ONNX Conversion with MNNConvert segmentation fault

[seanghay]

$ build git:(3.4.1) ./MNNConvert -f ONNX --modelFile /private/tmp/docaligner/out.onnx --MNNModel card.mnn --bizCode MNN
The device supports: i8sdot:1, fp16:1, i8mm: 1, sve2: 0, sme2: 0
Start to Convert Other Model Format To MNN Model..., target version: 3.4
[23:34:25] /tmp/docaligner/MNN/tools/converter/source/onnx/onnxConverter.cpp:46: ONNX Model ir version: 8
[23:34:25] /tmp/docaligner/MNN/tools/converter/source/onnx/onnxConverter.cpp:47: ONNX Model opset version: 16
[23:34:25] /tmp/docaligner/MNN/tools/converter/source/onnx/onnxConverter.cpp:149: Check it out ==> /neck/block.0/upsamples.3/Resize_output_0 has empty input, the index is 1
[23:34:25] /tmp/docaligner/MNN/tools/converter/source/onnx/onnxConverter.cpp:149: Check it out ==> /neck/block.0/upsamples.2/Resize_output_0 has empty input, the index is 1
[23:34:25] /tmp/docaligner/MNN/tools/converter/source/onnx/onnxConverter.cpp:149: Check it out ==> /neck/block.0/upsamples.1/Resize_output_0 has empty input, the index is 1
[23:34:25] /tmp/docaligner/MNN/tools/converter/source/onnx/onnxConverter.cpp:149: Check it out ==> /neck/block.0/upsamples.0/Resize_output_0 has empty input, the index is 1
[23:34:25] /tmp/docaligner/MNN/tools/converter/source/onnx/onnxConverter.cpp:149: Check it out ==> /neck/block.1/upsamples.3/Resize_output_0 has empty input, the index is 1
[23:34:25] /tmp/docaligner/MNN/tools/converter/source/onnx/onnxConverter.cpp:149: Check it out ==> /neck/block.1/upsamples.2/Resize_output_0 has empty input, the index is 1
[23:34:25] /tmp/docaligner/MNN/tools/converter/source/onnx/onnxConverter.cpp:149: Check it out ==> /neck/block.1/upsamples.1/Resize_output_0 has empty input, the index is 1
[23:34:25] /tmp/docaligner/MNN/tools/converter/source/onnx/onnxConverter.cpp:149: Check it out ==> /neck/block.1/upsamples.0/Resize_output_0 has empty input, the index is 1
[23:34:25] /tmp/docaligner/MNN/tools/converter/source/onnx/onnxConverter.cpp:149: Check it out ==> /neck/block.2/upsamples.3/Resize_output_0 has empty input, the index is 1
[23:34:25] /tmp/docaligner/MNN/tools/converter/source/onnx/onnxConverter.cpp:149: Check it out ==> /neck/block.2/upsamples.2/Resize_output_0 has empty input, the index is 1
[23:34:25] /tmp/docaligner/MNN/tools/converter/source/onnx/onnxConverter.cpp:149: Check it out ==> /neck/block.2/upsamples.1/Resize_output_0 has empty input, the index is 1
[23:34:25] /tmp/docaligner/MNN/tools/converter/source/onnx/onnxConverter.cpp:149: Check it out ==> /neck/block.2/upsamples.0/Resize_output_0 has empty input, the index is 1
Start to Optimize the MNN Net...
Reshape error: 0 -> 1
[1] 48897 segmentation fault ./MNNConvert -f ONNX --modelFile /private/tmp/docaligner/out.onnx --MNNModel

Model: https://drive.google.com/file/d/1jmr50_AysEws3dO3AHUxHwj6KMQug-6s/view?usp=sharing

[LudovicoYIN]

hi, @seanghay

Root cause:

• ONNX Resize optional empty inputs like [X, "", scales] were not handled correctly during import.
• Einsum("i,i...->...") was not lowered robustly in converter optimization.

I also checked the converted model against ONNX Runtime, and the result is now correct within floating-point tolerance:

• max_abs = 1.38298e-07
• mean_abs = 8.33486e-09
• rmse = 1.75169e-08
• np.allclose(rtol=1e-4, atol=1e-4) = True

[seanghay]

Thank you so much @LudovicoYIN !