changing the onnxwrapper script for gpu issue

mmdeploy/backend/onnxruntime/wrapper.py

@@ -50,18 +50,17 @@ def __init__(self,
             logger.warning(f'The library of onnxruntime custom ops does \
             not exist: {ort_custom_op_path}')
         device_id = parse_device_id(device)
-        is_cuda_available = ort.get_device() == 'GPU'
-        providers = [('CUDAExecutionProvider', {'device_id': device_id})] \
-            if is_cuda_available else ['CPUExecutionProvider']
+        providers = ['CPUExecutionProvider'] \
+            if device == 'cpu' else \
+            [('CUDAExecutionProvider', {'device_id': device_id})]
         sess = ort.InferenceSession(
             onnx_file, session_options, providers=providers)
         if output_names is None:
             output_names = [_.name for _ in sess.get_outputs()]
         self.sess = sess
         self.io_binding = sess.io_binding()
         self.device_id = device_id
-        self.is_cuda_available = is_cuda_available
-        self.device_type = 'cuda' if is_cuda_available else 'cpu'
+        self.device_type = 'cpu' if device == 'cpu' else 'cuda'
         super().__init__(output_names)
 
     def forward(self, inputs: Dict[str,
@@ -77,7 +76,7 @@ def forward(self, inputs: Dict[str,
         for name, input_tensor in inputs.items():
             # set io binding for inputs/outputs
             input_tensor = input_tensor.contiguous()
-            if not self.is_cuda_available:
+            if self.device_type == 'cpu':
                 input_tensor = input_tensor.cpu()
             # Avoid unnecessary data transfer between host and device
             element_type = input_tensor.new_zeros(

requirements/runtime.txt

@@ -2,5 +2,6 @@ h5py
 matplotlib
 numpy
 onnx>=1.8.0
+protobuf==3.20.0
 six
 terminaltables