When to expect DCN_v2 to come in tensorflow

[Vedanshu]

Hi all,
It’s been a long time since the release of dcn_v2 ([1811.11168] Deformable ConvNets v2: More Deformable, Better Results). Many pytorch models are already using it like (yolact++, sipmask, CARAFE, maYOLACT etc.) When can we expect dcnv2 to come in tensorflow?
There are some unofficial implementations (GitHub - smallsunsun1/custom_ops: Deformable Convolution Networks(deformable conv2D TF_implementation)) and unmerged pull requests (Add Modulated Deformable Convolution layer by Licht-T · Pull Request #2196 · tensorflow/addons · GitHub) also.

[Bhack]

We had a quite long DCN related thread at:

"Code must be uploaded along with Your Submission. Failure to provide code may result in Your Submission being disqualified."

So we are supposed to upload the code for a revolutionary, creative, high-impact app against this risk to reward profile? Tens of thousands of people will enter this contest given that it is Google and the large size of the prizes There are a less than miniscule number of slots with prizes that are of any monetary significance Our entry has to survive the very real hurdles of being seen by a judge who is tired, or without interest, or done in by the simple vagaries of luck where our entry coul…

We could collect this layer in:

GitHub - keras-team/keras-cv: Industry-strength Computer Vision workflows...

But it is a python only repo.

[Vedanshu]

But the speed of python won’t be same as that of c++.

[Bhack]

You could try to propose this in the mentioned Keras-cv repo.

I’ve never benchmarked a python implementation using only TF native ops e.g. like:

https://tensorlayer.readthedocs.io/en/latest/_modules/tensorlayer/layers/convolution/deformable_conv.html