InstanceSpec

InstanceSpec()

Objects rather than classes: which object every pixel belongs to, scored as instances.

The same images and the same U-shaped networks as semantic_segmentation, with the masks read one file per object instead of united into one. Two fields are new on a model - separation, what the network predicts, and decode, a list of ways to turn those channels into objects - and pyplatypus.spec.instances is where both are documented, measurements included.

What this task is honestly worth, said here because the tag’s name promises more. A reader who writes task: instance_segmentation gets a semantic model with a separation target, decoded into objects and scored by the Data Science Bowl’s own average precision. On that dataset the separation signal is worth +0.068 at best while a perfect mask is worth +0.2398, so the limit is the mask rather than the separation - and connected_components, which ignores the signal entirely, is always available as the row every other row is read against. The capability is real, the ceiling is measured, and both are in the documentation rather than in a footnote.

Methods

Name	Description
nothing_watches_a_validation_number_that_will_not_arrive	A callback cannot wait for val_loss in a run that has no validation set.

nothing_watches_a_validation_number_that_will_not_arrive

InstanceSpec.nothing_watches_a_validation_number_that_will_not_arrive()

A callback cannot wait for val_loss in a run that has no validation set.