prediction is wrong with multiple input with mini deployment=13

[xu-wang11]

🐞Describe the bug

A clear and brief description of what the bug is.

I have posted all the code and test data to the repo:

https://github.com/xu-wang11/SiamRPN-coreml.git

• input:
  • image: shape(1, 3, 271, 271)
  • kernel_reg: shape(20, 512, 4, 4)
  • kernel_cls: shape(10, 512, 4, 4)
• output:
  • out_reg: shape(1, 20, 19, 19)
  • out_cls: shape(1, 10, 19, 19)
    My pytorch model first is converted to onnx model, and convert the model to mlmodel

 mlmodel = convert(model=oonx_file, minimum_ios_deployment_target='13')
 mlmodel.save(coreml_file)

first I run the onnx model with onnxruntime

ort_session = onnxruntime.InferenceSession("siamrpn_update.onnx")
ort_inputs = {'x': instance_image.astype(np.float32), 'kernel_reg': x1['kernel_reg'], 'kernel_cls': x1['kernel_cls']}
ort_outs = ort_session.run(None, ort_inputs)

and the range of out_reg is (-0.27, 0.47)and the range of out_cls is (-10.37, 10.38)
then i run the mlmodel with coremltools

 x2 = update_model.predict({'x': instance_image, 'kernel_reg': x1['kernel_reg'], 'kernel_cls': x1['kernel_cls']},useCPUOnly=True)

and the range of out_reg is (-0.18, 0.25) and the range of out_cls is (-0.56, 0.75)

System environment (please complete the following information):

• coremltools version 3.3
• onnx-coreml version 1.3
• OS MacOS
• macOS version Catalina 10.15.4
• How you install python anaconda
• python version 3.6
• any other relevant information: