Faster ONNX inference with bindings (#17184)

Co-authored-by: UltralyticsAssistant <web@ultralytics.com>
Co-authored-by: Burhan <62214284+Burhan-Q@users.noreply.github.com>
Co-authored-by: Ultralytics Assistant <135830346+UltralyticsAssistant@users.noreply.github.com>
Co-authored-by: Glenn Jocher <glenn.jocher@ultralytics.com>

ultralytics/nn/autobackend.py

@@ -189,10 +189,32 @@ class AutoBackend(nn.Module):
                 check_requirements("numpy==1.23.5")
             import onnxruntime
 
-            providers = ["CUDAExecutionProvider", "CPUExecutionProvider"] if cuda else ["CPUExecutionProvider"]
+            providers = onnxruntime.get_available_providers()
+            if not cuda and "CUDAExecutionProvider" in providers:
+                providers.remove("CUDAExecutionProvider")
+            elif cuda and "CUDAExecutionProvider" not in providers:
+                LOGGER.warning("WARNING ⚠️ Failed to start ONNX Runtime session with CUDA. Falling back to CPU...")
+                device = torch.device("cpu")
+                cuda = False
+            LOGGER.info(f"Preferring ONNX Runtime {providers[0]}")
             session = onnxruntime.InferenceSession(w, providers=providers)
             output_names = [x.name for x in session.get_outputs()]
             metadata = session.get_modelmeta().custom_metadata_map
+            dynamic = isinstance(session.get_outputs()[0].shape[0], str)
+            if not dynamic:
+                io = session.io_binding()
+                bindings = []
+                for output in session.get_outputs():
+                    y_tensor = torch.empty(output.shape, dtype=torch.float16 if fp16 else torch.float32).to(device)
+                    io.bind_output(
+                        name=output.name,
+                        device_type=device.type,
+                        device_id=device.index if cuda else 0,
+                        element_type=np.float16 if fp16 else np.float32,
+                        shape=tuple(y_tensor.shape),
+                        buffer_ptr=y_tensor.data_ptr(),
+                    )
+                    bindings.append(y_tensor)
 
         # OpenVINO
         elif xml:
@@ -477,8 +499,22 @@ class AutoBackend(nn.Module):
 
         # ONNX Runtime
         elif self.onnx:
-            im = im.cpu().numpy()  # torch to numpy
-            y = self.session.run(self.output_names, {self.session.get_inputs()[0].name: im})
+            if self.dynamic:
+                im = im.cpu().numpy()  # torch to numpy
+                y = self.session.run(self.output_names, {self.session.get_inputs()[0].name: im})
+            else:
+                if not self.cuda:
+                    im = im.cpu()
+                self.io.bind_input(
+                    name="images",
+                    device_type=im.device.type,
+                    device_id=im.device.index if im.device.type == "cuda" else 0,
+                    element_type=np.float16 if self.fp16 else np.float32,
+                    shape=tuple(im.shape),
+                    buffer_ptr=im.data_ptr(),
+                )
+                self.session.run_with_iobinding(self.io)
+                y = self.bindings
 
         # OpenVINO
         elif self.xml: