scorevision: push artifact

miner.py

@@ -1,18 +1,21 @@
-# build-rev: 2026-05-28-02 (forces new HF SHA → new chute nickname; prior swan build failed on Chutes registry push)
-"""Open-source Detect-beverage miner v12 (new yolo11s weights + NMS-baked ONNX).
+# build-rev: 2026-05-28-v14 (yolo11s trained on validator-IDENTICAL SAM3-GT)
+"""Open-source Detect-beverage miner v14 (yolo11s trained on SAM3 validator-GT).
 
-New v12 ONNX weights: trained on combined dataset (375 fresh cross-consensus
-pseudo-GT from top 3 Beverage miners + 279 prior validator-pseudo). NMS-baked
-export — output shape `[1, 300, 6]` (xyxy, conf, cls). On holdout 73 imgs:
-ultralytics-val mAP50=0.879 (v8: 0.835).
+Trained on 329 frames labelled by the SAME SAM3 endpoint the validator uses to
+build pseudo-GT (prompts cup/bottle/can, thr 0.5, mosaic 0) — i.e. the actual
+scoring target, not peer miners' boxes. NMS-baked ONNX, output [1,300,6].
+
+On 50 SAM3-GT holdout (validator-target): mAP50=0.715 (navierstocks 0.673,
+v12 0.645); best composite UI 63.47% (nav 62.91%, v12 61.97%). Beats peers on
+detection; parity-plus on composite.
 
 Post-proc:
 - detect NMS-baked output and unpack to (N, 4+num_classes) one-hot scores
-- per-class conf filter `[0.60, 0.40, 0.60]` (best from sweep on v12 ONNX)
+- per-class conf filter `[0.70, 0.50, 0.50]` (best v14 sweep on SAM3-GT)
 - sane-box geometric filter (min_box_area=100, max_aspect_ratio=10)
 - per-class hard NMS @ iou=0.4 (redundant after baked NMS but safe)
-- cross-class dedup @ iou=0.7
-- TTA off (sweep showed flip-TTA + cluster-boost hurt UI on NMS-baked output)
+- cross-class dedup @ iou=0.7 (helps bottle↔can misclassification FP)
+- TTA off
 
 Contract: class `Miner` at HF root, `predict_batch(...) -> list[TVFrameResult]`.
 """
@@ -48,9 +51,9 @@ class Miner:
     input_size = 1280
     num_classes = 3                          # cup, bottle, can
 
-    # per-class conf — best v12 sweep without TTA (UI 80.32%, FP 0.832):
-    conf_thres = np.array([0.60, 0.40, 0.60], dtype=np.float32)
-    # rescue bonus disabled — v12 model strong enough not to need it
+    # per-class conf — best v14 sweep on SAM3-GT holdout (composite 63.47%):
+    conf_thres = np.array([0.70, 0.50, 0.50], dtype=np.float32)
+    # rescue bonus disabled
     rescue_bonus = np.array([0.0, 0.0, 0.0], dtype=np.float32)
 
     iou_thres = 0.40                         # per-class NMS (redundant after baked-NMS but safe)
@@ -78,7 +81,7 @@ class Miner:
         _ort_type = self.sess.get_inputs()[0].type   # "tensor(float16)" or fp32
         self.np_dtype = np.float16 if "float16" in _ort_type else np.float32
         active = self.sess.get_providers()[0]
-        print(f"✅ v12 ONNX beverage model loaded (provider={active}, dtype={self.np_dtype.__name__})")
+        print(f"✅ v14 ONNX beverage model loaded (provider={active}, dtype={self.np_dtype.__name__})")
 
         # Detect output format once
         out0 = self.sess.get_outputs()[0]
@@ -89,12 +92,12 @@ class Miner:
         try:
             dummy = np.zeros((self.input_size, self.input_size, 3), dtype=np.uint8)
             _ = self._infer(dummy)
-            print(f"✅ v12 ONNX warmup pass completed (provider={active})")
+            print(f"✅ v14 ONNX warmup pass completed (provider={active})")
         except Exception as e:
-            print(f"⚠️ v12 ONNX warmup pass failed (not fatal): {e}")
+            print(f"⚠️ v14 ONNX warmup pass failed (not fatal): {e}")
 
     def __repr__(self) -> str:
-        return f"BeverageONNXv12(in={self.input_size}, cls={self.num_classes})"
+        return f"BeverageONNXv14(in={self.input_size}, cls={self.num_classes})"
 
     # ---- preprocessing --------------------------------------------------
     def _letterbox(self, im: ndarray) -> tuple[ndarray, float]: