fix a problem for fully nan feature vectors, this is now handled but …

…not properly tested in calibration_smoothing.py

src/depiction/calibration/spectrum/calibration_smoothing.py

@@ -1,23 +1,52 @@
 from __future__ import annotations
 
+from typing import Callable
+
 import numpy as np
+import xarray
 from xarray import DataArray
 
 from depiction.image.spatial_smoothing_sparse_aware import SpatialSmoothingSparseAware
 
 
-# TODO should be refactored later
+def _apply_on_spatial_view(array: DataArray, fn: Callable[[DataArray], DataArray]) -> DataArray:
+    # adjust indexing for the rest of the function
+    array_flat = array.drop("i").set_xindex(["x", "y"])
 
+    # perform the computation on 2d view
+    array_2d = array_flat.unstack("i").transpose("y", "x", "c")
+    array_2d = fn(array_2d)
 
-def smooth_image_features(all_features: DataArray, kernel_size: int, kernel_std: float) -> DataArray:
-    features_flat = all_features.drop("i").set_xindex(["x", "y"])
-    features_2d = features_flat.unstack("i").transpose("y", "x", "c")
-    smoother = SpatialSmoothingSparseAware(
-        kernel_size=kernel_size,
-        kernel_std=kernel_std,
+    # trick: concatenate an additional channel that will indicate everything that was present before, including
+    #        fully nan elements because they should not disappear like an actual background
+    # this almost works but is broken:
+    is_nan_before = (
+        array_flat.isnull().all("c").astype(array_2d.dtype).expand_dims("c").unstack("i").transpose("y", "x", "c")
     )
-    result = smoother.smooth(features_2d, bg_value=np.nan)
-    # TODO a bit ugly...
-    result = result.stack(i=("x", "y")).dropna("i", how="all")
+    array_2d = xarray.concat([array_2d, is_nan_before], dim="c")
+
+    # stack back the 2d view
+    result = array_2d.stack(i=("x", "y")).dropna("i", how="all")
+
+    # remove the additional channel
+    result = result.isel(c=slice(0, -1))
+
+    # revert the indexing
     x, y = result.x.values, result.y.values
     return result.drop_vars(["i", "x", "y"]).assign_coords(x=("i", x), y=("i", y), i=np.arange(len(result.i)))
+
+
+# TODO should be refactored later
+
+# TODO test this case : a spectrum is all nan before, but present in the flat repr, it should not disappear like an actual background
+
+
+def smooth_image_features(all_features: DataArray, kernel_size: int, kernel_std: float) -> DataArray:
+    def fn(array_2d: DataArray) -> DataArray:
+        smoother = SpatialSmoothingSparseAware(
+            kernel_size=kernel_size,
+            kernel_std=kernel_std,
+        )
+        return smoother.smooth(array_2d, bg_value=np.nan)
+
+    return _apply_on_spatial_view(all_features, fn)