patchworks 0.11.17__tar.gz → 0.11.18__tar.gz

{patchworks-0.11.17 → patchworks-0.11.18}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: patchworks
-Version: 0.11.17
+Version: 0.11.18
 Summary: Tiled processing of arbitrarily large images with globally consistent labels
 Project-URL: Homepage, https://github.com/imcf/patchworks
 Project-URL: Issues, https://github.com/imcf/patchworks/issues
@@ -34,7 +34,7 @@ Requires-Dist: glasbey; extra == 'all'
 Requires-Dist: imaris-ims-file-reader; extra == 'all'
 Requires-Dist: ipykernel<7; extra == 'all'
 Requires-Dist: lxml-html-clean; extra == 'all'
-Requires-Dist: napari[all]>=0.5.5; extra == 'all'
+Requires-Dist: napari[all]>=0.7.1; extra == 'all'
 Requires-Dist: numpy<2.5; extra == 'all'
 Requires-Dist: nvidia-ml-py; extra == 'all'
 Requires-Dist: psutil; extra == 'all'
@@ -74,7 +74,7 @@ Provides-Extra: napari
 Requires-Dist: glasbey; extra == 'napari'
 Requires-Dist: ipykernel<7; extra == 'napari'
 Requires-Dist: lxml-html-clean; extra == 'napari'
-Requires-Dist: napari[all]>=0.5.5; extra == 'napari'
+Requires-Dist: napari[all]>=0.7.1; extra == 'napari'
 Requires-Dist: numpy<2.5; extra == 'napari'
 Provides-Extra: workflow
 Requires-Dist: snakemake-executor-plugin-slurm; extra == 'workflow'

{patchworks-0.11.17 → patchworks-0.11.18}/docs/guide/snakemake.md

@@ -194,6 +194,129 @@ also re-runs a step when its **code, params or software environment** change —
 so upgrading patchworks would re-do the conversion and overwrite an existing
 result. Keep `mtime` and reruns happen only when an output is missing or stale.
 
+## Running two segmentations (e.g. nuclei + cytoplasm)
+
+Every path the workflow writes — `tiles.json`, `stage.zarr`, per-tile `seg/`,
+the cached model, `labels.done` — lives under `work_dir/<label_name>/`, so
+running the workflow **twice with two configs against the same `work_dir`**
+never collides: each run gets its own private subdirectory, and both reuse
+the *same* already-converted `image.zarr` (conversion never re-runs).
+
+```yaml
+# config/config_nuclei.yaml
+input: "/data/scan.ims"
+work_dir: "/scratch/results"
+label_name: "nuclei_labels"
+channel: 1                # nuclear stain channel
+tile_shape: [16, 1024, 1024]
+cellpose:
+  model: "nuclei"
+  diameter: 15
+  do_3D: true
+```
+
+```yaml
+# config/config_cyto.yaml
+input: "/data/scan.ims"
+work_dir: "/scratch/results"   # same work_dir — image.zarr is reused
+label_name: "cyto_labels"
+channel: 0                # cytoplasm/membrane channel
+tile_shape: [16, 1024, 1024]   # keep this identical across configs — see below
+cellpose:
+  model: "cyto3"
+  diameter: 30
+  do_3D: true
+```
+
+Run them one after another (or as two independent SLURM submissions, even
+concurrently — they touch disjoint files):
+
+```bash
+snakemake --workflow-profile profile/slurm --configfile config/config_nuclei.yaml
+snakemake --workflow-profile profile/slurm --configfile config/config_cyto.yaml
+```
+
+Both land side by side in the same store:
+
+```text
+results/image.zarr/labels/nuclei_labels/
+results/image.zarr/labels/cyto_labels/
+```
+
+!!! tip "Keep `tile_shape` (and `level`) identical across configs"
+    Different segmentations of the same image can use different `channel` and
+    `cellpose:` settings freely, but keep `tile_shape`/`level` the same across
+    configs — the label arrays then share the exact same chunk layout, which
+    `label_relations()` (below) requires.
+
+### Relating labels across segmentations
+
+Once you have two segmentations of the same image (e.g. nuclei inside cells),
+`label_relations()` maps each label in one to the label it overlaps most in
+the other — by streaming both arrays chunk by chunk, so it scales to
+hundreds of thousands of objects without loading anything fully into RAM:
+
+```python
+import dask.array as da
+from patchworks import label_relations
+
+nuclei = da.from_zarr("results/image.zarr", component="labels/nuclei_labels/0")
+cells = da.from_zarr("results/image.zarr", component="labels/cyto_labels/0")
+
+table = label_relations(nuclei, cells)
+table[2]
+# {'match': 3, 'overlap_voxels': 4821, 'overlap_fraction': 0.94}
+# -> nucleus 2 belongs to cell 3, 94% of its voxels fall inside it
+```
+
+Save it as a table:
+
+```python
+import csv
+
+with open("nuclei_to_cell.csv", "w", newline="") as f:
+    w = csv.writer(f)
+    w.writerow(["nucleus_id", "cell_id", "overlap_voxels", "overlap_fraction"])
+    for nucleus_id, m in table.items():
+        w.writerow([nucleus_id, m["match"], m["overlap_voxels"], m["overlap_fraction"]])
+```
+
+## Measurements (fast, whole-volume regionprops)
+
+`skimage.measure.regionprops` needs the full labelled + intensity array in
+RAM — fine for one tile, not for a hundred-thousand-object OME-ZARR. Use
+[`dask-image`](https://image.dask.org)'s `ndmeasure`, which computes directly
+on the dask/zarr-backed arrays, chunk-parallel, without materializing the
+volume:
+
+```bash
+pip install dask-image
+```
+
+```python
+import dask.array as da
+from dask_image.ndmeasure import area, center_of_mass, mean, standard_deviation
+
+labels = da.from_zarr("results/image.zarr", component="labels/cyto_labels/0")
+image = da.from_zarr("results/image.zarr", component="0")[0]  # channel 0, level 0
+
+ids = da.unique(labels[labels > 0]).compute()
+areas = area(image, labels, ids).compute()               # voxel counts
+means = mean(image, labels, ids).compute()                # mean intensity
+stds = standard_deviation(image, labels, ids).compute()
+centroids = center_of_mass(image, labels, ids).compute()  # voxel coords (z, y, x)
+```
+
+Multiply `areas` by the voxel's physical volume and `centroids` by the pixel
+size (both read straight from the OME-ZARR's own `multiscales` metadata) to
+get µm-scale measurements.
+
+For interactively inspecting individual cells by clicking in the viewer
+(not all objects at once), the
+[napari-skimage-regionprops](https://github.com/haesleinhuepf/napari-skimage-regionprops)
+plugin's table widget works well — point it at a cropped region rather than
+the full volume, since it loads its input fully into memory.
+
 ## Custom segmentation function
 
 Not using Cellpose? Run **your own** per-tile function — no need to edit the

{patchworks-0.11.17 → patchworks-0.11.18}/pyproject.toml

@@ -59,15 +59,16 @@ bioio = [
 # imaris reads .ims files natively (HDF5, no JVM) for OME-ZARR conversion.
 imaris = ["imaris-ims-file-reader"]
 # napari enables the interactive viewer plugin.
-# - napari >= 0.5: NumPy 2.0 compatible (older 0.4.x uses np.array(copy=False),
-#   which raises under NumPy 2 -> ValidationError on Viewer()).
+# - napari >= 0.7.1: adds locked_data_level (manual multiscale resolution pin
+#   for 2D *and* 3D) — before this, 3D view always showed the coarsest
+#   pyramid level unconditionally, with no way to change it.
 # - numpy < 2.5: napari/numba don't support numpy 2.5 yet.
 # - ipykernel < 7: napari-console requires ipykernel < 7.
 # - lxml-html-clean: napari's notebook_display imports lxml.html.clean, split
 #   into a separate package in lxml >= 5.2 (else ImportError on Viewer()).
 # - glasbey: distinct high-contrast label LUTs for view_in_napari.
 napari = [
-    "napari[all]>=0.5.5",
+    "napari[all]>=0.7.1",
     "numpy<2.5",
     "ipykernel<7",
     "lxml-html-clean",

{patchworks-0.11.17 → patchworks-0.11.18}/src/patchworks/__init__.py

@@ -36,6 +36,7 @@ from ._distributed import create_stage, spatial_tiles, stage_tile
 from ._io import estimate_empty_tiles, load_ome_zarr
 from ._merge import merge_tile_labels
 from ._relabel import relabel_sequential_array, relabel_sequential_zarr
+from ._relations import label_relations
 
 try:
     __version__ = _pkg_version("patchworks")
@@ -52,6 +53,7 @@ __all__ = [
     "make_local_cluster",
     "relabel_sequential_array",
     "relabel_sequential_zarr",
+    "label_relations",
     "spatial_tiles",
     "create_stage",
     "stage_tile",

patchworks-0.11.18/src/patchworks/_relations.py

@@ -0,0 +1,142 @@
+"""Relate two label images by voxel overlap (e.g. nucleus -> containing cell)."""
+
+from __future__ import annotations
+
+import logging
+import os
+from concurrent.futures import ThreadPoolExecutor
+from pathlib import Path
+from typing import Union
+
+import dask.array as da
+import numpy as np
+
+logger = logging.getLogger(__name__)
+
+
+def _as_dask(source: Union["da.Array", str, Path], component: str) -> "da.Array":
+    if isinstance(source, (str, Path)):
+        return da.from_zarr(str(source), component=component)
+    return source
+
+
+def _chunk_pairs(a_block: np.ndarray, b_block: np.ndarray) -> np.ndarray:
+    """Non-background ``(a_id, b_id) -> voxel count`` rows for one chunk pair."""
+    mask = (a_block > 0) & (b_block > 0)
+    if not mask.any():
+        return np.empty((0, 3), dtype=np.int64)
+    pairs = np.stack([a_block[mask], b_block[mask]], axis=1).astype(np.int64)
+    uniq, counts = np.unique(pairs, axis=0, return_counts=True)
+    return np.concatenate([uniq, counts[:, None]], axis=1)
+
+
+def label_relations(
+    a: Union["da.Array", str, Path],
+    b: Union["da.Array", str, Path],
+    *,
+    a_component: str = "labels",
+    b_component: str = "labels",
+    n_workers: int | None = None,
+) -> dict[int, dict[str, float]]:
+    """Map each label in *a* to its best-overlapping label in *b*.
+
+    For every non-background label in *a* (e.g. a nucleus segmentation),
+    finds the label in *b* (e.g. a cell/cytoplasm segmentation of the same
+    image) it shares the most voxels with. Streams both arrays chunk by
+    chunk — memory is bounded by the number of distinct label pairs (one row
+    per touching (a, b) pair per chunk), not by volume size.
+
+    *a* and *b* must be two segmentations of the **same image** (identical
+    shape and chunk layout) — e.g. two runs of the Snakemake workflow with
+    different ``label_name``/``cellpose:`` config but the same ``tile_shape``.
+
+    Parameters
+    ----------
+    a, b : da.Array, str or Path
+        Two label arrays of identical shape and chunking. A path is read via
+        ``dask.array.from_zarr(path, component=...)``.
+    a_component, b_component : str, optional
+        Zarr array name inside *a*/*b* when they're store paths (default
+        ``"labels"``).
+    n_workers : int or None, optional
+        Parallel chunk workers. Default ``min(4, cpu_count)``.
+
+    Returns
+    -------
+    dict
+        ``{a_label: {"match": b_label, "overlap_voxels": int,
+        "overlap_fraction": float}}`` — one entry per *a* label that touches
+        at least one non-background *b* voxel. ``overlap_fraction`` is the
+        matched voxel count over *a* label's total voxel count (1.0 = fully
+        contained). Labels in *a* with zero overlap are omitted.
+
+    Examples
+    --------
+    >>> from patchworks import label_relations
+    >>> table = label_relations(
+    ...     "scan.zarr/labels/nuclei", "scan.zarr/labels/cells"
+    ... )  # doctest: +SKIP
+    >>> table[2]  # nucleus 2 sits inside cell 3  # doctest: +SKIP
+    {'match': 3, 'overlap_voxels': 4821, 'overlap_fraction': 0.94}
+    """
+    a = _as_dask(a, a_component)
+    b = _as_dask(b, b_component)
+    if a.shape != b.shape:
+        raise ValueError(f"shape mismatch: a={a.shape} b={b.shape}")
+    if a.chunks != b.chunks:
+        raise ValueError(
+            "a and b must share the same chunk layout "
+            f"(a={a.chunks} b={b.chunks}); rechunk one to match the other, "
+            "e.g. b = b.rechunk(a.chunks)"
+        )
+
+    n_blocks = a.numblocks
+    total = int(np.prod(n_blocks))
+    nw = n_workers if n_workers is not None else min(4, os.cpu_count() or 1)
+
+    def _one(flat_idx: int) -> np.ndarray:
+        idx = np.unravel_index(flat_idx, n_blocks)
+        return _chunk_pairs(np.asarray(a.blocks[idx]), np.asarray(b.blocks[idx]))
+
+    with ThreadPoolExecutor(max_workers=nw) as ex:
+        parts = list(ex.map(_one, range(total)))
+
+    rows = [p for p in parts if p.size]
+    if not rows:
+        return {}
+    all_pairs = np.concatenate(rows, axis=0)
+
+    # Merge duplicate (a_id, b_id) rows across chunks (a label can span
+    # several chunks) by sorting on a combined key and summing runs.
+    key = all_pairs[:, 0] * (int(all_pairs[:, 1].max()) + 1) + all_pairs[:, 1]
+    order = np.argsort(key, kind="stable")
+    all_pairs, key = all_pairs[order], key[order]
+    starts = np.concatenate([[0], np.flatnonzero(np.diff(key)) + 1])
+    merged_counts = np.add.reduceat(all_pairs[:, 2], starts)
+    merged = np.stack(
+        [all_pairs[starts, 0], all_pairs[starts, 1], merged_counts], axis=1
+    )
+
+    a_totals: dict[int, int] = {}
+    for a_id, _, count in merged:
+        a_id = int(a_id)
+        a_totals[a_id] = a_totals.get(a_id, 0) + int(count)
+
+    best: dict[int, tuple[int, int]] = {}
+    for a_id, b_id, count in merged:
+        a_id, b_id, count = int(a_id), int(b_id), int(count)
+        cur = best.get(a_id)
+        if cur is None or count > cur[1]:
+            best[a_id] = (b_id, count)
+
+    logger.info(
+        "label_relations: %d a-labels matched across %d chunks", len(best), total
+    )
+    return {
+        a_id: {
+            "match": b_id,
+            "overlap_voxels": count,
+            "overlap_fraction": count / a_totals[a_id],
+        }
+        for a_id, (b_id, count) in best.items()
+    }

{patchworks-0.11.17 → patchworks-0.11.18}/src/patchworks/plugins/napari.py

@@ -142,12 +142,14 @@ def _resolve_image(
     return source
 
 
-def _pyramid_scale(path: Union[str, Path], ndim: int) -> list[float] | None:
-    """Read the level-0 physical scale (µm/px) from an OME-ZARR's own metadata.
+def _pyramid_calibration(
+    path: Union[str, Path], ndim: int
+) -> tuple[list[float], list[str]] | tuple[None, None]:
+    """Read the level-0 physical scale and units from an OME-ZARR's metadata.
 
-    Without this, napari shows every axis with an implicit scale of 1, so a
-    volume with anisotropic voxels (e.g. z coarser than x/y) renders with the
-    wrong aspect ratio and no real-world units.
+    Without this, napari shows every axis with an implicit scale of 1 and
+    "pixel" units, so a volume with anisotropic voxels (e.g. z coarser than
+    x/y) renders with the wrong aspect ratio and no real-world units.
 
     Parameters
     ----------
@@ -161,15 +163,20 @@ def _pyramid_scale(path: Union[str, Path], ndim: int) -> list[float] | None:
 
     Returns
     -------
-    list of float or None
-        Physical size per axis, or ``None`` if the store carries no
-        calibration (napari then falls back to its uncalibrated default).
+    tuple
+        ``(scale, units)`` — physical size and unit name per axis — or
+        ``(None, None)`` if the store carries no calibration (napari then
+        falls back to its uncalibrated pixel default).
     """
     from .ome_zarr import _base_scale, _default_axes, _read_zarr_calibration
 
     axes = _default_axes(ndim)
     pixel_size = _read_zarr_calibration(str(path), axes)
-    return _base_scale(axes, pixel_size) if pixel_size else None
+    if not pixel_size:
+        return None, None
+    scale = _base_scale(axes, pixel_size)
+    units = ["micrometer"] * len(scale)
+    return scale, units
 
 
 def _inner_label_names(store: Union[str, Path]) -> list[str]:
@@ -271,6 +278,18 @@ def view_in_napari(
     napari.Viewer
         The viewer instance (useful when ``show=False``).
 
+    Notes
+    -----
+    In 3-D view (the cube icon, or ``viewer.dims.ndisplay = 3``), napari
+    always shows the **coarsest** pyramid level — there is no automatic
+    zoom-based switching in 3-D, only in 2-D. To pin a specific resolution
+    (needs ``napari>=0.7.1``)::
+
+        viewer.layers["labels"].locked_data_level = 0   # full resolution
+        viewer.layers["labels"].locked_data_level = None  # back to coarsest
+
+    A widget for this is also in the layer controls panel in napari>=0.7.1.
+
     Examples
     --------
     >>> view_in_napari("scan.zarr")  # auto-loads scan.zarr/labels/*  # doctest: +SKIP
@@ -279,13 +298,16 @@ def view_in_napari(
 
     img = _resolve_image(image, channel)
     img_ndim = img[0].ndim if isinstance(img, list) else img.ndim
-    img_scale = _pyramid_scale(image, img_ndim) if _is_zarr(image) else None
+    img_scale, img_units = (
+        _pyramid_calibration(image, img_ndim) if _is_zarr(image) else (None, None)
+    )
     viewer = napari.Viewer()
     viewer.add_image(
         img,
         name=image_name,
         multiscale=isinstance(img, list),
         scale=img_scale,
+        units=img_units,
         **add_image_kwargs,
     )
 
@@ -322,12 +344,17 @@ def view_in_napari(
     if labels is not None:
         lab = _resolve_labels(labels, labels_component)
         lab_ndim = lab[0].ndim if isinstance(lab, list) else lab.ndim
-        lab_scale = _pyramid_scale(labels, lab_ndim) if _is_zarr(labels) else None
+        lab_scale, lab_units = (
+            _pyramid_calibration(labels, lab_ndim)
+            if _is_zarr(labels)
+            else (None, None)
+        )
         viewer.add_labels(
             lab,
             name=labels_name,
             multiscale=isinstance(lab, list),
             scale=lab_scale,
+            units=lab_units,
             **label_kwargs,
         )
     elif _is_zarr(image):
@@ -341,11 +368,13 @@ def view_in_napari(
             store = f"{image}/labels/{name}"
             levels = _multiscale_levels(store, None)
             lab = [lvl.astype("int32") for lvl in levels]
+            lab_scale, lab_units = _pyramid_calibration(store, lab[0].ndim)
             viewer.add_labels(
                 lab,
                 name=name,
                 multiscale=True,
-                scale=_pyramid_scale(store, lab[0].ndim),
+                scale=lab_scale,
+                units=lab_units,
                 **label_kwargs,
             )
             logger.info("auto-loaded labels/%s from %s", name, image)

patchworks-0.11.18/tests/test_relations.py

@@ -0,0 +1,43 @@
+import dask.array as da
+import numpy as np
+import pytest
+
+from patchworks import label_relations
+
+
+def test_label_relations_majority_overlap():
+    # nucleus 1 sits fully inside cell 10; nucleus 2 straddles cells 20/21
+    # with more voxels in 20 -> should match 20, not 21.
+    a = np.zeros((4, 10), dtype=np.int32)
+    a[0:2, 0:2] = 1
+    a[0:2, 4:8] = 2
+    b = np.zeros((4, 10), dtype=np.int32)
+    b[0:3, 0:3] = 10
+    b[0:2, 4:6] = 20  # 4 voxels overlap with label 2
+    b[0:2, 6:8] = 21  # 4 voxels overlap with label 2 -> tie, but let's skew it
+    b[0:2, 6:7] = 20  # tip the tie: label 20 now has 6 voxels vs 21's 2
+
+    table = label_relations(da.from_array(a, chunks=(2, 5)), da.from_array(b, chunks=(2, 5)))
+
+    assert table[1]["match"] == 10
+    assert table[1]["overlap_voxels"] == 4
+    assert table[1]["overlap_fraction"] == 1.0
+
+    assert table[2]["match"] == 20
+    assert table[2]["overlap_voxels"] == 6
+
+
+def test_label_relations_no_overlap_omitted():
+    a = np.zeros((2, 2), dtype=np.int32)
+    a[0, 0] = 1
+    b = np.zeros((2, 2), dtype=np.int32)  # all background, no overlap anywhere
+
+    table = label_relations(da.from_array(a, chunks=(2, 2)), da.from_array(b, chunks=(2, 2)))
+    assert table == {}
+
+
+def test_label_relations_chunk_mismatch_raises():
+    a = da.zeros((4, 4), chunks=(2, 4), dtype=np.int32)
+    b = da.zeros((4, 4), chunks=(4, 4), dtype=np.int32)
+    with pytest.raises(ValueError, match="chunk layout"):
+        label_relations(a, b)

{patchworks-0.11.17 → patchworks-0.11.18}/workflow/README.md

@@ -75,6 +75,17 @@ The GPU request lives in `profile/slurm/config.yaml` under
 `set-resources: segment:` (`--gres=gpu:1`). Raise `jobs:` to use more GPUs at
 once.
 
+## Multiple segmentations (nuclei + cytoplasm, …)
+
+Every intermediate path is namespaced under `work_dir/<label_name>/`, so
+running the workflow **twice with two configs** (different `label_name`,
+`channel`, `cellpose:`) against the **same `work_dir`** is safe — both reuse
+the one converted `image.zarr` and land side by side in
+`image.zarr/labels/<label_name>/`. See the
+[full guide](https://imcf.one/patchworks/guide/snakemake/#running-two-segmentations-eg-nuclei--cytoplasm)
+for the two-config recipe and `patchworks.label_relations()` for mapping one
+segmentation onto the other (e.g. nucleus → containing cell).
+
 ## Output
 
 `<work_dir>/image.zarr` — the image plus `labels/<label_name>/` (multi-scale,

{patchworks-0.11.17 → patchworks-0.11.18}/workflow/Snakefile

@@ -26,4 +26,4 @@ localrules:
 
 rule all:
     input:
-        f"{WORK}/labels.done",
+        f"{RUN}/labels.done",

{patchworks-0.11.17 → patchworks-0.11.18}/workflow/config/config.yaml

@@ -30,6 +30,10 @@ empty_threshold: null # null → Otsu; or a number
 
 # ---- segmentation -----------------------------------------------------------
 method: "cellpose" # "cellpose" (GPU), "threshold" (no GPU; testing), "custom"
+# Also namespaces this run's intermediate files under work_dir/<label_name>/,
+# so a second segmentation (different label_name, e.g. nuclei vs cytoplasm)
+# can safely target the same work_dir — see docs/guide/snakemake.md
+# "Running two segmentations".
 label_name: "cellpose_labels"
 cellpose:
   model: "nuclei"

{patchworks-0.11.17 → patchworks-0.11.18}/workflow/rules/common.smk

@@ -7,8 +7,17 @@ IMAGE = f"{WORK}/image.zarr"
 # directory) lets Snakemake skip conversion when the store already exists and
 # avoids wiping the whole store on a re-run (same trick as imcf/sopa).
 IMAGE_OK = f"{IMAGE}/zarr.json"
-TILES = f"{WORK}/tiles.json"
-STAGE = f"{WORK}/stage.zarr"
+
+# Everything below is per-segmentation, namespaced under WORK/<label_name>/, so
+# running the workflow twice with two configs (different label_name, e.g.
+# "nuclei_labels" and "cell_labels") against the *same* work_dir never
+# collides — each gets its own tiles/stage/seg/model/labels.done, and both
+# read the *same* already-converted image.zarr. See docs/guide/snakemake.md
+# "Running two segmentations" for the two-config recipe.
+LABEL_NAME = config.get("label_name", "labels")
+RUN = f"{WORK}/{LABEL_NAME}"
+TILES = f"{RUN}/tiles.json"
+STAGE = f"{RUN}/stage.zarr"
 # Completion sentinel for the stage store. Tracking a touch()ed marker instead
 # of directory(STAGE) keeps Snakemake from deleting/recreating the store on a
 # re-run and avoids directory-mtime quirks (same touch() discipline as sopa).
@@ -17,16 +26,18 @@ STAGE_OK = f"{STAGE}.done"
 
 # Logs: one shared file for the sequential CPU steps (convert/prepare/merge),
 # one file per tile for the GPU segment jobs.
-LOGS = f"{WORK}/logs"
+LOGS = f"{RUN}/logs"
 STEPLOG = f"{LOGS}/steps.log"
 
 # Marker that the segmentation model is cached locally. Produced by a local
 # rule (runs on the networked submit host) so offline GPU nodes never download.
-MODEL_OK = f"{WORK}/model.ready"
+# Namespaced per-run too: two configs using different models must each fetch
+# their own, rather than the second silently reusing the first's marker.
+MODEL_OK = f"{RUN}/model.ready"
 
 
 def occupied_done(wildcards):
     """Per-tile markers for the occupied tiles (resolved after the checkpoint)."""
     tiles = checkpoints.prepare.get().output.tiles
     occupied = json.loads(Path(tiles).read_text())["occupied"]
-    return [f"{WORK}/seg/{i}.done" for i in occupied]
+    return [f"{RUN}/seg/{i}.done" for i in occupied]

{patchworks-0.11.17 → patchworks-0.11.18}/workflow/rules/merge.smk

@@ -4,7 +4,7 @@ rule merge:
     input:
         occupied_done,
     output:
-        touch(f"{WORK}/labels.done"),
+        touch(f"{RUN}/labels.done"),
     log:
         STEPLOG,
     script:

{patchworks-0.11.17 → patchworks-0.11.18}/workflow/rules/segment.smk

@@ -35,7 +35,7 @@ rule segment:
         image=IMAGE_OK,
         model=MODEL_OK,
     output:
-        f"{WORK}/seg/{{index}}.done",
+        f"{RUN}/seg/{{index}}.done",
     log:
         f"{LOGS}/segment/{{index}}.log",
     script:

{patchworks-0.11.17 → patchworks-0.11.18}/workflow/scripts/_pw.py

@@ -87,20 +87,26 @@ def open_image(work_dir, channel, level):
     )
 
 
-def stage_path(work_dir):
-    """Path of the staged-labels store.
+def stage_path(work_dir, label_name):
+    """Path of the staged-labels store for one segmentation run.
+
+    Namespaced under ``<work_dir>/<label_name>/`` so two segmentations
+    (different ``label_name``) can target the same ``work_dir`` without
+    colliding — see docs/guide/snakemake.md "Running two segmentations".
 
     Parameters
     ----------
     work_dir : str or Path
         Workflow output directory.
+    label_name : str
+        This run's ``label_name`` (from config).
 
     Returns
     -------
     str
-        ``<work_dir>/stage.zarr``.
+        ``<work_dir>/<label_name>/stage.zarr``.
     """
-    return str(Path(work_dir) / "stage.zarr")
+    return str(Path(work_dir) / label_name / "stage.zarr")
 
 
 def load_tiles_json(path):

{patchworks-0.11.17 → patchworks-0.11.18}/workflow/scripts/merge.py

@@ -17,8 +17,9 @@ from _pw import stage_path, start_log
 start_log(snakemake.log[0])  # noqa: F821
 cfg = snakemake.config  # noqa: F821
 work_dir = cfg["work_dir"]
+label_name = cfg.get("label_name", "labels")
 image_store = str(Path(work_dir) / "image.zarr")
-merged_store = str(Path(work_dir) / "_merged.zarr")
+merged_store = str(Path(work_dir) / label_name / "_merged.zarr")
 
 # merge_tile_labels defaults to min(4, cpu_count) workers, so it ignores
 # whatever cpus_per_task the "merge" rule was actually allocated in the SLURM
@@ -26,7 +27,7 @@ merged_store = str(Path(work_dir) / "_merged.zarr")
 # the cores it's paying for; merge_workers: in config.yaml can still override.
 default_workers = int(os.environ.get("SLURM_CPUS_PER_TASK", os.cpu_count() or 4))
 merged = merge_tile_labels(
-    stage_path(work_dir),
+    stage_path(work_dir, label_name),
     write_to=merged_store,
     input_component="staged",
     sequential_labels=cfg.get("sequential_labels", True),
@@ -36,18 +37,18 @@ merged = merge_tile_labels(
 group = write_labels(
     image_store,
     merged,
-    name=cfg.get("label_name", "labels"),
+    name=label_name,
     n_levels=int(cfg.get("pyramid_levels", 5)),
     downscale=int(cfg.get("pyramid_downscale", 2)),
     overwrite=True,
 )
 
 shutil.rmtree(merged_store, ignore_errors=True)
-shutil.rmtree(stage_path(work_dir), ignore_errors=True)
+shutil.rmtree(stage_path(work_dir, label_name), ignore_errors=True)
 # Also drop the checkpoint's completion sentinel (stage.zarr.done): the
 # "prepare" rule's stage=touch(STAGE_OK) output must not outlive the store it
 # claims exists, or a future rerun (e.g. re-segmenting for new labels) skips
 # "prepare" and "segment" tries to open a stage.zarr that's already gone.
-Path(f"{stage_path(work_dir)}.done").unlink(missing_ok=True)
+Path(f"{stage_path(work_dir, label_name)}.done").unlink(missing_ok=True)
 print(f"[patchworks] labels written to {group}")
 open(snakemake.output[0], "w").close()  # noqa: F821

{patchworks-0.11.17 → patchworks-0.11.18}/workflow/scripts/prepare_tiles.py

@@ -16,6 +16,8 @@ from _pw import open_image, stage_path, start_log
 start_log(snakemake.log[0])  # noqa: F821
 cfg = snakemake.config  # noqa: F821
 work_dir = cfg["work_dir"]
+label_name = cfg.get("label_name", "labels")
+Path(work_dir, label_name).mkdir(parents=True, exist_ok=True)
 image = open_image(work_dir, cfg["channel"], cfg["level"])
 
 ts = cfg.get("tile_shape", "auto")
@@ -46,9 +48,9 @@ if cfg.get("skip_empty", True):
     occ = info["occupancy"].ravel()  # row-major, matches spatial_tiles
     occupied = [i for i in range(len(tiles)) if occ[i]]
 
-create_stage(stage_path(work_dir), image.shape, tile_shape)
+create_stage(stage_path(work_dir, label_name), image.shape, tile_shape)
 
-Path(work_dir, "tiles.json").write_text(
+Path(work_dir, label_name, "tiles.json").write_text(
     json.dumps(
         {
             "tile_shape": list(tile_shape),

{patchworks-0.11.17 → patchworks-0.11.18}/workflow/scripts/segment_tile.py

@@ -12,6 +12,7 @@ start_log(snakemake.log[0])  # noqa: F821
 cfg = snakemake.config  # noqa: F821
 index = int(snakemake.wildcards.index)  # noqa: F821
 work_dir = cfg["work_dir"]
+label_name = cfg.get("label_name", "labels")
 
 manifest = load_tiles_json(snakemake.input.tiles)  # noqa: F821
 image = open_image(work_dir, cfg["channel"], cfg["level"])
@@ -19,7 +20,7 @@ image = open_image(work_dir, cfg["channel"], cfg["level"])
 stage_tile(
     image,
     build_fn(cfg),
-    stage_path(work_dir),
+    stage_path(work_dir, label_name),
     index,
     tile_shape=tuple(manifest["tile_shape"]),
     overlap=int(manifest["overlap"]),