AML (add-on functionality / should still be able to do the same things with the local codebase):

• AML training pipeline (mostly done, just needs to run)
• introduce data versioning functionality (through AML)
  • snapshot output of prepare_* scripts along with the experiment run in the outputs folder in each experiment (the way TensorFlow logs are currently being snapshotted)

Models:

• add back Dutch section-based dataset (better performance for high-resolution models, faster scoring)
  • high-res models like HRNet / SEResNet Unet

Data:

• Parihaka as a stretch goal
• entering Parihaka-based competition