Expose detections_per_img and topk_candidates in config

src/deepforest/conf/config.yaml

@@ -11,7 +11,8 @@ batch_size: 1
 architecture: 'retinanet'
 nms_thresh: 0.05
 score_thresh: 0.1
-
+detections_per_img: 300
+topk_candidates: 1000
 # Set model name to None to initialize from scratch
 model:
     name: 'weecology/deepforest-tree'

src/deepforest/conf/schema.py

@@ -138,6 +138,8 @@ class Config:
 
     nms_thresh: float = 0.05
     score_thresh: float = 0.1
+    detections_per_img: int = 300
+    topk_candidates: int = 1000
     model: ModelConfig = field(default_factory=ModelConfig)
 
     log_root: str = "./"

src/deepforest/models/retinanet.py

@@ -199,6 +199,8 @@ def create_model(
                 nms_thresh=self.config.nms_thresh,
                 score_thresh=self.config.score_thresh,
                 label_dict=label_dict,
+                detections_per_img=self.config.detections_per_img,
+                topk_candidates=self.config.topk_candidates,
             )
         else:
             # Pre 2.0 compatibility, the score_threshold used to be stored under retinanet.score_thresh
@@ -214,6 +216,8 @@ def create_model(
                 label_dict=label_dict,
                 nms_thresh=self.config.nms_thresh,
                 score_thresh=self.config.score_thresh,
+                detections_per_img=self.config.detections_per_img,
+                topk_candidates=self.config.topk_candidates,
                 **hf_args,
             )
 

tests/test_main.py

@@ -1173,6 +1173,7 @@ def test_custom_log_root(m, tmpdir):
     version_dir = version_dirs[0]
     assert version_dir.join("hparams.yaml").exists(), "hparams.yaml not found"
 
+
 def test_huggingface_model_loads_correct_label_dict():
     """Regression test for #1286:
     HuggingFace models should load correct label_dict from config.json.
@@ -1194,3 +1195,24 @@ def test_huggingface_model_loads_correct_label_dict():
 
     actual = set(m.label_dict.keys())
     assert actual == expected, f"Expected {expected}, got {actual}"
+
+def test_detections_per_img_and_topk_candidates_config():
+    """Test that detections_per_img and topk_candidates can be configured
+    and are passed through to the underlying model."""
+    m = main.deepforest()
+
+    # Check default values
+    assert m.config.detections_per_img == 300
+    assert m.config.topk_candidates == 1000
+
+    # Test custom values
+    m.config.detections_per_img = 500
+    m.config.topk_candidates = 2000
+
+    assert m.config.detections_per_img == 500
+    assert m.config.topk_candidates == 2000
+
+    # Verify values are passed to actual model
+    m.create_model()
+    assert m.model.detections_per_img == 500
+    assert m.model.topk_candidates == 2000