SSD MobileNetV2 MnasFPN gets converted successfully but inference fails at Conv2D op

[codethief]

Describe the bug
SSD MobileNetV2 MnasFPN gets converted successfully by tf2onnx 1.8.2 buts fails to execute using ONNXRuntime 1.6.0.

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04): Ubuntu 20.04
• Tensorflow Version: 1.15.3
• Python version: 3.6.12

To Reproduce
I trained a model based upon ssd_mobilenet_v2_mnasfpn_coco from the TFv1 Model Zoo.

Afterwards I converted it using tf2onnx:

python3 -m tf2onnx.convert --saved-model path/to/saved/model --output model.onnx --opset 11

which seemed to work perfectly. However, running inference with this model then resulted in:

>           return self._sess.run(output_names, input_feed, run_options)
E           onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : Non-zero status code returned while running Conv node. Name:'WeightSharedConvolutionalBoxPredictor_1/PredictionTower/conv2d_0/separable_conv2d/depthwise:0_nchwc' Status Message: Input channels C is not equal to kernel channels * group. C: 48 kernel channels: 1 group: 1

/home/user/opt/miniconda3/envs/my-conda-env/lib/python3.6/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py:124: Fail

Expected behavior
tf2onnx should have complained when converting the model or output a working model.

Additional context
The issue looks similar to #1100 but I'm not entirely sure. #824 also seems related but the suggestions there didn't help me as I'm not using use_explicit_padding: true in my pipeline.config in the first place.