Fix CI compatibility and notebook links

Author: wrignj08

README.md

@@ -107,7 +107,7 @@ If you are working with cloudy imagery, either:
 
 This matters because OWM optimises its detection thresholds both **locally** (per region/patch) and **globally** (across the whole scene). Cloud and cloud-shadow pixels are out-of-distribution and can skew those optimisations, so bad data in one part of a scene can degrade the water prediction in other, otherwise-clean parts. Masking those pixels to no-data removes them from the optimisation entirely.
 
-[OmniCloudMask](https://github.com/DPIRD-DMA/OmniCloudMask) is a good choice for the masking step. See the [cloudy Sentinel-2 example](https://github.com/DPIRD-DMA/OmniWaterMask/blob/main/examples/Sentinel-2%20example.ipynb) for an end-to-end mask-then-infer workflow.
+[OmniCloudMask](https://github.com/DPIRD-DMA/OmniCloudMask) is a good choice for the masking step. See the [cloudy Sentinel-2 example](https://github.com/DPIRD-DMA/OmniWaterMask/blob/main/examples/Sentinel%202%20cloudy%20example.ipynb) for an end-to-end mask-then-infer workflow.
 
 
 ## Parameters
@@ -159,7 +159,7 @@ Example notebooks are available in the [examples/](https://github.com/DPIRD-DMA/
 
 -   [NAIP example](https://github.com/DPIRD-DMA/OmniWaterMask/blob/main/examples/NAIP%20example.ipynb) — Water segmentation on NAIP aerial imagery from HuggingFace
 -   [Sentinel-2 example](https://github.com/DPIRD-DMA/OmniWaterMask/blob/main/examples/Sentinel%202%20example.ipynb) — Water segmentation on a Sentinel-2 mosaic using [s2mosaic](https://github.com/DPIRD-DMA/s2mosaic)
--   [Cloudy Sentinel-2 example](https://github.com/DPIRD-DMA/OmniWaterMask/blob/main/examples/Sentinel-2%20example.ipynb) — Masking clouds with [OmniCloudMask](https://github.com/DPIRD-DMA/OmniCloudMask) before running OWM on a cloudy AWS scene
+-   [Cloudy Sentinel-2 example](https://github.com/DPIRD-DMA/OmniWaterMask/blob/main/examples/Sentinel%202%20cloudy%20example.ipynb) — Masking clouds with [OmniCloudMask](https://github.com/DPIRD-DMA/OmniCloudMask) before running OWM on a cloudy AWS scene
 
 ## Changelog
 

examples/NAIP example.ipynb

@@ -87,7 +87,7 @@
     {
      "data": {
       "application/vnd.jupyter.widget-view+json": {
-       "model_id": "15c5ab4ffab546538750e0eed14c8f3d",
+       "model_id": "b2ecf3ebe7e64707aaa2f12927561f7f",
        "version_major": 2,
        "version_minor": 0
       },

examples/Sentinel 2 cloudy example.ipynb

@@ -14,6 +14,8 @@
 except ImportError:
     omniwatermask_version = "0.0.0+unknown"
 
+gdown_download: Any = getattr(gdown, "download")  # noqa: B009
+
 
 def download_file_from_google_drive(file_id: str, destination: Path) -> None:
     """
@@ -24,7 +26,7 @@ def download_file_from_google_drive(file_id: str, destination: Path) -> None:
         destination (Path): The local path where the file should be saved.
     """
     url = f"https://drive.google.com/uc?id={file_id}"
-    gdown.download(url, str(destination), quiet=False)
+    gdown_download(url, str(destination), quiet=False)
 
 
 def download_file_from_hugging_face(destination: Path) -> None:
@@ -60,11 +62,25 @@ def download_file(file_id: str, destination: Path, source: str) -> None:
         )
 
 
+def _release_version(version: str) -> str:
+    """Return only the public release portion of a version string.
+
+    Tag-based versioning produces dev/dirty suffixes between releases
+    (e.g. "0.5.1.dev3+g1a2b3c4"). Keying the model cache on the full
+    string would create a new empty directory for every commit and force
+    a re-download. Stripping the ".dev*"/"+local" suffix keeps the cache
+    stable between releases while still refreshing on real version bumps.
+    """
+    return version.split("+")[0].split(".dev")[0]
+
+
 def get_model_data_dir() -> Path:
     """Get the user data directory for model files"""
     data_dir = Path(
         platformdirs.user_data_dir(
-            "omniwatermask", version=omniwatermask_version, ensure_exists=True
+            "omniwatermask",
+            version=_release_version(omniwatermask_version),
+            ensure_exists=True,
         )
     )
     return data_dir

examples/Sentinel 2 example.ipynb

@@ -4,6 +4,7 @@
 import geopandas as gpd
 import numpy as np
 import rasterio as rio
+from numpy.typing import NDArray
 from rasterio import features
 from rasterio.transform import from_bounds
 
@@ -39,11 +40,11 @@ def resample_input(
 
 
 def export_to_disk(
-    array: np.ndarray,
+    array: NDArray[Any],
     export_path: Path,
     source_path: Path,
     layer_names: list[str],
-    nodata_mask: Optional[np.ndarray] = None,
+    nodata_mask: Optional[NDArray[Any]] = None,
 ) -> None:
     """Export the array to disk as a GeoTIFF.
 
@@ -76,7 +77,7 @@ def export_to_disk(
 
 def rasterize_vector(
     gdf: gpd.GeoDataFrame, reference_profile: dict[str, Any], all_touched: bool = False
-) -> np.ndarray:
+) -> NDArray[Any]:
     """Rasterize a GeoDataFrame into a binary array using the reference rio profile.
 
     With ``all_touched=False`` (the default) a pixel is only set when its centre

examples/Sentinel-2 example.ipynb

@@ -8,6 +8,7 @@
 import numpy as np
 import rasterio as rio
 import torch
+from numpy.typing import NDArray
 from omnicloudmask import predict_from_array
 from scipy.optimize import minimize_scalar
 
@@ -253,7 +254,7 @@ def multi_scale_optimisation(
 
 
 def get_NDWI(
-    input_bands: np.ndarray, mosaic_device: Union[str, torch.device]
+    input_bands: NDArray[Any], mosaic_device: Union[str, torch.device]
 ) -> torch.Tensor:
     input_bands_tensor = torch.from_numpy(input_bands.astype(np.float16)).to(
         mosaic_device
@@ -267,7 +268,7 @@ def get_NDWI(
 
 def make_composite_output(
     input_dict: dict[str, Optional[torch.Tensor]],
-) -> tuple[np.ndarray, list[str]]:
+) -> tuple[NDArray[Any], list[str]]:
     output_layers = []
     layer_names = []
     # Get the shape of the first non-None layer
@@ -289,7 +290,7 @@ def make_composite_output(
 
 
 def integrate_water_detection_methods(
-    input_bands: np.ndarray,
+    input_bands: NDArray[Any],
     input_path: Path,
     cache_dir: Path,
     inference_dtype: torch.dtype,
@@ -311,7 +312,7 @@ def integrate_water_detection_methods(
     mosaic_device: Union[str, torch.device] = "cpu",
     no_data_value: int = 0,
     optimise_model: bool = True,
-) -> tuple[np.ndarray, list[str], Optional[np.ndarray]]:
+) -> tuple[NDArray[Any], list[str], Optional[NDArray[Any]]]:
     """Combine the NDWI, model predictions and vector targets.
 
     Returns the stacked output array, the per-band layer names, and an optional

omniwatermask/download_models.py

@@ -25,6 +25,9 @@ dependencies = [
     "tqdm>=4.0",
     "platformdirs>=4.0.0",
     "fastai>=2.7",
+    # spaCy 3.8.14 advertises Python 3.14 support but does not publish an
+    # installable cp314 artifact for the GitHub Actions Linux runner.
+    "spacy!=3.8.14",
 ]
 license = "MIT"
 license-files = ["LICENSE"]