How to benchmark models with custom MTK ops

NeuroPilot NeuroPilot - Analytical AI
, ,
1

The tflite model generated using the mtk converter tool contain custom MTK ops like MTKEXT_CONV_2D which are not supported by the official tflite benchmarking tool (android_aarch64_benchmark_model). How to benchmark such models?

ERROR: Encountered unresolved custom op: MTKEXT_CONV_2D.
See instructions: https://www.tensorflow.org/lite/guide/ops_custom
ERROR: Node number 2 (MTKEXT_CONV_2D) failed to prepare.
ERROR: Failed to apply NNAPI delegate.
ERROR: Benchmarking failed.
3

Hi GeorgeC,

Thanks for reaching out!

When converting the model, we recommend setting the operator export specification to builtin_ignore_version. This prompts the conversion tool to prioritize standard TFLite built-in operators and avoids generating custom MTK operators.

mtk_pytorch_converter \
    --input_script_module_file="model.pt" \
    --tflite_op_export_spec="builtin_ignore_version" \
    # other existing arguments...

For more details on mtk_pytorch_converter parameters, please refer to this link.

Best,
Jun