Add support for Less in GNMT

[RanACohen]

When I use GNMT from https://drive.google.com/drive/folders/1cWYUg1V1FFgFo8dF-W0hR2tyezCzY_UV

I get:
runtimeError: [ONNXRuntimeError] : 9 : NOT_IMPLEMENTED : Could not find an implementation for the node Less(9)

[RanACohen]

I am using the latest master from today: commit f25847b

[RanACohen]

This is for float64 input, I will create a MR for this.
I implemented it locally.

[RanACohen]

@snnn How do I create a MR? I don't seem to have permission

[neginraoof]

How is your implementation different from this PR? #1604
Since I'm able to run the ONNX exported model with no issues.

[RanACohen]

this model have double (float 64)
just added Less with double support.

[neginraoof]

We found the missing input type at this line:
https://github.com/NVIDIA/DeepLearningExamples/blob/3d3ff3e168a391e55cb28c3d5a4df69c9d4f4100/PyTorch/Translation/GNMT/seq2seq/models/attention.py#L81
This refers to int64 type. Can you confirm why we need float64?

[snnn]

Hi @RanACohen,

1. Fork this repo into your account. see https://help.github.com/en/articles/working-with-forks

2. Create a PR from your repo
   https://help.github.com/en/articles/creating-a-pull-request-from-a-fork

[RanACohen]

@neginraoof In the GNMT I posted in the drive link above (made by NVidia I think) there is a casting to float64 (double) and then Less.

[snnn]

Where did you get the model? How did you convert it? Is it the same one in mlperf https://github.com/mlperf/inference ?

Welcome to submit your change. I can help review.

[RanACohen]

@snnn same source as in #1583

[RanACohen]

@snnn Thank you Sun, please review #1722

[neginraoof]

@RanACohen The onnx model you are trying to run in the Google Drive is created by me. This is the onnx model exported from a fork of GNMT model as a workaround. Now that 'int64' is supported for 'Less' op, you should be able to run the model with no issues. I exported the onnx model from GNMT master branch and uploaded it again (replaces the files). You can try running the new one. There's no need to use double for 'Less' op in this model. Sorry if the uploaded onnx model confused you.

[RanACohen]

@neginraoof Thanks a lot! Does not hurt to extend the runtime with Less(double) in any case 😃
I will use the updated one.