the CentroidContext mixin to train dendritic networks that use the centroid/prototype method to (a) construct a prototype context signal for training, and (b) infer a previously-seen context during inference,
the PermutedMNISTTaskIndices mixin that is only compatible with permutedMNIST experiments, and significantly speeds up initialization since no data transformations are applied when task indices are computed, and
(minor) new configs added to the projects/dendrites/permutedMNIST folder.
The train & evaluate functions in dendrite model utils have also been updated to accommodate the centroid setting.
This PR introduces
CentroidContext
mixin to train dendritic networks that use the centroid/prototype method to (a) construct a prototype context signal for training, and (b) infer a previously-seen context during inference,PermutedMNISTTaskIndices
mixin that is only compatible with permutedMNIST experiments, and significantly speeds up initialization since no data transformations are applied when task indices are computed, andprojects/dendrites/permutedMNIST
folder.The train & evaluate functions in dendrite model utils have also been updated to accommodate the centroid setting.