fix(data): attribute preprocessing-off warning to data module and reword for non-image configs

contributor-configs/tree-xgboost-shap/assessment.yaml

@@ -21,6 +21,10 @@ data:
   name: synthetic-tabular
   source: ${hydra:runtime.cwd}/contributor-configs/tree-xgboost-shap/artifacts/features.csv
   forward_batch_size: 8
+  # Declares the modality so the resolver skips the image-only "no preprocessing"
+  # warning (image preprocessing is meaningless for tabular features).
+  input_metadata:
+    kind: tabular
 
 transparency:
   treeshap:

src/raitap/data/preprocessing.py

@@ -324,9 +324,9 @@ def _suppress_off_warning(data_cfg: DataConfig, *, acknowledge_off: bool = False
 
 
 _OFF_WARNING = (
-    "Image preprocessing is OFF.\n"
+    "No image preprocessing set in config.\n"
     "RAITAP is forwarding your images to the model unchanged.\n"
-    "Most pretrained image models expect normalized inputs — without\n"
+    "Most pretrained image models expect normalized inputs — without "
     "preprocessing, your accuracy results may be silently incorrect.\n"
     "\n"
     "Set both knobs in your config:\n"
@@ -338,15 +338,17 @@ def _suppress_off_warning(data_cfg: DataConfig, *, acknowledge_off: bool = False
     "      preprocessing: ./resize.py\n"
     "      model_input_transformation: ./normalize.py\n"
     "\n"
+    "Not using image data? Set data.input_metadata.kind (e.g. tabular) in your config.\n"
+    "\n"
     "If you've already preprocessed your images, silence this message:\n"
     "  - Python API:   pass acknowledge_preprocessing_off=True to raitap.run(...)\n"
     "  - CLI:          re-run with --acknowledge-preprocessing-off"
 )
 
 _MODEL_OFF_WARNING = (
-    "data.model_input_transformation is OFF.\n"
-    "Data-side preprocessing will run, but no transformation is applied at\n"
-    "the model boundary. Pretrained models that expect normalized inputs\n"
+    "No data.model_input_transformation set in config.\n"
+    "Data-side preprocessing will run, but no transformation is applied at "
+    "the model boundary. Pretrained models that expect normalized inputs "
     "(ImageNet mean/std, etc.) will produce silently incorrect predictions.\n"
     "\n"
     "Set the model-side knob:\n"

src/raitap/data/tests/test_preprocessing.py

@@ -32,7 +32,7 @@ def test_both_off_no_keys_warns(monkeypatch: pytest.MonkeyPatch) -> None:
     assert result.model_module is None
     assert not result.is_active
     assert len(result.warnings) == 1
-    assert "preprocessing is OFF" in result.warnings[0]
+    assert "No image preprocessing set in config" in result.warnings[0]
 
 
 def test_both_off_suppression_via_kwarg(monkeypatch: pytest.MonkeyPatch) -> None:
@@ -83,7 +83,7 @@ def test_both_off_kind_explicit_none_still_warns(monkeypatch: pytest.MonkeyPatch
     assert result.data_origin == "off"
     assert result.model_origin == "off"
     assert len(result.warnings) == 1
-    assert "preprocessing is OFF" in result.warnings[0]
+    assert "No image preprocessing set in config" in result.warnings[0]
 
 
 # ---------------------------------------------------------------------------
@@ -103,7 +103,7 @@ def test_model_input_transformation_off_warns_when_data_side_set(
     assert result.model_origin == "off"
     assert result.data_module is not None
     assert result.model_module is None
-    assert any("model_input_transformation is OFF" in w for w in result.warnings)
+    assert any("No data.model_input_transformation set in config" in w for w in result.warnings)
 
 
 def test_model_input_transformation_off_suppressed_for_tabular(
@@ -363,7 +363,7 @@ def test_custom_file_data_side_only(monkeypatch: pytest.MonkeyPatch, tmp_path: P
     assert result.model_module is None
     assert result.data_file_path == fixture.resolve()
     # model side off + image data → warning about model side
-    assert any("model_input_transformation is OFF" in w for w in result.warnings)
+    assert any("No data.model_input_transformation set in config" in w for w in result.warnings)
 
 
 def test_custom_file_both_sides_from_same_file(

src/raitap/models/model.py

@@ -162,7 +162,10 @@ def _apply_preprocessing(
             backend.model = nn.Sequential(model_module, backend.model)
 
     for warning in resolved.warnings:
-        raitap_log.warn(warning)
+        # Preprocessing is a data-module concept (see the "Preprocessing:" info
+        # line below); attribute the warning to data, not the pipeline frame the
+        # deferred-log replay would otherwise infer.
+        raitap_log.warn(warning, module=Module.data)
 
     if resolved.is_active:
         # Deferred by ``_run_pipeline`` (this runs inside ``raitap_log.deferred()``)