functions for the primary functions needed. Users will still be expected to write their own lambda function with a consistent input signature (sqs records containing a tile body).
- Inputs: tile endpoint, geotiff, ndarray, with various compression options
- Tiling: should support the user sending XYZ input tiles (anything else?) into the SQS queue, acquiring imagery based on these bounds, and then subdividing these prior to GPU prediction as desired ("megatiles")
- Prediction: wrapping/handling of the requests to the TFServing/fastai-serving endpoint
- Saving: support save to S3 or RDS in various formats (jsonb, single value, image, compressed image)
Rather than provide specific configuration options, we'll provide a host of utility functions for the primary functions needed. Users will still be expected to write their own lambda function with a consistent input signature (sqs records containing a tile body).