Training plan

[APJansen]

Once the other issues are addressed, I propose that we decide how we split the data into train/val/test, and then run some hyperparameter searches for the supervised learning of:

• local model:
  • number of layers
  • units per layer
  • learning rate
• 2D UNet
  • number of blocks
  • start channels
  • batch norm
  • learning rate
  • kernel size?
  • block depth?
• 3D Unet
  • as above

Perhaps all with learning rate decay as well.

This will be an interesting comparison in itself, will hopefully give us a model that performs well at least on similar samples. It should close the supervised part and be a good starting point for unsupervised fine tuning.

[chyalexcheng]

Sounds like a very good plan. Parallel to hypertuning, I will test the synthetic dataset from simulations.