Model works with target_ios='12' but not target_ios='13'

[SumatPurewal]

🐞Describe the bug

I am using inceptionv3 as a feature extractor. The inceptionv3 subnetwork works fine if I set target ios to 12 but when I set the target_ios to 13 then the model converts fine however when I attempt to run a prediction on it in the iPhone11 simulator or an iPhoneXS then the app crashes. Running predictions through coremltools also result in a segfault.

If I visualize the model in netron it looks fine however if I try to do the same using coremltool's visualize_spec() method then I get the following error:
ValueError: Input inception_features : Multi array must be of length 1 or 3

To Reproduce

link: https://github.com/SumatPurewal/CoreML3-Conversion-Bug
The repo readme describes how to reproduce the bug. Should be straightforward.

System environment:

• coremltools version (3.0):
• onnx-coreml version (1.0):
• OS (MacOS Catalina):
• How you install python (system):
• python version (2.7.15):

Additional context

I personally don't mind using coreml2 for the inception network however I need operations like dynamic shape inference and gemm for the rest of my network that are only available in coreml3. The link to the github repo shows a simple network that attempts to compress the inception features by taking the channel wise mean. This network also throws no errors or warnings when converting with target_ios=13 but returns nil in the app and presents the same "Multi array must be of length 1 or 3" when I try to visualize it using coremltools

[dingzeyuli]

I run into the same issue. Is there any fix to this?