cellmap_flow.portable.targets_calls

What the browser calls for training targets: the function web/src/finetune/target-call.ts runs through chunkmirage’s Engine.call("cellmap_flow.portable.targets_calls:targets", args, arrays) (calls.py says the shape). It decodes the chunk’s labels, calls the maths in targets.py and encodes the answer; the maths is not here.

Functions

targets(args, arrays)

A painted chunk's target and mask (targets.binary, or

Module Contents

cellmap_flow.portable.targets_calls.targets(args, arrays)

A painted chunk’s target and mask (targets.binary, or targets.affinity for an affinity model). args: shape (z, y, x), labels_dtype, channels (the model’s output channels), and either offsets (null for a binary model, else one [dz, dy, dx] or null per channel from the first, as a caller that knows them gives them) or channel_names, from which the server’s rule decides (label_it.affinity_offsets: a model with a *_aff* channel is an affinity model, its first affinity channels along z, y and x), so the browser, which has only its manifest’s names, trains on what the server would. select_channel (absent or null for every channel) is the one channel of a binary model the paint is for: its target alone, one channel, as the server’s cli makes it for --select-channel (refused, as there, for an affinity model, each of whose channels has a target of its own). arrays: the chunk’s labels (z, y, x). Answers the channels with the target and the mask, each float32 (channels, z, y, x).