0.3.0

CHANGELOG.md

@@ -4,6 +4,8 @@ All notable changes to cellmap-flow. The format follows [Keep a Changelog](https
 
 ## Unreleased
 
+## 0.3.0 — 2026-10-22
+
 ### Added
 - **The paint logic written once, in Python, and run by the browser too.** `cellmap_flow/portable/` holds the Python both modes run, importable with the standard library, numpy, scipy, pydantic and Pillow alone (`tests/portable/test_boundary.py` holds it to that): the rank score (`portable/rank.py`, moved whole from `finetune/rank.py`, which keeps the import path) and the maths of *Use this model*'s paint carry (`portable/carry.py`, out of `carry_paint` in `finetune/session/volume.py`, which keeps the zarr reads and writes around it). The browser app bundles those files as text and runs them in chunkmirage's Pyodide worker (`Inference.python`: chunkmirage's `Engine.call` by module path, the one hook added to chunkmirage for it, pending there as PR *consumer-python*); Pyodide and numpy load in the background when a dataset opens, about 17 MB once per browser. The TypeScript ports `web/src/finetune/rank.ts` and `carry.ts` are gone with their unit tests, and the rank routes are tested on the same Python in node's Pyodide (`web/test/unit/pyodide.ts`; skipped without the network). The pilot of *One implementation per behaviour* in `web/ARCHITECTURE.md`.
 - **The dashboard's contract, written once and generated for the browser.** `cellmap_flow/schema.py` exports a JSON Schema of every request model (`dashboard/requests.py`), the table of every route with where it is answered (both modes, a browser stub, the server only, the browser only) and the constants both modes share (the rank score's, Undo's depth, the AI annotation's caps, the Finetune form's defaults); `npm run gen` in web/ writes `web/src/generated/cellmap-flow.{schema.json,ts}` and `docs/api.md`, and a test fails when they are stale, when a Flask route is added without saying where it is answered, or when the browser's router registers a route the table does not give it or misses one it does. The browser takes the shared numbers from the generated file. The rule it serves, *One implementation per behaviour*, is in `web/ARCHITECTURE.md`.
@@ -114,6 +116,7 @@ All notable changes to cellmap-flow. The format follows [Keep a Changelog](https
 - **The Flask inference server** (`cellmap_flow/server.py`, `CellMapFlowServer`) and `CELLMAP_FLOW_ENGINE`: chunkmirage serves every model. A model's own environment (`env:`) must have chunkmirage installed, as every environment in `pixi.toml` has; a server started where it is missing fails saying so, where it fell back to the Flask server with a warning. The dashboard draws only chunkmirage layers (`zarr://<server>/<model><settings>/zarr`, channels first), so this dashboard no longer draws a server from an older cellmap-flow, nor an older dashboard this release's servers. With the server go `cellmap_flow serve --debug` (Flask's dev-server flag, which the chunkmirage server ignored), `CELLMAP_FLOW_RAW_READ_CONCURRENCY` (`CELLMAP_FLOW_RAW_CACHE_BYTES` stays), the request-cancellation plumbing in `inference.runner` (chunkmirage drops the chunks nobody waits for itself), `ImageDataInterface`'s `concurrency_limit` and `cache_bytes` arguments (the server was their only user; `io.source.open_array` keeps them) and the `flask-cors` dependency. Flask itself stays: the dashboard is a Flask app.
 
 ### Fixed
+- **The dashboard says when painting is unavailable.** A pip install has upstream neuroglancer, whose client lacks the paint fork's brush, flood fill and MinIO writes (the fork is not on PyPI; pixi installs it). `cellmap_flow view` and the other dashboard commands now log one warning at start, with the install line, instead of a Draw tab that comes up without a brush; `cellmap_flow doctor` says the same. The doctor's install line names the commit pixi pins, not the fork's branch, and a test keeps the two together.
 - **A `cellmap` model is served at the shape its export runs at.** cellmap_models traces the TorchScript at the metadata's `inference_input_shape`, and a traced UNet runs only at that shape; the model type read the training `input_shape` instead, so `jrc_mus-livers_16nm_to_8nm_mito` and `_peroxisome` (216³ in training, 288³ exported) failed every chunk with "padded input size per channel: (1 x 1 x 1)". The export is traced at whichever shape the exporter passed, and the metadata does not say which (`mito_aff_unet_setup_16` is the other way round: traced at its training 178³, it answers 56³ to the inference 378³), so the model type now asks the model: a forward at the training shape that answers the training output means the training shapes, anything else the inference shapes.
 - **daisy is pinned below 2.** daisy 2.0.0 (released 2026-10-21) is a rewrite without `daisy.tcp`, `daisy.Server` and `daisy.Client`, which blockwise uses; the pin keeps installs on the 1.2 line until blockwise is ported.
 - **A model's `scale` is honoured wherever it is started.** A config with `scale` launched from the Models tab (ticked again after a YAML started it, or run by *Add a model* or *Use this model*) was served at the session's data path, the scale ignored; only `cellmap_flow infer`, `yaml` and blockwise applied it. Every launcher now reads the level the scale names (`serving.launch.launch`).

cellmap_flow/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.2.3"
+__version__ = "0.3.0"
 __version_info__ = tuple(int(i) for i in __version__.split("."))
 
 # Plugins (~/.cellmap_flow/plugins/*.py) are not loaded here: the commands

cellmap_flow/cli/doctor.py

@@ -28,8 +28,13 @@
 # would not work: the fork keeps upstream's version string.
 NEUROGLANCER_FORK_MARKER = "Paint Value"
 
+# The fork's commit pixi installs (pixi.toml, the finetune feature's
+# pypi-dependencies); tests/web/test_pins.py keeps the two the same, so a pip
+# user who follows the doctor gets the client that deploys. The install
+# builds the client with npm and a C++ extension with the compiler at hand.
+NEUROGLANCER_FORK_COMMIT = "fae54db8c667176a602a34f0d9d1028956009836"
 NEUROGLANCER_FORK_INSTALL = (
-    "pip install git+https://github.com/briossant/neuroglancer@feature/voxel-annotation"
+    f"pip install git+https://github.com/briossant/neuroglancer@{NEUROGLANCER_FORK_COMMIT}"
 )
 
 # MinIO stopped publishing prebuilt community server binaries: dl.min.io
@@ -130,6 +135,20 @@ def _check_neuroglancer_fork():
     )
 
 
+def painting_unavailable():
+    """One line for the dashboard's start when the installed neuroglancer
+    cannot paint (_check_neuroglancer_fork), with the install line; None when
+    it can. dashboard.app.create_and_run_app logs it, so a pip install, which
+    has upstream neuroglancer, hears at start why the Draw tab has no brush."""
+    result = _check_neuroglancer_fork()
+    if result.status == "ok":
+        return None
+    return (
+        f"Painting is unavailable: {result.detail}. To paint, install the fork: "
+        f"{result.fix} (cellmap_flow doctor checks the rest of the environment)."
+    )
+
+
 def _check_binary(name, fix):
     path = shutil.which(name)
     if path is None:

cellmap_flow/dashboard/app.py

@@ -115,11 +115,18 @@ def create_and_run_app(neuroglancer_url=None, server=None):
     the plugins, whose normalizers, postprocessors and model types its
     pages offer, if nothing has yet.
     """
+    from cellmap_flow.cli.doctor import painting_unavailable
     from cellmap_flow.plugins import load_plugins
 
     if not logging.getLogger().handlers:
         configure_logging()
     load_plugins()
+    # A pip install has upstream neuroglancer, whose client lacks the paint
+    # fork's brush, flood fill and MinIO writes: say so here, with the
+    # install line, rather than let the Draw tab come up without a brush.
+    unavailable = painting_unavailable()
+    if unavailable:
+        logger.warning(unavailable)
     get_session().neuroglancer_url = neuroglancer_url
     if server is None:
         server = bind_dashboard()

pixi.toml

@@ -15,7 +15,7 @@
 # serve and finetune on Apple's GPU). CUDA is Linux-only.
 [workspace]
 name = "cellmap-flow"
-version = "0.2.3"
+version = "0.3.0"
 description = "Realtime prediction using neuroglancer"
 channels = ["nvidia", "conda-forge", "pytorch"]
 platforms = ["linux-64", "osx-arm64"]

pyproject.toml

@@ -86,8 +86,11 @@ finetune = [
     "tensorboard",           # Training curves; lora_trainer silently skips logging without it
     # MinIO server and client are required for annotations but not available on PyPI.
     # Install via conda-forge: mamba install minio-server minio-client -c conda-forge
-    # Neuroglancer with voxel annotation support is required but not yet on PyPI.
-    # Install via: pip install git+https://github.com/briossant/neuroglancer@feature/voxel-annotation
+    # Painting needs the neuroglancer fork with the voxel-annotation tools
+    # (google/neuroglancer#858, not on PyPI; a URL here would make this
+    # package unpublishable). pixi installs it; with pip, install the commit
+    # pixi.toml pins: pip install
+    # git+https://github.com/briossant/neuroglancer@fae54db8c667176a602a34f0d9d1028956009836
 ]
 
 # AI-assisted annotation in the finetune tab (docs/ai_annotate.md): the

runnables.yaml

@@ -1,6 +1,6 @@
 name: CellMap Flow
 description: Real-time neural network inference on EM microscopy data.
-version: "0.2.3"
+version: "0.3.0"
 
 requirements:
   # pixi.lock is lockfile version 7, which pixi reads from 0.68.0 on.

tests/web/test_pins.py

@@ -1,5 +1,6 @@
-"""chunkmirage is pinned twice, and the two pins move together.
-pyproject.toml's is a version range, the release the Python inference server
+"""Pins that move together: chunkmirage's two, and the neuroglancer fork's two.
+
+chunkmirage is pinned twice. pyproject.toml's is a version range, the release the Python inference server
 is installed from: PyPI refuses an upload whose dependencies name a URL, so
 a git pin there would keep cellmap-flow from being released.
 web/chunkmirage.lock's is a commit, the one the browser engine is cloned
@@ -31,3 +32,17 @@ def test_the_server_takes_a_chunkmirage_release_and_the_browser_a_commit():
     )
     commits = LOCK_PIN.findall((ROOT / "web" / "chunkmirage.lock").read_text())
     assert len(commits) == 1, f"web/chunkmirage.lock should pin one full commit, found {commits}"
+
+
+def test_the_doctors_neuroglancer_fork_install_names_the_commit_pixi_installs():
+    """The paint fork is installed by pixi (pixi.toml's finetune feature) and,
+    for a pip install, by the line `cellmap_flow doctor` prints: the two name
+    one commit, or a pip user who follows the doctor gets a different client
+    from the one that deploys. pyproject.toml's comment names it too."""
+    from cellmap_flow.cli.doctor import NEUROGLANCER_FORK_COMMIT, NEUROGLANCER_FORK_INSTALL
+
+    pixi = tomllib.loads((ROOT / "pixi.toml").read_text())
+    pinned = pixi["feature"]["finetune"]["pypi-dependencies"]["neuroglancer"]["rev"]
+    assert NEUROGLANCER_FORK_COMMIT == pinned
+    assert NEUROGLANCER_FORK_INSTALL.endswith(f"@{pinned}")
+    assert pinned in (ROOT / "pyproject.toml").read_text()