How much of each class a mask covers

Description

A quick sanity check before training: a foreground of half a percent is a problem the loss has to be chosen for, and it is better to know before the run than after.

Usage

mask_coverage(mask, labels = NULL)

Arguments

mask	Class indices, counted from 1.
labels	Optional names for the classes.

Value

A data frame with one row per class.

Examples

library("rplatypus")

mask <- matrix(1L, 8, 8); mask[2:4, 2:4] <- 2L
mask_coverage(mask, labels = c("background", "nucleus"))

  class      label pixels fraction
1     1 background     55 0.859375
2     2    nucleus      9 0.140625

# A class with 0 pixels is the thing to look for: its target channel is zero
# everywhere, so there is no gradient towards it and the model is never shown
# what it is being asked to find.