Initialize subclass attributes on classifier load

Adds initialization of subclass-specific attributes (such as default_threshold, min_predictions, max_predictions, and label_thresholds) in AdaptiveClassifier.load to ensure proper state after loading. Updates tests to handle ONNX and non-ONNX models correctly when setting eval mode and loading classifiers.

Author: codelion

src/adaptive_classifier/classifier.py

@@ -832,6 +832,17 @@ def _from_pretrained(
             classifier.strategic_optimizer = None
             classifier.strategic_evaluator = None
 
+            # Initialize subclass-specific attributes (e.g., for MultiLabelAdaptiveClassifier)
+            # These will be overwritten if the subclass has its own initialization logic
+            if not hasattr(classifier, 'default_threshold'):
+                classifier.default_threshold = 0.5
+            if not hasattr(classifier, 'min_predictions'):
+                classifier.min_predictions = 1
+            if not hasattr(classifier, 'max_predictions'):
+                classifier.max_predictions = None
+            if not hasattr(classifier, 'label_thresholds'):
+                classifier.label_thresholds = {}
+
             if classifier.config.enable_strategic_mode:
                 classifier._initialize_strategic_components()
         else:

tests/test_classifier.py

@@ -60,15 +60,16 @@ def test_save_load(base_classifier, sample_data):
     torch.manual_seed(42)
     np.random.seed(42)
     random.seed(42)
-    
+
     texts, labels = sample_data
     base_classifier.add_examples(texts, labels)
-    
+
     with tempfile.TemporaryDirectory() as tmpdir:
         save_path = Path(tmpdir) / "test_classifier"
-        
-        # Ensure model is in eval mode before saving
-        base_classifier.model.eval()
+
+        # Ensure model is in eval mode before saving (if not ONNX)
+        if not base_classifier.use_onnx and hasattr(base_classifier.model, 'eval'):
+            base_classifier.model.eval()
         if base_classifier.adaptive_head is not None:
             base_classifier.adaptive_head.eval()
 
@@ -81,13 +82,14 @@ def test_save_load(base_classifier, sample_data):
         assert (save_path / "examples.json").exists()
         assert (save_path / "README.md").exists()
 
-        # Load with same device
-        loaded_classifier = AdaptiveClassifier.load(save_path, device=base_classifier.device)
+        # Load with same device (disable ONNX for deterministic comparison)
+        loaded_classifier = AdaptiveClassifier.load(save_path, device=base_classifier.device, use_onnx=False)
         assert loaded_classifier is not None
         assert loaded_classifier.label_to_id == base_classifier.label_to_id
-        
-        # Ensure loaded model is also in eval mode
-        loaded_classifier.model.eval()
+
+        # Ensure loaded model is also in eval mode (if not ONNX)
+        if not loaded_classifier.use_onnx and hasattr(loaded_classifier.model, 'eval'):
+            loaded_classifier.model.eval()
         if loaded_classifier.adaptive_head is not None:
             loaded_classifier.adaptive_head.eval()