Train Unsupervised Embeddings with 3DUnet

[frandorr]

• Consolidate 2 year South England Dataset.
• Try training with ddp instead of dp (if using dp should check the code for modifications)
• Investigate Batch Norm in ddp (should use Sync Batch Norm?)
• Log the images with sigmoid applied
• Add metrics callbacks
• Retrain full model for many epochs
• Create unsupervised UNet Architecture to train with Moco
• Train Unet3d Encoder with Moco

[doloresgarcia]

• Explore what backbone we want to use using supervised
  • 3d-CNN (WIP) [tested as a dummy test but not continued ]
  • U-Net (WIP) [training]
• Use that backbone to train embeddings [having some bugs training on the vm]
• Use these embeddings to train a downstream tasks and compare performance [ will check this after training the embeddings]