v7: yolo11s+validator-pseudo fp16 (mAP50 0.83 on holdout)

yolo11s fine-tuned on Roboflow + 70 validator-pseudo frames (top-1 peer labels, cs=0.683). fp16 ONNX 19.27MB. miner.py auto-detects fp16 input dtype.

best.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:4c4604cea567e2ab85bd9f9db3ec027ef57d0d0ef0047108d4c069d8a6debd71
-size 11110309
+oid sha256:8beaa959e6efb3b0f82ac0876fe14bfe1d8d5dfaeecb39ca31d3430473bd3ed9
+size 19271662

miner.py

@@ -59,8 +59,12 @@ class Miner:
             sess_options=so,
         )
         self.inp = self.sess.get_inputs()[0].name
+        # ONNX может быть экспортирован в fp16 (для лимита репо ≤30MB) —
+        # кастим вход в тот же dtype, иначе INVALID_ARGUMENT на sess.run.
+        _ort_type = self.sess.get_inputs()[0].type  # e.g. "tensor(float16)"
+        self.np_dtype = np.float16 if "float16" in _ort_type else np.float32
         active = self.sess.get_providers()[0]
-        print(f"✅ ONNX beverage model loaded (provider={active})")
+        print(f"✅ ONNX beverage model loaded (provider={active}, dtype={self.np_dtype.__name__})")
 
         # Eager CUDA EP allocation: ORT lazily binds CUDA on first sess.run,
         # so without this the validator's first /predict eats the cold-bind
@@ -89,8 +93,12 @@ class Miner:
 
     def _infer(self, im_bgr: ndarray) -> ndarray:
         lb, s = self._letterbox(im_bgr)
-        x = lb[:, :, ::-1].transpose(2, 0, 1)[None].astype(np.float32) / 255.0
+        x = (lb[:, :, ::-1].transpose(2, 0, 1)[None].astype(np.float32) / 255.0
+             ).astype(self.np_dtype)
         out = self.sess.run(None, {self.inp: x})[0][0]      # (4+nc, N)
+        # ONNX fp16 → numpy float16 в out; для последующего NMS на CPU
+        # удобнее float32, кастим обратно
+        out = np.asarray(out, dtype=np.float32)
         p = out.T if out.shape[0] < out.shape[1] else out    # (N, 4+nc)
         boxes = p[:, :4].copy()
         scores = p[:, 4:4 + self.num_classes]