v2.2

API updates:

• Add support for adding not given modality as None
• Add support for additional arguments in nn.FeaturedData and nn.MAGNET2
• Introducing builder package
• Rename Mid*Fusion module to Mean*Fusion
• Skip loss calculation in losses.MAGLoss when a modality is not given

Other updates:

• Minor bugs fixed
• Typing improvement

v2.1.1

Updates:

• Typing improvement

v2.1

API Updates:

• Implement torchmanager_monai.Manager as general manager for monai
• Introduced losses.MAGMSLoss
• Missing modalities training support added during unpacking data in MonaiManager
• Monai dependency is now optional with torchmanager_monai as its extra package (MonaiManager, networks.UNETR, and networks.UNETRWithMultiModality will be defined as NotImplemented)
• Rename magnet.managers.monai.Manager to magnet.manager.monai.SegmentationManager
• Rename torchmanager_monai.MonaiManager to SegmentationManager which inherits torchmanager_monai.Monai

Other updates:

• Deprecated HeMIS implementation
• Minor bugs fixed
• Remove losses.protocols.DistillatedData

v2.0.2

Updates:

• Change the default value of return_features in build_v2_unet function to True

v2.0.1

Updates:

• Deprecated HeMIS implementation
• Minor bugs fixed
• Remove losses.protocols.DistillatedData

v2.0

API updates:

• Add self distillation losses for MAG-MS framework (losses.MAGFeatureDistillationLoss and losses.MAGSelfDistillationLoss)
• Add a copy_encoder boolean flag to control if use the same initialized weights for modality specific encoders
• Introducing MAGNET2 architecture with feature fusion
• Handling available targets options in HeMIS when target_dict does not give all available modalities but inputs contain all the modalities
• Python framework renamed to magms
• The target property can now be set as a list of target modalities in both TargetingManager and MAGNET

v1.1.3

Updates:

• Typing improvement

v1.1.2

Updates:

• Handling a list of torch.Tensor when aggregating a CumulativeIteration monai metric.
• Minor bugs fixed

v1.1.1

Updates:

• Minor bugs fixed
• Targets control improvement

v1.1

API updates:

• A traditional torch.data.Dataset and torch.data.DataLoader can now be used to train with all modalities inputs instead of using magnet.data.TargetingDataset and magnet.data.TargetingDataLoader
• Add a copy_modality boolean flag control for build method
• Add HeMIS network implementation as a comparison
• Deprecate UNETRWithDictOutput for dictionary outputs
• Introducing magnet.losses packages where magnet.losses.MAGLoss can be used to combine the training of Magnet in one iteration
• Introducing networks.unetr package
• Targeting protocol now accepts optional keys in target_dict property for all modalities
• The target property can now be set as a list of target modalities in both TargetingManager and MAGNET