patchworks 0.5.0__tar.gz → 0.7.0__tar.gz

{patchworks-0.5.0 → patchworks-0.7.0}/.github/workflows/release.yml

@@ -48,3 +48,17 @@ jobs:
 
       - name: Publish to PyPI
         uses: pypa/gh-action-pypi-publish@release/v1
+
+  # Rebuild the org-wide pdoc apidocs site so it picks up the new version.
+  apidocs:
+    needs: release
+    runs-on: ubuntu-latest
+    steps:
+      - name: Trigger imcf.github.io apidocs rebuild
+        uses: peter-evans/repository-dispatch@v3
+        with:
+          # Fine-grained PAT with "Contents: write" on imcf/imcf.github.io,
+          # stored as the APIDOCS_DISPATCH_TOKEN secret in this repo.
+          token: ${{ secrets.APIDOCS_DISPATCH_TOKEN }}
+          repository: imcf/imcf.github.io
+          event-type: dispatch-event

{patchworks-0.5.0 → patchworks-0.7.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: patchworks
-Version: 0.5.0
+Version: 0.7.0
 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
@@ -127,11 +127,15 @@ def my_fn(tile):
     return label(tile > threshold_otsu(tile)).astype("int32")
 
 
-result = tile_process("image.zarr", my_fn, compute=True)
+result = tile_process("image.zarr", my_fn)
 ```
 
-Done. `result` is a NumPy array of integer labels, same spatial shape as the
-input, with globally unique IDs across all tiles.
+Done. `result` is a **lazy dask array** of integer labels (call `.compute()`
+for a NumPy array), same spatial shape as the input, with globally unique IDs
+across all tiles. By default the labels are also written **into the input
+store** at `image.zarr/labels/labels/` as a multi-scale pyramid, so the image
+and its segmentation live in one OME-ZARR. Pass `write_to="labels.zarr"` to
+write a separate store instead.
 
 ---
 
@@ -203,6 +207,26 @@ tile_process("image.zarr", my_custom_fn, tile_shape=(1, 512, 512))
 
 ---
 
+## Convert to OME-ZARR & view in napari
+
+Optional plugins close the loop: convert any image (Imaris `.ims`, CZI, LIF,
+ND2, OME-TIFF, … via bioio) to a pyramidal, **calibrated** OME-ZARR, then view
+the image and its labels in napari.
+
+```python
+from patchworks.plugins.ome_zarr import to_ome_zarr
+from patchworks.plugins.napari import view_in_napari
+
+to_ome_zarr("scan.ims", "scan.zarr")          # lazy, OOM-safe, keeps µm calibration
+view_in_napari("scan.zarr", labels="scan.zarr/labels/labels")
+```
+
+Pyramids downsample **X/Y only** (Z kept full-res) and are built level-by-level
+from disk, so terabyte volumes convert in bounded RAM. See the
+[OME-ZARR & napari guide](https://imcf.one/patchworks/guide/ome_zarr_napari/).
+
+---
+
 ## Common patterns
 
 ### Auto-size tiles from available memory
@@ -288,8 +312,8 @@ merged = merge_tile_labels(
 
 ## How tiling and merging work
 
-See [docs/how-it-works.md](docs/how-it-works.md) for a full explanation.
-Short version:
+See the [Merging labels guide](https://imcf.one/patchworks/guide/merging/) for
+a full explanation. Short version:
 
 1. Image is split into tiles (with optional overlap for boundary context).
 2. Your function is called independently on each tile. Dask handles parallelism
@@ -319,10 +343,15 @@ tiles where the dask-image approach stalls.
 
 ## Documentation
 
-- [Quick Start](docs/quickstart.md)
-- [API Reference](docs/api-reference.md)
-- [How It Works](docs/how-it-works.md)
-- [Examples](docs/examples/)
+Full docs, guides and tutorials: **<https://imcf.one/patchworks/>**
+
+- [Getting Started](https://imcf.one/patchworks/getting_started/)
+- [User Guide](https://imcf.one/patchworks/guide/tiling/) — tiling, merging,
+  empty-tile skipping, GPU/distributed, OME-ZARR & napari, pitfalls
+- [Examples](https://imcf.one/patchworks/examples/cellpose_2d/) — Cellpose,
+  StarDist, custom functions, standalone merge
+- [API Reference](https://imcf.one/patchworks/api/tile_process/) ·
+  [pdoc API](https://imcf.one/apidocs/patchworks/)
 
 ---
 
@@ -335,7 +364,11 @@ Optional:
 - `psutil` — accurate RAM sizing for `tile_shape="auto"`
 - `nvidia-ml-py` — accurate GPU VRAM sizing
 - `tqdm` — progress bars
-- `cellpose` — Cellpose plugin
+- `cellpose` — Cellpose plugin (`patchworks[cellpose]`)
+- `bioio` + readers — convert CZI/LIF/ND2/OME-TIFF/… to OME-ZARR
+  (`patchworks[bioio]`)
+- `imaris-ims-file-reader` — convert Imaris `.ims` (`patchworks[imaris]`)
+- `napari` — interactive viewer plugin (`patchworks[napari]`)
 
 ---
 

{patchworks-0.5.0 → patchworks-0.7.0}/README.md

@@ -61,11 +61,15 @@ def my_fn(tile):
     return label(tile > threshold_otsu(tile)).astype("int32")
 
 
-result = tile_process("image.zarr", my_fn, compute=True)
+result = tile_process("image.zarr", my_fn)
 ```
 
-Done. `result` is a NumPy array of integer labels, same spatial shape as the
-input, with globally unique IDs across all tiles.
+Done. `result` is a **lazy dask array** of integer labels (call `.compute()`
+for a NumPy array), same spatial shape as the input, with globally unique IDs
+across all tiles. By default the labels are also written **into the input
+store** at `image.zarr/labels/labels/` as a multi-scale pyramid, so the image
+and its segmentation live in one OME-ZARR. Pass `write_to="labels.zarr"` to
+write a separate store instead.
 
 ---
 
@@ -137,6 +141,26 @@ tile_process("image.zarr", my_custom_fn, tile_shape=(1, 512, 512))
 
 ---
 
+## Convert to OME-ZARR & view in napari
+
+Optional plugins close the loop: convert any image (Imaris `.ims`, CZI, LIF,
+ND2, OME-TIFF, … via bioio) to a pyramidal, **calibrated** OME-ZARR, then view
+the image and its labels in napari.
+
+```python
+from patchworks.plugins.ome_zarr import to_ome_zarr
+from patchworks.plugins.napari import view_in_napari
+
+to_ome_zarr("scan.ims", "scan.zarr")          # lazy, OOM-safe, keeps µm calibration
+view_in_napari("scan.zarr", labels="scan.zarr/labels/labels")
+```
+
+Pyramids downsample **X/Y only** (Z kept full-res) and are built level-by-level
+from disk, so terabyte volumes convert in bounded RAM. See the
+[OME-ZARR & napari guide](https://imcf.one/patchworks/guide/ome_zarr_napari/).
+
+---
+
 ## Common patterns
 
 ### Auto-size tiles from available memory
@@ -222,8 +246,8 @@ merged = merge_tile_labels(
 
 ## How tiling and merging work
 
-See [docs/how-it-works.md](docs/how-it-works.md) for a full explanation.
-Short version:
+See the [Merging labels guide](https://imcf.one/patchworks/guide/merging/) for
+a full explanation. Short version:
 
 1. Image is split into tiles (with optional overlap for boundary context).
 2. Your function is called independently on each tile. Dask handles parallelism
@@ -253,10 +277,15 @@ tiles where the dask-image approach stalls.
 
 ## Documentation
 
-- [Quick Start](docs/quickstart.md)
-- [API Reference](docs/api-reference.md)
-- [How It Works](docs/how-it-works.md)
-- [Examples](docs/examples/)
+Full docs, guides and tutorials: **<https://imcf.one/patchworks/>**
+
+- [Getting Started](https://imcf.one/patchworks/getting_started/)
+- [User Guide](https://imcf.one/patchworks/guide/tiling/) — tiling, merging,
+  empty-tile skipping, GPU/distributed, OME-ZARR & napari, pitfalls
+- [Examples](https://imcf.one/patchworks/examples/cellpose_2d/) — Cellpose,
+  StarDist, custom functions, standalone merge
+- [API Reference](https://imcf.one/patchworks/api/tile_process/) ·
+  [pdoc API](https://imcf.one/apidocs/patchworks/)
 
 ---
 
@@ -269,7 +298,11 @@ Optional:
 - `psutil` — accurate RAM sizing for `tile_shape="auto"`
 - `nvidia-ml-py` — accurate GPU VRAM sizing
 - `tqdm` — progress bars
-- `cellpose` — Cellpose plugin
+- `cellpose` — Cellpose plugin (`patchworks[cellpose]`)
+- `bioio` + readers — convert CZI/LIF/ND2/OME-TIFF/… to OME-ZARR
+  (`patchworks[bioio]`)
+- `imaris-ims-file-reader` — convert Imaris `.ims` (`patchworks[imaris]`)
+- `napari` — interactive viewer plugin (`patchworks[napari]`)
 
 ---
 

{patchworks-0.5.0 → patchworks-0.7.0}/docs/examples/custom.md

@@ -17,7 +17,7 @@ def threshold_fn(tile: np.ndarray) -> np.ndarray:
     return label(tile > thr).astype("int32")
 
 
-result = tile_process("image.zarr", threshold_fn, compute=True)
+result = tile_process("image.zarr", threshold_fn)
 ```
 
 ## Gaussian + morphological operations
@@ -86,12 +86,12 @@ from patchworks import tile_process
 
 # From any array-like source
 arr = da.from_array(my_numpy_array, chunks=(1, 1024, 1024))
-result = tile_process(arr, my_fn, compute=True)
+result = tile_process(arr, my_fn)
 
 # From tifffile
 import tifffile
 import dask.array as da
 
 arr = da.from_array(tifffile.imread("image.tif", aszarr=True))
-result = tile_process(arr, my_fn, compute=True)
+result = tile_process(arr, my_fn)
 ```

{patchworks-0.5.0 → patchworks-0.7.0}/docs/examples/custom_method.py

@@ -41,8 +41,7 @@ result = tile_process(
     my_fn,
     tile_shape=(1, 512, 512),
     overlap=16,
-    compute=True,
     progress=True,
 )
 
-print(f"Found {result.max()} objects")
+print(f"Found {int(result.max().compute())} objects")

{patchworks-0.5.0 → patchworks-0.7.0}/docs/getting_started.md

@@ -89,9 +89,11 @@ objects spanning tile boundaries are merged into a single label.
     ```python
     from patchworks import tile_process
 
-    result = tile_process("image.zarr", my_fn, compute=True)
+    # returns a lazy dask array; labels are also written into image.zarr by
+    # default (image.zarr/labels/labels/, as a pyramid)
+    result = tile_process("image.zarr", my_fn)
     print(result.shape)  # (z, y, x)
-    print(result.max())  # number of objects found
+    print(int(result.max().compute()))  # number of objects found
     ```
 
 === "From a dask array"
@@ -101,7 +103,7 @@ objects spanning tile boundaries are merged into a single label.
     from patchworks import tile_process
 
     arr = da.from_zarr("image.zarr")
-    result = tile_process(arr, my_fn, compute=True)
+    result = tile_process(arr, my_fn)
     ```
 
 === "Stream to zarr (recommended for large images)"

{patchworks-0.5.0 → patchworks-0.7.0}/docs/guide/ome_zarr_napari.md

@@ -50,6 +50,13 @@ to_ome_zarr("scan.czi", "scan.zarr", n_levels=5)   # via bioio
 to_ome_zarr("scan.ims", "scan.zarr")               # Imaris, native HDF5
 ```
 
+!!! note "Imaris pyramids are rebuilt, not reused"
+    `.ims` files carry their own resolution pyramid, but `to_ome_zarr` reads
+    only the **full-resolution** level and **builds a fresh NGFF pyramid** from
+    it. This guarantees a consistent pyramid (XY-only, nearest-neighbour,
+    calibrated) rather than inheriting Imaris's own downsampling scheme. It
+    costs some extra compute, but the build is lazy and OOM-safe.
+
 ### Pixel calibration
 
 The physical voxel size is read from the input — bioio's `physical_pixel_sizes`,
@@ -96,13 +103,17 @@ write_labels("scan.zarr", my_labels, name="nuclei")
 layer in one call. OME-ZARR pyramids are handed to napari as a lazy multi-scale
 list, so even huge stores open instantly and only on-screen data is fetched.
 
+Because `tile_process` writes labels **into** the store by default, you usually
+need no `labels=` argument at all — `view_in_napari` auto-loads every label
+image found under `scan.zarr/labels/`:
+
 ```python
 from patchworks.plugins.napari import view_in_napari
 
-# one store holding both image and labels/<name>:
-view_in_napari("scan.zarr", labels="scan.zarr/labels/labels")
+# auto-loads scan.zarr/labels/* as Labels layers:
+view_in_napari("scan.zarr")
 
-# or a separate plain label store written with write_to=:
+# or point at a separate plain label store written with write_to=:
 view_in_napari("scan.zarr", labels="labels.zarr")
 ```
 
@@ -120,7 +131,7 @@ from patchworks.plugins.napari import view_in_napari
 tile_process("scan.zarr", fn, progress=True)
 
 # 2. inspect image + labels together, straight from the one store
-view_in_napari("scan.zarr", labels="scan.zarr/labels/labels")
+view_in_napari("scan.zarr")  # labels auto-loaded from scan.zarr/labels/
 ```
 
 Plugging in a different segmentation method is just swapping `fn` — any

patchworks-0.7.0/docs/guide/performance.md

@@ -0,0 +1,76 @@
+# Performance & memory safety
+
+`tile_process` is built so a run **adapts to whatever machine it lands on** and
+can't run out of RAM/VRAM or freeze the box — without you tuning anything.
+
+## Automatic, machine-aware concurrency
+
+The staging step (running your `fn` once per tile to a temp store) and the
+merge step are sized to the host automatically:
+
+- **GPU** (`use_gpu=True`) → **one tile at a time**, so concurrent evaluations
+  can never exhaust VRAM.
+- **CPU** → as many tiles in flight as fit **80 % of available RAM** (estimated
+  from the tile size), and always **leaving one core free** so the machine
+  stays responsive — it never pins every core.
+
+The RAM figure is read live via `psutil`; without it, a conservative default is
+used instead of guessing high.
+
+## Live progress dashboard (GPU runs)
+
+A single-GPU run still gets a **Dask dashboard**: patchworks spins up a tiny
+1-worker / 1-thread in-process cluster, which keeps GPU evaluations serial (no
+VRAM contention) while exposing the dashboard so you can watch tiles stream
+through. The URL is logged at the start of staging:
+
+```text
+INFO:patchworks._core:Dask dashboard for this run: http://127.0.0.1:8787/status
+```
+
+This needs `distributed` (and `bokeh` for the UI) installed; if they are
+missing, patchworks logs a warning and falls back to the threaded scheduler
+(no dashboard, same result). A cluster you start yourself
+(`make_local_cluster`) is used as-is instead.
+
+## Overriding the worker count
+
+```python
+from patchworks import tile_process
+
+# let patchworks pick (recommended)
+tile_process("scan.zarr", fn)
+
+# or cap it yourself (staging threads + merge processes)
+tile_process("scan.zarr", fn, max_workers=8)
+```
+
+`max_workers` bounds both staging and merging. A running **distributed client**
+manages its own concurrency, so the override is skipped there — configure the
+cluster's memory limits instead.
+
+## Why it won't OOM or freeze
+
+| Resource | Guard |
+|----------|-------|
+| RAM | concurrent tiles × tile size × overhead ≤ 80 % of available RAM |
+| VRAM | GPU path runs one tile at a time |
+| CPU | always leaves at least one core free |
+| Disk I/O | each pyramid/stage level is streamed chunk-by-chunk; no whole volume in memory |
+
+The staging graph itself is kept small — a single fused `map_overlap`
+(halo → `fn` → trim) rather than three separate passes — and there is **no**
+extra read-back of the staged data.
+
+## Getting more speed
+
+- `tile_shape="auto"` sizes tiles to free RAM (or VRAM with `use_gpu=True`).
+- `skip_empty=True` with `estimate_empty_tiles()` skips background tiles.
+- A Dask **distributed** cluster (`make_local_cluster`) parallelises across
+  workers/GPUs; patchworks then defers concurrency to the cluster.
+
+!!! note "What doesn't help here"
+    The merge and relabel steps are already vectorised NumPy + SciPy (C-level)
+    with no per-voxel Python loop, and the pipeline is I/O-bound — so `numba`,
+    `cupy`, `arrow` and `xarray` bring essentially nothing. The real levers are
+    tile size, concurrency (above) and zarr chunking.

{patchworks-0.5.0 → patchworks-0.7.0}/docs/index.md

@@ -53,7 +53,7 @@ def my_fn(tile):
     return label(tile > threshold_otsu(tile)).astype("int32")
 
 
-result = tile_process("image.zarr", my_fn, compute=True)
+result = tile_process("image.zarr", my_fn)
 ```
 
 Any function. Any image.

{patchworks-0.5.0 → patchworks-0.7.0}/mkdocs.yml

@@ -38,6 +38,7 @@ nav:
   - Merging labels: guide/merging.md
   - Empty tile skipping: guide/skip_empty.md
   - GPU & distributed: guide/gpu_distributed.md
+  - Performance & memory: guide/performance.md
   - OME-ZARR & napari: guide/ome_zarr_napari.md
   - Pitfalls: guide/pitfalls.md
 - Examples:

{patchworks-0.5.0 → patchworks-0.7.0}/src/patchworks/_chunks.py

@@ -57,6 +57,51 @@ def _get_available_memory() -> int:
         return 8 * 1024**3
 
 
+def safe_worker_count(
+    tile_nbytes: int,
+    *,
+    use_gpu: bool = False,
+    fn_overhead: int = 4,
+    ram_fraction: float = 0.8,
+) -> int:
+    """Concurrent tiles that fit the machine without OOM or a CPU freeze.
+
+    Bounds the threaded scheduler by two limits and takes the smaller:
+
+    * **CPU** — leaves at least one core free so the box stays responsive
+      (never pins every core).
+    * **RAM** — at most ``ram_fraction`` of available memory, assuming each
+      in-flight tile needs ``fn_overhead`` copies (halo + output + temporaries).
+
+    On GPU the answer is always 1: one evaluation at a time so concurrent
+    tiles can never exhaust VRAM. Without ``psutil`` it returns a conservative
+    default rather than guessing high.
+
+    Parameters
+    ----------
+    tile_nbytes : int
+        Size of one tile in bytes (``prod(tile_shape) * dtype.itemsize``).
+    use_gpu : bool, optional
+        Whether tiles are processed on the GPU.
+    fn_overhead : int, optional
+        Assumed peak number of tile-sized buffers alive per worker.
+    ram_fraction : float, optional
+        Fraction of available RAM the staging step may use.
+
+    Returns
+    -------
+    int
+        Worker-thread count (always >= 1).
+    """
+    cpu_cap = max(1, (os.cpu_count() or 1) - 1)
+    if use_gpu:
+        return 1
+    avail = _get_available_memory()
+    per_tile = max(1, int(tile_nbytes) * max(1, fn_overhead))
+    mem_cap = max(1, int(avail * ram_fraction) // per_tile)
+    return max(1, min(cpu_cap, mem_cap))
+
+
 def _get_gpu_memory() -> int:
     """Return free GPU VRAM in bytes. Falls back to 8 GiB default."""
     try:

{patchworks-0.5.0 → patchworks-0.7.0}/src/patchworks/_core.py

@@ -11,7 +11,7 @@ from typing import Any, Callable, Union
 import dask.array as da
 import numpy as np
 
-from ._chunks import auto_tile_shape
+from ._chunks import auto_tile_shape, safe_worker_count
 from ._cluster import _client_is_in_process, _distributed_client
 from ._io import _auto_empty_threshold, load_ome_zarr
 from ._merge import zarr_native_merge
@@ -56,6 +56,7 @@ def tile_process(
     channel: int | None = 0,
     level: int = 0,
     use_gpu: bool = False,
+    max_workers: int | None = None,
     progress: bool = False,
     write_to: Union[str, Path, None] = None,
     output_component: str = "labels",
@@ -114,6 +115,13 @@ def tile_process(
         Pyramid level when *image* is a path (0 = full resolution).
     use_gpu:
         When ``tile_shape="auto"``, size tiles against GPU VRAM instead of RAM.
+        Also forces staging to one tile at a time (no VRAM contention).
+    max_workers:
+        Cap the worker threads/processes used for staging and merging. ``None``
+        (default) auto-sizes to the machine: bounded by available RAM (tile
+        size) and CPU (leaves one core free) so a run can neither OOM nor pin
+        every core. Ignored when a distributed client is active (it manages its
+        own concurrency).
     progress:
         Show a progress bar during the tile-writing and relabel steps.
     write_to:
@@ -283,11 +291,6 @@ def tile_process(
         for ax, c in enumerate(image.chunks)
     }
 
-    if overlap > 0:
-        # boundary="none" is required: only this boundary mode composes with
-        # trim_overlap to recover the original shape. "reflect" keeps the halo.
-        image = da.overlap.overlap(image, depth=_depth, boundary="none")
-
     # Wrap fn with optional empty-tile skipping
     _skip_thr = empty_threshold
     if skip_empty and _skip_thr is None:
@@ -303,28 +306,67 @@ def tile_process(
             logger.debug("process tile %s shape=%s", loc, block.shape)
         return fn(block)
 
-    labeled = image.map_blocks(
-        active_fn,
-        dtype=np.int32,
-        meta=np.empty((0,) * image.ndim, dtype=np.int32),
-    )
-
-    # Trim the overlap halo so staged tiles have clean boundaries for the
-    # boundary-slab scan. Without this the scan reads halo-expanded chunks and
-    # the merged output is larger than the input.
+    _meta = np.empty((0,) * image.ndim, dtype=np.int32)
     if overlap > 0:
-        labeled = da.overlap.trim_overlap(
-            labeled, depth=_depth, boundary="none"
+        # One fused pass: add the halo, run fn, trim it back off. map_overlap
+        # materialises only the halos it needs (no separate overlapped array)
+        # and keeps the task graph small. boundary="none" + trim recovers the
+        # original shape, so the boundary-slab scan reads clean tiles.
+        labeled = da.map_overlap(
+            active_fn,
+            image,
+            depth=_depth,
+            boundary="none",
+            trim=True,
+            dtype=np.int32,
+            meta=_meta,
         )
+    else:
+        labeled = image.map_blocks(active_fn, dtype=np.int32, meta=_meta)
 
-    # With no distributed client the threaded scheduler runs many tiles at
-    # once. For GPU that means several evals sharing one device → CUDA OOM.
-    # Pin to a single worker thread so evals run serially. A distributed
-    # client manages its own concurrency, so skip the override there.
+    # Bound staging concurrency to the machine so it can neither OOM nor pin
+    # every core:
+    #   - GPU → 1 eval at a time (no VRAM contention),
+    #   - CPU → as many tiles as fit RAM, leaving one core free.
+    # A distributed client manages its own concurrency, so skip the override.
     import dask as _dask
 
+    _tile_nbytes = int(np.prod(labeled.chunksize)) * labeled.dtype.itemsize
+    _temp_cluster = None
+    _temp_client = None
     if _active is None and use_gpu:
-        _sched_ctx: Any = _dask.config.set(scheduler="threads", num_workers=1)
+        # Single-GPU runs still get a live Dask dashboard: a 1-worker /
+        # 1-thread in-process cluster keeps GPU evals serial (no VRAM
+        # contention) while exposing the dashboard for progress.
+        try:
+            from dask.distributed import Client, LocalCluster
+
+            _temp_cluster = LocalCluster(
+                n_workers=1, threads_per_worker=1, processes=False
+            )
+            _temp_client = Client(_temp_cluster)
+            logger.info(
+                "Dask dashboard for this run: %s",
+                _temp_client.dashboard_link,
+            )
+        except Exception as exc:  # no distributed/bokeh → threaded fallback
+            logger.warning(
+                "Could not start a dashboard cluster (%s); "
+                "falling back to the threaded scheduler.",
+                exc,
+            )
+
+    if _distributed_client() is None:
+        _workers = (
+            max_workers
+            if max_workers is not None
+            else safe_worker_count(_tile_nbytes, use_gpu=use_gpu)
+        )
+        _workers = max(1, min(_workers, os.cpu_count() or 1))
+        logger.info("Staging with %d worker thread(s)", _workers)
+        _sched_ctx: Any = _dask.config.set(
+            scheduler="threads", num_workers=_workers
+        )
     else:
         _sched_ctx = _nullcontext()
 
@@ -345,26 +387,15 @@ def tile_process(
     logger.info("Staging tiles to %s …", stage_path)
     with _sched_ctx:
         _stage_to_zarr(labeled, stage_path, "staged", progress)
+    if _temp_client is not None:
+        _temp_client.close()
+        _temp_cluster.close()
     labeled = da.from_zarr(stage_path, component="staged")
 
-    if skip_empty and _skip_thr is not None:
-
-        def _tile_max(block: np.ndarray) -> np.ndarray:
-            return np.full((1,) * block.ndim, int(block.max()), dtype=np.int32)
-
-        _tile_maxes = labeled.map_blocks(
-            _tile_max,
-            dtype=np.int32,
-            chunks=tuple(tuple(1 for _ in c) for c in labeled.chunks),
-        ).compute()
-        _n_skip = int((_tile_maxes == 0).sum())
-        logger.info(
-            "skip_empty: %d/%d tiles ran fn, %d skipped (max<=%.4g)",
-            int(_tile_maxes.size) - _n_skip,
-            int(_tile_maxes.size),
-            _n_skip,
-            _skip_thr,
-        )
+    # NB: no post-staging skip-count pass here — counting skipped tiles by
+    # re-reading the whole staged store off disk would double the I/O of the
+    # entire run just for a log line. Use estimate_empty_tiles() up front for
+    # that figure instead.
 
     def _cleanup_stage():
         if not keep_stage:
@@ -373,7 +404,9 @@ def tile_process(
             shutil.rmtree(stage_path, ignore_errors=True)
             logger.info("Removed stage store %s", stage_path)
 
-    _nw = min(4, os.cpu_count() or 1)
+    # Merge runs in worker processes (each holds one chunk + an mmap'd LUT);
+    # size it to RAM/CPU like staging, capped so we don't spawn a process storm.
+    _nw = max_workers or max(1, min(safe_worker_count(_tile_nbytes), 8))
 
     # Default: input is a .zarr store and no explicit write_to → labels go back
     # *into* the input store under the NGFF labels/<name>/ group with an auto

{patchworks-0.5.0 → patchworks-0.7.0}/src/patchworks/plugins/napari.py

@@ -14,13 +14,12 @@ napari is an optional, GUI-heavy dependency. Install it with
 Usage
 -----
 >>> from patchworks import tile_process
->>> from patchworks.plugins.ome_zarr import to_ome_zarr
 >>> from patchworks.plugins.napari import view_in_napari
 >>>
->>> tile_process("scan.zarr", fn, write_to="labels.zarr")
->>> to_ome_zarr("scan.zarr", "scan_pyramid.zarr")        # optional, for speed
->>>
->>> view_in_napari("scan_pyramid.zarr", labels="labels.zarr")
+>>> # labels are written into scan.zarr/labels/ by default …
+>>> tile_process("scan.zarr", fn)
+>>> # … so the viewer finds and overlays them with no labels= argument:
+>>> view_in_napari("scan.zarr")
 """
 
 from __future__ import annotations
@@ -86,6 +85,15 @@ def _resolve_image(
     return source
 
 
+def _inner_label_names(store: Union[str, Path]) -> list[str]:
+    """Names registered under an OME-ZARR's NGFF ``labels/`` group, if any."""
+    try:
+        grp = zarr.open_group(f"{store}/labels", mode="r")
+    except Exception:
+        return []
+    return list(grp.attrs.get("labels", []))
+
+
 def _resolve_labels(
     source: Union[da.Array, str, Path], component: str
 ) -> Union[da.Array, list[da.Array]]:
@@ -123,7 +131,10 @@ def view_in_napari(
     labels : da.Array, str, Path or None
         Label array to overlay. A plain ``.zarr`` store written by
         ``tile_process`` is read from its ``labels_component``; an OME-ZARR
-        pyramid is shown multi-scale; ``None`` shows the image only.
+        pyramid is shown multi-scale. ``None`` (default) **auto-loads** every
+        label image stored inside the OME-ZARR under ``labels/<name>/`` — the
+        place ``tile_process`` writes them by default — each as its own Labels
+        layer. (Falls back to image-only if there are none.)
     channel : int or None, optional
         Channel to display from the image (``None`` keeps all channels).
     labels_component : str, optional
@@ -145,7 +156,7 @@ def view_in_napari(
 
     Examples
     --------
-    >>> view_in_napari("scan.zarr", labels="labels.zarr")  # doctest: +SKIP
+    >>> view_in_napari("scan.zarr")  # auto-loads scan.zarr/labels/*  # doctest: +SKIP
     """
     napari = _require_napari()
 
@@ -161,6 +172,15 @@ def view_in_napari(
     if labels is not None:
         lab = _resolve_labels(labels, labels_component)
         viewer.add_labels(lab, name=labels_name)
+    elif _is_zarr(image):
+        # No labels given → auto-overlay every label image stored inside the
+        # OME-ZARR under labels/<name>/ (the default place tile_process writes
+        # them), each as its own multi-scale Labels layer.
+        for name in _inner_label_names(image):
+            levels = _multiscale_levels(f"{image}/labels/{name}", None)
+            lab = [lvl.astype("int32") for lvl in levels]
+            viewer.add_labels(lab if len(lab) > 1 else lab[0], name=name)
+            logger.info("auto-loaded labels/%s from %s", name, image)
 
     if show:
         napari.run()

{patchworks-0.5.0 → patchworks-0.7.0}/tests/test_core.py

@@ -247,3 +247,27 @@ def test_estimate_empty_tiles():
     assert info["n_tiles"] == 4
     assert info["n_occupied"] == 2
     assert info["empty_fraction"] == 0.5
+
+
+def test_safe_worker_count_bounds():
+    import os
+
+    from patchworks._chunks import safe_worker_count
+
+    # GPU → always serial (no VRAM contention)
+    assert safe_worker_count(10**6, use_gpu=True) == 1
+    # Absurdly large tile → memory-bound to 1
+    assert safe_worker_count(10**15) == 1
+    # Tiny tile → CPU-bound, leaves a core free, always >= 1
+    n = safe_worker_count(1024)
+    assert 1 <= n <= max(1, (os.cpu_count() or 1) - 1)
+
+
+def test_tile_process_max_workers():
+    import dask.array as da
+
+    from patchworks import tile_process
+
+    arr = da.from_array(_make_image((2, 32, 32)), chunks=(1, 32, 32))
+    result = tile_process(arr, _label_fn, max_workers=1).compute()
+    assert result.shape == (2, 32, 32)

{patchworks-0.5.0 → patchworks-0.7.0}/tests/test_napari.py

@@ -39,3 +39,32 @@ def test_require_napari_message(monkeypatch):
             nplugin._require_napari()
     else:
         assert nplugin._require_napari() is napari
+
+
+def test_inner_label_discovery(tmp_path):
+    """Labels written into a store are discoverable for auto-overlay."""
+    import numpy as np
+
+    from patchworks.plugins.ome_zarr import to_ome_zarr, write_labels
+
+    store = to_ome_zarr(
+        np.zeros((8, 8, 8), "uint16"), tmp_path / "scan.zarr", n_levels=2
+    )
+    write_labels(store, np.ones((8, 8, 8), "int32"), name="cells", n_levels=2)
+
+    assert nplugin._inner_label_names(store) == ["cells"]
+    levels = nplugin._multiscale_levels(f"{store}/labels/cells", None)
+    assert len(levels) == 2
+    assert levels[1].shape == (8, 4, 4)  # Z preserved, XY downsampled
+
+
+def test_inner_label_discovery_none(tmp_path):
+    """A store without labels yields an empty list (image-only view)."""
+    import numpy as np
+
+    from patchworks.plugins.ome_zarr import to_ome_zarr
+
+    store = to_ome_zarr(
+        np.zeros((8, 8, 8), "uint16"), tmp_path / "img.zarr", n_levels=1
+    )
+    assert nplugin._inner_label_names(store) == []