Merge pull request #351 from MannLabs/fix_346

Fix 346

Author: sophiamaedler

.gitignore

@@ -8,6 +8,8 @@ config_test.yml
 segmentation_workflows/
 scportrait_data/
 
+#do not track vscode settings
+.vscode/
 
 #do not track output generated by sphinx-gallery
 sg_execution_times.rst

src/scportrait/pipeline/segmentation/workflows/_model_caches.py

@@ -12,12 +12,9 @@ def _make_zenodo_download_link(record_id: str, filename: str) -> str:
     """
     Construct a direct download URL for a file stored in a Zenodo record.
 
-    Parameters
-    ----------
-    record_id : str
-        The Zenodo record identifier (e.g., "1234567").
-    filename : str
-        The exact filename stored in the Zenodo record (case sensitive).
+    Args:
+        record_id : The Zenodo record identifier (e.g., "1234567").
+        filename : The exact filename stored in the Zenodo record (case sensitive).
 
     Returns
     -------
@@ -28,7 +25,11 @@ def _make_zenodo_download_link(record_id: str, filename: str) -> str:
 
 
 def _scportrait_cache_model_path(basename: str) -> None:
-    """Download a model from a public Nextcloud share into Cellpose's model cache if missing."""
+    """Download a model from a public Zenodo share into Cellpose's model cache if missing.
+
+    Args:
+
+    """
     MODEL_DIR.mkdir(parents=True, exist_ok=True)
 
     url = _make_zenodo_download_link(

tests/conftest.py

@@ -0,0 +1,62 @@
+import shutil
+
+import numpy as np
+import pandas as pd
+import pytest
+from anndata import AnnData
+from spatialdata import SpatialData
+from spatialdata.datasets import blobs
+
+rng = np.random.default_rng()
+
+
+@pytest.fixture
+def h5sc_object() -> AnnData:
+    # Two cells, two channels, small images
+    cell_ids = [101, 102, 107, 109]
+    n_cells = 4
+    channel_names = np.array(["seg_all_nucleus", "ch0", "ch1"])
+    channel_mapping = np.array(["mask", "image", "image"])  # or whatever mapping your code expects
+    n_channels = len(channel_names)
+    H, W = 10, 10
+
+    # --- obs ---
+    obs = pd.DataFrame({"scportrait_cell_id": cell_ids}, index=np.arange(n_cells))
+
+    # --- var (channel metadata) ---
+    var = pd.DataFrame(index=np.arange(n_channels).astype("str"))
+    var["channels"] = channel_names
+    var["channel_mapping"] = channel_mapping
+
+    adata = AnnData(obs=obs, var=var)
+    adata.obsm["single_cell_images"] = rng.random((n_cells, n_channels, H, W))
+    adata.uns["single_cell_images"] = {
+        "channel_mapping": channel_mapping,
+        "channel_names": channel_names,
+        "compression": "lzf",
+        "image_size": np.int64(H),
+        "n_cells": np.int64(n_cells),
+        "n_channels": np.int64(n_channels),
+        "n_image_channels": np.int64(n_channels - 1),
+        "n_masks": np.int64(1),
+    }
+
+    yield adata
+
+
+@pytest.fixture()
+def sdata(tmp_path) -> SpatialData:
+    sdata = blobs()
+    # Write to temporary location
+    sdata_path = tmp_path / "sdata.zarr"
+    sdata.write(sdata_path)
+    yield sdata
+    shutil.rmtree(sdata_path)
+
+
+@pytest.fixture
+def sdata_with_labels() -> SpatialData:
+    sdata = blobs()
+    sdata["table"].obs["labelling_categorical"] = sdata["table"].obs["instance_id"].astype("category")
+    sdata["table"].obs["labelling_continous"] = (sdata["table"].obs["instance_id"] > 10).astype(float)
+    return sdata

tests/unit_tests/pipeline/test_featurization.py

@@ -0,0 +1,64 @@
+import pytest
+import torch
+
+from scportrait.pipeline.featurization import CellFeaturizer  # adjust import to your project
+
+
+def test_cell_featurizer(tmp_path):
+    # temp directory unique to this test
+    out_dir = tmp_path / "featurization"
+    out_dir.mkdir()
+
+    config = {
+        "batch_size": 100,
+        "dataloader_worker_number": 10,
+    }
+
+    f = CellFeaturizer(
+        config=config,
+        directory=str(out_dir),
+        project_location=None,
+        overwrite=True,
+    )
+
+    # Image: 1 batch, 1 channel, 4x4
+    img = torch.zeros((1, 1, 4, 4), dtype=torch.float32)
+
+    # Define a "cell" region and a "nucleus" region inside it
+    nucleus_mask = torch.zeros((1, 1, 4, 4), dtype=torch.float32)
+    cell_mask = torch.zeros((1, 1, 4, 4), dtype=torch.float32)
+
+    cell_mask[..., 1:3, 1:3] = 1.0  # 2x2 cell block
+    nucleus_mask[..., 1:2, 1:2] = 1.0  # 1x1 nucleus (top-left of cell)
+
+    # Intensities: nucleus=10, cytosol=1
+    img[cell_mask.bool()] = 1.0
+    img[nucleus_mask.bool()] = 10.0
+
+    # Build label stack
+    labels = torch.cat([nucleus_mask, cell_mask], dim=1)  # shape (1, 2, 4, 4)
+
+    # Run featurization on concatenated masks and images
+    feats = f.calculate_statistics(torch.cat([labels, img], dim=1))
+    column_names = f._generate_column_names(n_masks=2, channel_names=["ch0"])
+    feat_map = dict(zip(column_names, feats[0].tolist(), strict=True))
+
+    # Key regression assertion:
+    # If masks accidentally become all-True, these will match (or be very close).
+    nuc_mean = feat_map["ch0_mean_nucleus"]
+    cyto_mean = feat_map["ch0_mean_cytosol"]
+
+    assert nuc_mean > cyto_mean
+    assert abs(nuc_mean - cyto_mean) > 1.0
+
+
+def test_mask_bool_integrity_not_corrupted():
+    mask = torch.tensor([[True, False], [False, True]])
+    before_true = mask.sum().item()
+
+    # This simulates the buggy behavior:
+    # mask[mask == 0] = torch.nan  # would cast nan->True and change mask
+    # Instead, ensure your implementation never does this.
+
+    assert mask.dtype == torch.bool
+    assert mask.sum().item() == before_true

tests/unit_tests/plotting/test_sdata.py

@@ -10,14 +10,6 @@
 from scportrait.plotting import sdata as plotting
 
 
-@pytest.fixture
-def sdata():
-    sdata = blobs()
-    sdata["table"].obs["labelling_categorical"] = sdata["table"].obs["instance_id"].astype("category")
-    sdata["table"].obs["labelling_continous"] = (sdata["table"].obs["instance_id"] > 10).astype(float)
-    return sdata  # provides images and labels used in tests
-
-
 @pytest.mark.parametrize(
     "channel_names, palette, return_fig, show_fig",
     [
@@ -26,9 +18,9 @@ def sdata():
         ([0, 1], None, False, False),
     ],
 )
-def test_plot_image(sdata, channel_names, palette, return_fig, show_fig):
+def test_plot_image(sdata_with_labels, channel_names, palette, return_fig, show_fig):
     fig = plotting.plot_image(
-        sdata=sdata,
+        sdata=sdata_with_labels,
         image_name="blobs_image",
         channel_names=channel_names,
         palette=palette,
@@ -51,9 +43,9 @@ def test_plot_image(sdata, channel_names, palette, return_fig, show_fig):
         (None, None),  # test only mask overlay without image
     ],
 )
-def test_plot_segmentation_mask(sdata, selected_channels, background_image):
+def test_plot_segmentation_mask(sdata_with_labels, selected_channels, background_image):
     fig = plotting.plot_segmentation_mask(
-        sdata=sdata,
+        sdata=sdata_with_labels,
         masks=["blobs_labels"],
         background_image=background_image,
         selected_channels=selected_channels,
@@ -73,9 +65,9 @@ def test_plot_segmentation_mask(sdata, selected_channels, background_image):
         (True, "labelling_continous"),
     ],
 )
-def test_plot_labels(sdata, vectorized, color):
+def test_plot_labels(sdata_with_labels, vectorized, color):
     fig = plotting.plot_labels(
-        sdata=sdata,
+        sdata=sdata_with_labels,
         label_layer="blobs_labels",
         vectorized=vectorized,
         color=color,

tests/unit_tests/tools/sdata/test_write.py

@@ -7,16 +7,6 @@
 from spatialdata.datasets import blobs
 
 
-@pytest.fixture()
-def sdata(tmp_path):
-    sdata = blobs()
-    # Write to temporary location
-    sdata_path = tmp_path / "sdata.zarr"
-    sdata.write(sdata_path)
-    yield sdata
-    shutil.rmtree(sdata_path)
-
-
 @pytest.fixture()
 def sdata_path(tmp_path):
     # Write to temporary location