In some cases we've seen SpaceNet competitors evaluate their loss functions on a channel-by-channel basis and sum the loss function outputs afterward. Particularly for implementations that don't scale to a 3rd dimension well (many focal loss implementations being a prime example), this is useful.
Describe the solution you'd like
The ability to provide an argument through the config yml file as well as in the function that collects and runs loss functions to evaluate on individual channels and then sum the outputs rather than across the entire image.
In some cases we've seen SpaceNet competitors evaluate their loss functions on a channel-by-channel basis and sum the loss function outputs afterward. Particularly for implementations that don't scale to a 3rd dimension well (many focal loss implementations being a prime example), this is useful.
Describe the solution you'd like The ability to provide an argument through the config yml file as well as in the function that collects and runs loss functions to evaluate on individual channels and then sum the outputs rather than across the entire image.