[one-cmds] policy with one-import-onnx extension

[seanshpark]

When we first introduced one-import-onnx-ext,

• we executed one-import-onnx-ext when installed, without option
• then introduced --disable_ext option, not to run extension

one-import-onnx-ext was introduced cause

• onnx-tensorflow isn't maintained anymore and there is no upgrade
• some of the models has problems with conversion, and the number will grow in the future

But for the user experience and to reduce complexity,

• had a offline talk with @lemmaa
• 1/ run the default conversion path, onnx-tensorflow -> tf2tfliteV2 -> tflite2circle
• 2/ if this path fails, run the extension
  • but if --disable_ext is given, don't run the extension
  • characteristics of this option will change
  • internally, as our extension is not officially released yet, I assume we can change this
• 3/ introduce --force_ext (or some better name) to skip default path and run extension if available
  • or --use-ext, --skip-default
  • introducing this is optional

[seanshpark]

what to test

• [ ] onne-tf is ok
• [ ] onnx-tf fails but onnx-import-onnx-ext is ok
• [ ] onnx-tf is ok but onnx-import-onnx-ext fails
• [ ] onnx-tf and onnx-import-onnx-ext both fails

[shs-park]

Just for curious,

When both succeed, is there any cases where onnx-import-onnx-ext will have a slightly better result than onnx-tf? If so, I think the result of onnx-tf will be used by priority though. Will it be okay?

[seanshpark]

test model where onnx-tf fails

• onnx_tf_error_model.zip
• this is portion of model.zip from https://github.com/onnx/onnx-tensorflow/issues 1078

[seanshpark]

When both succeed, is there any cases where onnx-import-onnx-ext will have a slightly better result than onnx-tf?

Mostly the same but yes, there is one(or some?) internal model from our customer.

If so, I think the result of onnx-tf will be used by priority though. Will it be okay?

for this case, option --force_ext can be used to skip onnx-tf