Feature Catalogue

Plug-AI allows you to configure a complete pipeline using options from a predefined catalogue. You can independantly select and configure:

  • a dataset and dataloader

  • a model to train or use

  • an optimizer

  • a learning rate scheduler

  • a criterion (loss)

  • a metric for evaluation

  • a pipeline (training, evaluation, inference)

Datasets and dataloader

Models

  • DynUnet

  • unetr

  • ModSegNet

nnU-Net is also available but follows a special pipeline non-compatible with the rest of the options.

Optimizers

By default, the optimizer used is SGD but you can decide to use any of these following optimizers:

  • Adadelta,

  • Adagrad,

  • Adam,

  • AdamW,

  • Adamax,

  • ASGD,

  • NAdam,

  • RAdam,

  • RMSprop,

  • Rprop,

  • SGD

Learning rate schedulers

By default, the learning rate is static and no learning rate scheduler is used but you can decide to use any of the following schedulers:

  • StepLR

  • MultiStepLR

  • ConstantLR

  • LinearLR

  • ExponentialLR

  • CosineAnnealingLR

  • ReduceLROnPlateau

  • CyclicLR

  • OneCycleLR

  • CosineAnnealingWarmRestarts

Criterions (loss)

The available loss are:

  • DiceLoss

  • GeneralizedDiceLoss

  • DiceCELoss

  • DiceFocalLoss

  • GeneralizedDiceFocalLoss

  • FocalLoss

  • TverskyLoss

Metrics for evaluation

The available metrics are:

  • MeanDice

  • MeanIoU

Pipelines (training, evaluation, inference)

  • Training

  • Evaluation

  • Inference