Wrong results on ANE when set minimum_ios_deployment_target=ios12

[yizhaoyanbo]

🐞Describe the bug

I have trained a fcn-mobilenetv2 model using pytorch and have converted the model from onnx to mlmodel successfully both with minimum_ios_deployment_target=ios12 and minimum_ios_deployment_target=ios13 .

When ios12 target is set, I get correct prediction result using cpu-only and cpu+gpu, but get wrong result when using apple's ANE.

When ios13 target is set, I get correct results on the iphonexs with ios 13.3, cpu-only, cpu+gpu and ANE.

In the test, all the inputs are same, such as an image with fixed value.

So It seems that the convert tool may have some bugs to convert the model with ios12 target.

To Reproduce

Here is the onnx and mlmodel. I put it on the google drive:
https://drive.google.com/drive/folders/165ykvpEvM8SWe3CniuCv2ZWb1vv2_rO7?usp=sharing

System environment (please complete the following information):

• coremltools version (e.g., 3.0b5): 3.3
• onnx-coreml version (e.g. 1.0b2): 1.3
• OS (e.g., MacOS, Linux): Macos
• macOS version (if applicable): 10.15.3
• How you install python (anaconda, virtualenv, system): anaconda
• python version (e.g. 3.7): 3.7
• any other relevant information: