Serve models

[favyen2]

Start exploring how to serve models.

[favyen2]

Todo:

• Do something to make sure we do not run multiple serving jobs for the same location.
• Have scripts to start Beaker jobs for all serving jobs that are not already completed for a given time range / dataset/model.
• Test the rslearn.serve code.

[favyen2]

btw this is just for batch mode of applying the models where we start lots of beaker jobs, each job covers a different part of the world, the job will acquire the image inputs needed and then run the model and upload the result to GCS

[favyen2]

Smaller tasks:

• Implement our own detection metric since the mAP one takes too long to compute.
• Implement rslearn.train.tasks.detection.DetectionTask.process_output
• Ensure rslp.serve works for wind turbine model for one window
• Add job launcher to run the model at every window for a particular timestamp (on Beaker)
• Add some sort of job tracking (maybe just based on output files) so that the job launcher can be executed multiple times without redoing work that was previously completed.
• Connect it to the existing remote_sensing_data_hub infrastructure (for now).

[favyen2]

Think about:

• Where should auxiliary code in model pipelines go? Some of this is per-location, that should go with the per-location job somehow. Some of this is global, like smoothing points over time, so that should be in some global post-processing place. All of it should be in rslp/ and should be testable and discoverable/deployable from one common format and all.