network todos

[kwagstyl]

To improve the network:

• nonlinear transformation layer from embedding to internal latent space to reorganise language embedding

• loss BCE

• output logits vs sigmoid

• check that the deconvolutions work - ?checkerboard artefact

• deep supervision

• other ways to improve existing architecture

• ?FILM layers

• augmentations - language (introduce those Alice) and mask

• intermediate layer to learn things

• skip connections

• symmetrical atlases need handling

• giving the MNI as an input (essential unet with additional info) (important step for native space)

• improved validation metrics (DICE), images?