cellmap_flow.portable.fill_calls

What the browser calls for Label it’s box tools: the functions web/src/backend/routes/view-labels.ts’s Seed from Prediction, All Background, Split Objects and Undo run through Inference.python("cellmap_flow.portable.fill_calls:<name>", args, arrays), in the shape chunkmirage.calls says: fn(args, arrays), args the JSON arguments, arrays the bytes sent beside them, in order; a JSON value, or a pair of JSON and arrays, back. Each decodes its arrays, calls fill.py and encodes the answer; the maths is not here, and the server’s routes (dashboard/routes/finetune/view_labels.py, through finetune/session/fill.py) call the same functions around MinIO.

The route reads the box from the paint store and writes back what comes out, and keeps the undo stack, as the AI annotation’s accept does (ai_annotate_calls.accept): nothing here reads or keeps a volume. A box travels as the store holds it: shape (z, y, x) and dtype (the volume’s) in args, its bytes in C order, always arrays[0]. A write answers the box to write as its one array, or no array when nothing changes, so the route writes nothing and pushes no undo entry then.

A refusal the server answers 400 (new ids that do not fit in the volume’s dtype, a ValueError there) comes back as {"error": message}, as label_it_calls answers its own, for the route to answer as the server does. A ModuleNotFoundError (scipy, which Split Objects labels its components with) is never caught: it is how chunkmirage’s worker learns to load scipy, before it calls again.

Functions

box(args, arrays)	The annotation voxels a box in world nm covers (fill.box_voxels).
seed(args, arrays)	A seed written into the box's unannotated voxels: fill.seed_labels
background(args, arrays)	Every unannotated voxel of the box background (1), written as
split(args, arrays)	The box's foreground relabelled by connected component
undo(args, arrays)	The box put back to before an action where it still holds what the

Module Contents

cellmap_flow.portable.fill_calls.box(args, arrays)

The annotation voxels a box in world nm covers (fill.box_voxels). args: volume (its output_voxel_size and dataset_offset_nm), offset_nm and shape_nm (the box’s lower corner and size, z, y, x) and volume_shape (the volume’s voxels). Answers {"lo", "hi"}, or null when no voxel of the volume is in the box.

cellmap_flow.portable.fill_calls.seed(args, arrays)

A seed written into the box’s unannotated voxels: fill.seed_labels of the objects, written as fill.paint writes without overwrite. args: the box’s shape and dtype, and instance_target (the model trains on instances, so new ids count up on a volume of more than 8-bit ids: fill.counts_up). arrays: the box, then the objects over it (uint32, 0 background, as label_it_calls.seed answers them). Answers the counts (n_foreground, n_background, n_overwritten) with the box to write; or the refusal, when the new ids do not fit.

cellmap_flow.portable.fill_calls.background(args, arrays)

Every unannotated voxel of the box background (1), written as fill.paint writes without overwrite: All Background. args: the box’s shape and dtype; arrays: the box. Answers as seed.

cellmap_flow.portable.fill_calls.split(args, arrays)

The box’s foreground relabelled by connected component (fill.relabel_objects) and written over its foreground voxels (fill.rewrite): Split Objects. args: the box’s shape and dtype, and connectivity and per_slice, as the route has read them. arrays: the box. Answers the counts (objects, split, merged) and changed, the voxels whose label changed, with the box to write; or the refusal, when the split-off pieces’ new ids do not fit.

cellmap_flow.portable.fill_calls.undo(args, arrays)

The box put back to before an action where it still holds what the action left (fill.restore): Undo. args: the box’s shape and dtype; arrays: the box as it is now, then the undo entry’s before and after, each shaped as the box. Answers restored, the voxels put back, with the box to write.