target

Various runtimes might have bugs or don't fully support the onnx spec.

For that reason tensorflow-onnx has a --target option which triggers workarounds for issues we are aware of.

You find supported targets by calling python -m tf2onnx.convert without arguments. The usage will contain the possible targets, like:

[--target {rs4,rs5,rs6,caffe2}]

You can combine multiple targets as comma separated list.

Currently we support:

--target caffe2 - workarounds for missing ops and broadcast semantic. This was for older versions of caffe2 before it was merged into pytorch.

--target rs4 - rs4 targets the pre release for winml on windows, workarounds for missing ops and broadcast semantic

--target rs5 - rs5 targets winml in the october-2018 windows update, workarounds for missing type support

--target rs6 - rs6 targets winml in next windows update, workarounds for missing type support