fix: pytorch 1.10 generates ONNX with ir_version==7, which should be 6

tools/pytorch2onnx_kneron.py

@@ -124,6 +124,10 @@ def pytorch2onnx(model,
     model.forward = origin_forward
     # NOTE: optimizing onnx for kneron inference
     m = onnx.load(output_file)
+    # NOTE: PyTorch 1.10.x exports onnx ir_version == 7 for opset 11,
+    #       but should be ir_version == 6
+    if opset_version == 11:
+        m.ir_version = 6
     m = torch_exported_onnx_flow(m, disable_fuse_bn=False)
     onnx.save(m, output_file)
 
@@ -143,7 +147,9 @@ def pytorch2onnx(model,
         ]
         net_feed_input = list(set(input_all) - set(input_initializer))
         assert (len(net_feed_input) == 1)
-        sess = rt.InferenceSession(output_file)
+        sess = rt.InferenceSession(
+            output_file, providers=['CPUExecutionProvider']
+        )
         onnx_result = sess.run(
             None, {net_feed_input[0]: img.detach().numpy()})[0]
         # show segmentation results