ai.onnxruntime.OrtException: Unsupported type - FLOAT16

[JiangWork]

When upgrade onnxruntime version to 1.16.x, if the onnx output type is float16, it will throw following exception:

Caused by: ai.onnxruntime.OrtException: Unsupported type - FLOAT16
	at ai.onnxruntime.TensorInfo.makeCarrier(TensorInfo.java:288)
	at ai.onnxruntime.OnnxTensor.getValue(OnnxTensor.java:86)

Seems it misses FLOAT16 case here:
https://github.com/microsoft/onnxruntime/blob/main/java/src/main/java/ai/onnxruntime/TensorInfo.java#L311

Related PR: #16703

[Craigacp]

Don't use getValue for FP16, use getShortBuffer then either work with it as shorts or use the Fp16Conversions utilities to convert it to floats.

It's noted in the PR that FP16 is currently unsupported for the methods which use arrays. I guess it could return a short array for the time being but in the future when Java has value types it would hopefully return an fp16 type.

[Craigacp]

Ahh. So if you were already using this then you were happy with it returning float arrays for fp16 values? I'd forgot that that used to work and it got lost in the refactor. I'll put that back in with a test to catch it.

[JiangWork]

Thanks for replying. When i used old runtime version, it will return float for float16. I will try this new API.

[Craigacp]

I've put up a PR to fix the problem - #18937. Note, the behaviour isn't identical to before as the old fp16 conversion code was very bad at handling denormals, infinities and NaNs (in that it returned incorrect fp32 values in those cases). I've added tests which compare the output against the expected result so it should behave correctly now.