Add ability to regress towards a heatmap / distilation / student-teacher networks.
This issue addresses the roadmap item:
-
Distilation: Easy distilation (i.e. student / teacher networks) by settting heatmaps from existing predictions as truth targets.
The idea is that we can specify "some process" that generates a heatmap, and then we train a new network to regress towards that predicted heatmap.
Supporting this will include:
- Specifying the channels in a kwcoco file to regress towards
- Sampling those channels in the dataloader and passing them as "truth" objects in the output batch items
- Adding a head that can regress towrads a target with the same channel specification.