neu_vol.ops.relabel¶

Give every occupied region of a sparse volume its own range of label ids.

Ground truth annotated chunk by chunk comes back with each chunk numbered from 1, so the same integer means a different cell in every chunk, and a mesh of one id is a dozen unrelated cells. This renumbers each occupied region into a range of its own.

The regions come from stored-chunk occupancy (neu_vol.ops.occupancy), so they are pairwise disjoint; the run is serial, each range starting where the last ended.
Boxes are not tightened: mode downsampling can drop a stray voxel, so a tightened box could leave level-0 data with its old id. The chunk-aligned box covers every stored chunk, and its writes are whole chunks.
The mapping is an output, written by default: it is the only way back from a new id to its region and original label.

Single-scale, like neu_vol.ops.write: run neu-vol downsample --start-level <level> afterwards.

Attributes¶

DEFAULT_MAX_REGION_BYTES

Functions¶

`apply_relabel`(plan, *[, dry_run, overwrite, map_path]) → dict	Carry out a `plan_relabel()`, returning the mapping and per-region summary.
`default_map_path`(destination, level) → str	`<last path component>.relabel-<level>.json`, in the working directory.
`plan_relabel`(volume, *[, out, in_place, level, ...]) → dict	Resolve the geometry of the relabel. Reads chunk keys only; writes nothing.
`relabel`(volume, *[, out, in_place, level, block_size, ...]) → dict	Plan and apply in one call. See `plan_relabel()` and `apply_relabel()`.

Module Contents¶

neu_vol.ops.relabel.apply_relabel(plan: dict, *, dry_run: bool = False, overwrite: bool = False, map_path: str | None = None) → dict¶

Carry out a plan_relabel(), returning the mapping and per-region summary.

With dry_run the reads still happen — the mapping is only knowable from the voxels — but nothing is written and no destination is created.

neu_vol.ops.relabel.default_map_path(destination: str, level: int) → str¶

<last path component>.relabel-<level>.json, in the working directory.

A default, so the mapping cannot be lost by forgetting a flag.

neu_vol.ops.relabel.plan_relabel(volume: str, *, out: str | None = None, in_place: bool = False, level: int = 0, block_size: int | None = None, max_region_bytes: int = DEFAULT_MAX_REGION_BYTES, fmt: str | None = None) → dict¶

Resolve the geometry of the relabel. Reads chunk keys only; writes nothing.

The id mapping needs the voxels, so apply_relabel() resolves it (and reports it without writing when dry_run=True).

neu_vol.ops.relabel.relabel(volume: str, *, out: str | None = None, in_place: bool = False, level: int = 0, block_size: int | None = None, max_region_bytes: int = DEFAULT_MAX_REGION_BYTES, dry_run: bool = False, overwrite: bool = False, map_path: str | None = None) → dict¶: Plan and apply in one call. See plan_relabel() and apply_relabel().

neu_vol.ops.relabel.DEFAULT_MAX_REGION_BYTES = 8589934592¶