Image transformers that do not use domain transformations

Some image transformation methods need the whole image because they do
both parameter estimation (=learning) and actual transformation (using
learned parameters) in a single operation. Therefore, when transforming,
instances are not independent and thus this can not be solved with
domain transformation in Orange.

orangecontrib/snom/preprocess/utils.py

@@ -27,6 +27,38 @@ def image_transformer(self, data, image_opts):
         raise NotImplementedError
 
 
+class NoComputeValue:
+    def __call__(self, data):
+        return np.full(len(data), np.nan)
+
+
+class PreprocessImageOpts2DOnlyWhole(PreprocessImageOpts):
+    def __call__(self, data, image_opts):
+        at = data.domain[image_opts["attr_value"]].copy(compute_value=NoComputeValue())
+        odata = data
+        domain = domain_with_single_attribute_in_x(at, data.domain)
+        data = data.transform(domain)
+        if len(data):
+            with data.unlocked(data.X):
+                data.X[:, 0] = odata.get_column(image_opts["attr_value"], copy=True)
+        try:
+            hypercube, _, indices = get_ndim_hyperspec(
+                data, (image_opts["attr_x"], image_opts["attr_y"])
+            )
+            image = hypercube[:, :, 0]
+            transformed = self.transform_image(image)
+            col = transformed[indices].reshape(-1)
+        except InvalidAxisException:
+            col = np.full(len(data), np.nan)
+        if len(data):
+            with data.unlocked(data.X):
+                data.X[:, 0] = col
+        return data
+
+    def transform_image(self, image):
+        raise NotImplementedError
+
+
 def axes_to_ndim_linspace(coordinates):
     # modified to avoid domains as much as possible
     ls = []

orangecontrib/snom/widgets/preprocessors/background_fit.py

@@ -6,15 +6,13 @@
 from pySNOM.images import BackgroundPolyFit
 
 from orangecontrib.snom.preprocess.utils import (
-    PreprocessImageOpts2D,
-    CommonDomainImage2D,
+    PreprocessImageOpts2DOnlyWhole,
 )
 from orangecontrib.snom.widgets.preprocessors.registry import preprocess_image_editors
 
 
-class _BackGroundFitCommon(CommonDomainImage2D):
-    def __init__(self, xorder, yorder, domain, image_opts):
-        super().__init__(domain, image_opts)
+class BackGroundFit(PreprocessImageOpts2DOnlyWhole):
+    def __init__(self, xorder=1, yorder=1):
         self.xorder = xorder
         self.yorder = yorder
 
@@ -25,15 +23,6 @@ def transform_image(self, image):
         return d
 
 
-class BackGroundFit(PreprocessImageOpts2D):
-    def __init__(self, xorder=1, yorder=1):
-        self.xorder = xorder
-        self.yorder = yorder
-
-    def image_transformer(self, data, image_opts):
-        return _BackGroundFitCommon(self.xorder, self.yorder, data.domain, image_opts)
-
-
 class BackGroundFitEditor(BaseEditorOrange):
     name = "Polynomial background fit"
     qualname = "orangecontrib.snom.background_fit_test"

orangecontrib/snom/widgets/preprocessors/linelevel.py

@@ -8,28 +8,18 @@
 
 from orangecontrib.snom.widgets.preprocessors.registry import preprocess_image_editors
 from orangecontrib.snom.preprocess.utils import (
-    PreprocessImageOpts2D,
-    CommonDomainImage2D,
+    PreprocessImageOpts2DOnlyWhole,
 )
 
 
-class _LineLevelCommon(CommonDomainImage2D):
-    def __init__(self, method, domain, image_opts):
-        super().__init__(domain, image_opts)
+class LineLevelProcessor(PreprocessImageOpts2DOnlyWhole):
+    def __init__(self, method="median"):
         self.method = method
 
     def transform_image(self, image):
         return LineLevel(method=self.method).transform(image)
 
 
-class LineLevelProcessor(PreprocessImageOpts2D):
-    def __init__(self, method="median"):
-        self.method = method
-
-    def image_transformer(self, data, image_opts):
-        return _LineLevelCommon(self.method, data.domain, image_opts)
-
-
 class LineLevelEditor(BaseEditorOrange):
     name = "Line leveling"
     qualname = "orangecontrib.snom.line_level_test"