opset_version uncompatible -- fail to load ort file on Android,

[theta666]

I just converted the onnx file to ort file according to onnxruntime official webset, and when loading it(create session) in code with pre-build runtime on Android, AndroidStudio reports this error message: Failed to find kernel for Pow(7) (node Pow_874), I checked the onnxruntime-mobile type sheet support, of which the Pow opset is {12,13}, SO is there any way to re-create onnx file with newer opset_version?

[snakers4]

@theta666
Which model did you try?
Can you maybe try this - #37 (comment) ?

@adamnsandle
Which opset did you use during export, 11?

[snakers4]

Can you try with new ONNX model?

[theta666]

I tried the onnx model inside 3.1, the new onnx model fails to run on android due to similar reason:
detail:

# convert command
python3 -m onnxruntime.tools.convert_onnx_models_to_ort --optimization_level=all ./silero_vad.onnx

operation list during conversion:

ai.onnx;1;GlobalAveragePool,LayerNormalization
ai.onnx;7;LSTM
ai.onnx;11;Conv
ai.onnx;13;Add,Concat,Div,Equal,Gather,Identity,If,Log,MatMul,Mul,Neg,Pad,Pow,ReduceMean,Relu,Reshape,Shape,Sigmoid,Slice,Softmax,Sqrt,Squeeze,Transpose,Unsqueeze
com.microsoft;1;FusedConv,FusedMatMul

when loading the converted ort file on android, it shows this error:

ai.onnxruntime.OrtException: Error code - ORT_FAIL - message: inference_session.cc:1217 AssignNodesToEpsFromHashesImpl Failed to find kernel def hash (5860801277476873352) in kernel registries for LSTM(7) node with name 'LSTM_600'.

it seems that the "com.microsoft.onnxruntime:onnxruntime-mobile:1.9.0" does not support lstm operation in onnx model

[snakers4]

is it supported in some newer opsets maybe?

[theta666]

no, according to the official doc, https://onnxruntime.ai/docs/reference/operators/mobile_package_op_type_support_1.10.html
maybe there's a way of convert lstm to custom op, I'm not professional on this

[snakers4]

looks like they do not support LSTM in mobile yet
basically LSTM is just a combination of linear layers and gates, but I do not believe we would have time and resources to implement custom LSTM just for the sake of being exportable to ONNX mobile