Cap regional failure oversampling

backend/tests/test_failure_driven_yolo_sampling.py

@@ -55,7 +55,7 @@ def test_sampling_repeats_only_failed_region_train_tiles() -> None:
         "background": {"pure_empty_false_positives": 2},
     }
     paths, metadata = MODULE.build_sampling(
-        summary=summary, manifest=manifest, assessment=assessment
+        summary=summary, manifest=manifest, assessment=assessment, max_region_share=1.0
     )
     assert paths.count(str(Path("/tmp/fl-pos.png").resolve())) == 3
     assert paths.count(str(Path("/tmp/fl-neg.png").resolve())) == 4
@@ -87,7 +87,7 @@ def test_sampling_can_use_calibration_before_test_is_opened() -> None:
         "background": None,
     }
     paths, metadata = MODULE.build_sampling(
-        summary=summary, manifest=manifest, assessment=assessment
+        summary=summary, manifest=manifest, assessment=assessment, max_region_share=1.0
     )
     assert len(paths) == 3
     assert metadata["failure_evidence_source"] == "calibration"
@@ -120,6 +120,7 @@ def test_precision_correction_can_balance_positive_and_negative_tiles() -> None:
         assessment=assessment,
         precision_positive_repeat=2,
         negative_repeat=3,
+        max_region_share=1.0,
     )
 
     assert paths.count(str(Path("/tmp/fl-pos.png").resolve())) == 2
@@ -157,7 +158,9 @@ def test_sampling_targets_failed_calibration_contexts_without_using_protected_ti
         },
     }
 
-    paths, metadata = MODULE.build_sampling(summary=summary, manifest=manifest, assessment=assessment)
+    paths, metadata = MODULE.build_sampling(
+        summary=summary, manifest=manifest, assessment=assessment, max_region_share=1.0
+    )
 
     assert paths.count(str(Path("/tmp/industry-pos.png").resolve())) == 5
     assert paths.count(str(Path("/tmp/industry-neg.png").resolve())) == 6
@@ -165,3 +168,42 @@ def test_sampling_targets_failed_calibration_contexts_without_using_protected_ti
     assert not any("protected" in path for path in paths)
     assert metadata["weak_recall_contexts"] == ["flanders:industrial"]
     assert metadata["weak_precision_contexts"] == ["flanders:industrial"]
+
+
+def test_region_cap_drops_only_repeats_and_preserves_every_unique_tile() -> None:
+    manifest = {"samples": [
+        {"sample_slug": "fl", "split": "train", "region": "flanders", "context": "industrial"},
+        {"sample_slug": "wa", "split": "train", "region": "wallonia", "context": "rural-town"},
+        {"sample_slug": "br", "split": "train", "region": "brussels", "context": "dense-urban"},
+    ]}
+    summary = {"tiles": [
+        {"sample_slug": "fl", "split": "train", "label_count": 2, "image_path": f"/tmp/fl-{index}.png"}
+        for index in range(4)
+    ] + [
+        {"sample_slug": "wa", "split": "train", "label_count": 2, "image_path": f"/tmp/wa-{index}.png"}
+        for index in range(2)
+    ] + [
+        {"sample_slug": "br", "split": "train", "label_count": 2, "image_path": f"/tmp/br-{index}.png"}
+        for index in range(2)
+    ]}
+    assessment = {
+        "status": "continue_training_loop",
+        "gates": {"min_region_f1": .45, "min_region_precision": .5, "min_region_recall": .4,
+                  "max_pure_empty_false_positives": 0},
+        "calibration": {"regions": {
+            "flanders": {"f1": .2, "precision": .3, "recall": .2},
+            "wallonia": {"f1": .6, "precision": .6, "recall": .6},
+            "brussels": {"f1": .6, "precision": .6, "recall": .6},
+        }},
+    }
+
+    paths, metadata = MODULE.build_sampling(
+        summary=summary, manifest=manifest, assessment=assessment,
+        positive_repeat=5, max_region_share=.65,
+    )
+
+    assert all(str(Path(f"/tmp/fl-{index}.png").resolve()) in paths for index in range(4))
+    assert metadata["pre_cap_entries_by_region"]["flanders"] == 20
+    assert metadata["sampled_entries_by_region"]["flanders"] == 7
+    assert metadata["dropped_region_repeat_count"] == 13
+    assert metadata["sampled_entries_by_region"]["flanders"] / len(paths) <= .65