修改python示例程序

Python/Classifier_io_binding.py

@@ -62,32 +62,46 @@ def ort_seg_dcu(model_path,image):
     
     provider_options=[{'device_id':'0','migraphx_fp16_enable':'true'}]
     dcu_session = ort.InferenceSession(model_path, providers=['MIGraphXExecutionProvider'], provider_options=provider_options)
-    output_data = np.empty(dcu_session.get_outputs()[0].shape).astype(np.float32)
     
-    inputData={}
-    outputData={}
-    inputData[dcu_session.get_inputs()[0].name]=ort.OrtValue.ortvalue_from_numpy(image, device_type='cuda')
-    outputData[dcu_session.get_outputs()[0].name]=ort.OrtValue.ortvalue_from_numpy(output_data, device_type='cuda')
+    images = [image]
+    input_nodes = dcu_session.get_inputs()
+    input_names = [i_n.name for i_n in input_nodes]
+    output_nodes = dcu_session.get_outputs()
+    output_names = [o_n.name for o_n in output_nodes]
+
+    input_dict = {}
+    for i_d, i_n in zip(images, input_names):
+        input_dict[i_n] = i_d
 
     io_binding = dcu_session.io_binding()
-    io_binding.bind_input(
-        name=dcu_session.get_inputs()[0].name,
-        device_type=inputData[dcu_session.get_inputs()[0].name].device_name(),
-        device_id=0,
-        element_type=np.float32,
-        shape=inputData[dcu_session.get_inputs()[0].name].shape(),
-        buffer_ptr=inputData[dcu_session.get_inputs()[0].name].data_ptr())
 
+    inputData = {}
+    for key in input_dict.keys():
+        inputData[key] = ort.OrtValue.ortvalue_from_numpy(input_dict[key], device_type='cuda')
+        io_binding.bind_input(
+            name = key,
+            device_type = inputData[key].device_name(),
+            device_id = 0,
+            element_type = np.float32,
+            shape = inputData[key].shape(),
+            buffer_ptr = inputData[key].data_ptr())
+        
+    outputData = {}
+    output_data = {}
+    for index, o_n in enumerate(output_names):
+        output_data[o_n] = np.empty(dcu_session.get_outputs()[index].shape).astype(np.float32)
+        outputData[o_n] = ort.OrtValue.ortvalue_from_numpy(output_data[o_n], device_type='cuda')
         io_binding.bind_output(
-        name=dcu_session.get_outputs()[0].name,
-        device_type=outputData[dcu_session.get_outputs()[0].name].device_name(),
-        device_id=0,
-        element_type=np.float32,
-        shape=outputData[dcu_session.get_outputs()[0].name].shape(),
-        buffer_ptr=outputData[dcu_session.get_outputs()[0].name].data_ptr())
+            name = o_n,
+            device_type = outputData[o_n].device_name(),
+            device_id = 0,
+            element_type = np.float32,
+            shape = outputData[o_n].shape(),
+            buffer_ptr = outputData[o_n].data_ptr())
 
     dcu_session.run_with_iobinding(io_binding)
-    scores=np.array(io_binding.copy_outputs_to_cpu()[0])
+    result = io_binding.copy_outputs_to_cpu()[0]
+    scores = np.array(result)
     print("ort result.shape:",scores.shape)
 
     return scores