Are there some contrib ops that are not yet implemented?

[anirudhsundar]

Hi all,

I'm new to onnxruntime and am just a user of onnxruntime, so not sure if I'm missing something obvious here. I was trying to use the contrib op com.microsoft.MulInteger. I created a com.microsoft.MulInteger node in an onnx graph by setting the op type to MulInteger and domain to com.microsoft. When I try to run this graph with sample inputs, I'm seeing the below error

Current thread 0x00007fe6ad7ab740 (most recent call first):
  File "/home/temp/anaconda3/lib/python3.7/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 321 in _create_inference_session
  File "/home/temp/anaconda3/lib/python3.7/site-packages/onnxruntime/capi/onnxruntime_inference_collection.py", line 283 in __init__
  File "/home/temp/anaconda3/lib/python3.7/site-packages/onnxruntime/backend/backend.py", line 107 in prepare
  File "test_ort.py", line 137 in get_onnxruntime_output
  File "test_ort.py", line 200 in test_mulinteger

Grepping through the source shows only a couple of files where the word MulInteger is used. One looks like the op definition schema in quantization_defs.cc and the other is the documentation itself in ContribOperators.md.

• Could someone explain how contrib operators are added?
• Is it that a new op definition is initially created and documented, and then later the implementation of the op is added?

Thanks in advance.

[yufenglee]

ReduceSumInteger and MulInteger were designed but not implemented. Initially we experiment those ops and want to integrate them to ONNX as standard ops. We now tends to use the Quantize/Dequantize (fakequant in other framework) format to represent quantized model. We will remove those ops schema.

[anirudhsundar]

Thanks for the reply @yufenglee