Upload export_medsiglip.py with huggingface_hub

export_medsiglip.py

@@ -233,24 +233,20 @@ def export_to_onnx(model, config: Config, quantize: bool = True):
     onnx_path = config.output_dir / "laborview_medsiglip.onnx"
     onnx_quant_path = config.output_dir / "laborview_medsiglip_int8.onnx"
 
-    print(f"Exporting to ONNX (legacy exporter): {onnx_path}")
+    print(f"Exporting to ONNX: {onnx_path}")
 
-    # Trace the model first
+    # Export directly without tracing (for new PyTorch exporter compatibility)
     with torch.no_grad():
-        traced = torch.jit.trace(wrapper, dummy_input)
-
-    # Export traced model to ONNX (fixed batch size for compatibility)
-    torch.onnx.export(
-        traced,
-        dummy_input,
-        str(onnx_path),
-        export_params=True,
-        opset_version=14,  # Use stable opset version
-        do_constant_folding=True,
-        input_names=['pixel_values'],
-        output_names=['seg_probs', 'plane_pred']
-        # No dynamic_axes - use fixed batch size of 1 for mobile
-    )
+        torch.onnx.export(
+            wrapper,
+            (dummy_input,),
+            str(onnx_path),
+            export_params=True,
+            opset_version=14,
+            do_constant_folding=True,
+            input_names=['pixel_values'],
+            output_names=['seg_probs', 'plane_pred']
+        )
 
     # Verify ONNX model
     onnx_model = onnx.load(str(onnx_path))