Did you copied the code or you implemented the code?

Did you copied the code or you implemented the code?

Will add copyright.

Please note , this op doesn't support mini-batch, so it will cause a lot of problems. I'd prefer to avoid using it.

synced offline, h and w are fixed across all images after preprocessing step, so it can do mini-batch.

The code looks a lot similar to roialign. Can we reuse that instead of duplicating?

The code looks a lot similar to roialign. Can we reuse that instead of duplicating?

The algorithm is totally different. The similar code happened in input validation, so I change this part accordingly.

Since you are adding a new contrib op you should also update this file: https://github.com/microsoft/onnxruntime/blob/master/docs/ContribOperators.md

(Please note: this file is auto-generated so use this script (https://github.com/microsoft/onnxruntime/blob/master/tools/python/gen_doc.py) to generate it)

Since you are adding a new contrib op you should also update this file: https://github.com/microsoft/onnxruntime/blob/master/docs/ContribOperators.md
(Please note: this file is auto-generated so use this script (https://github.com/microsoft/onnxruntime/blob/master/tools/python/gen_doc.py) to generate it)

Generated, thanks