Implements a multitask training API in .trainer.MultiTaskTrainer
Adds expected gradient estimation as a new interpretability scheme with .interpret.ExpectedGradient. We have tested expected gradients and found they are superior to our previous integrated gradient approach.
Adds an API for model interpretation in .api.scnym_interpret.
Enables training with a domain adversary on purely supervised data through the MultiTaskTrainer. For instance, you can now pass only a training set with domain labels and take advantage of the domain adversary, whereas previously our software required a target dataset to enable adaptation.
Adds balanced_classes and weighted_classes options to the scnym_api configuration options. These options may be useful for datasets with large class imbalances between the source and target domains, but we have not rigorously evaluated their impact on performance.
.trainer.MultiTaskTrainer.interpret.ExpectedGradient. We have tested expected gradients and found they are superior to our previous integrated gradient approach..api.scnym_interpret.MultiTaskTrainer. For instance, you can now pass only a training set with domain labels and take advantage of the domain adversary, whereas previously our software required a target dataset to enable adaptation.balanced_classesandweighted_classesoptions to thescnym_apiconfiguration options. These options may be useful for datasets with large class imbalances between the source and target domains, but we have not rigorously evaluated their impact on performance.