Deformable Convolutional Op Support

[rootkitchao]

System information

• TensorFlow version (you are using):2.0 alpha
• TensorFlow Addons version:0,21
• Is it in the tf.contrib (if so, where):no
• Are you willing to contribute it (yes/no):yes,if i know how to do that
• Are you willing to maintain it going forward? (yes/no):yes

Describe the feature and the current behavior/state.
Deformable Convolutional Network(https://arxiv.org/pdf/1703.06211.pdf,https://arxiv.org/pdf/1703.06211.pdf) has achieved good results in the du detection task.But the current version of Tensorflow does not seem to implement deformable convolution operations.The efficiency of some third-party implementations of deformable convolution operations is lower than that of the mxnet version(https://github.com/msracver/Deformable-ConvNets).I hope that Tensorflow can add native deformable convolution operations in future versions.
Thanks.
Will this change the current api? How?
Yes,Add a new op
Who will benefit with this feature?
Any one want to use tensorflow for object detection.
Any Other info.

[Squadrick]

Here's a python only implementation of v1 for Keras which can be ported to TF-Keras. This repository is the official implementation with CUDA code; it's written using MXNet but it can be ported over to TensorFlow.

Here are the papers:

1. Deformable Convolutional Networks
2. Deformable ConvNets v2

I can start working on this if needed. I'll get the python implementation for v2 up and running first, followed by the proper CUDA/C++ implementation.

[rootkitchao]

Here's a python only implementation of v1 for Keras which can be ported to TF-Keras. This repository is the official implementation with CUDA code; it's written using MXNet but it can be ported over to TensorFlow.
Here are the papers:

Deformable Convolutional Networks
Deformable ConvNets v2

I can start working on this if needed. I'll get the python implementation for v2 up and running first, followed by the proper CUDA/C++ implementation.

Thanks for reply.And here is a Mask RCNN implemented by Tensorflow that may help test.

[facaiy]

@rootkitchao Welcome, William. I think you can go first because you seems willing to contribute the codes, right?

Are you willing to contribute it (yes/no):yes,if i know how to do that

Sorry for the delay, and thank @Squadrick for the reference paper and implementation :-)

[Squadrick]

@rootkitchao Any update? I can work on this if you aren't.

[Squadrick]

Using native CUDA for deform conv is dependant on #118 .

[jlhervy]

I tried to look into it, but it is too hard for me to do as I don't have experience in CUDA.

I wonder how hard would it be to port this implementation (tf1.4) to tf2.0 :
https://github.com/Zardinality/TF-deformable-conv

This is the first version of the paper, but ideally it would be nice to have deformable-conv V2 :
https://arxiv.org/abs/1811.11168
For which there is an implementation for pytorch, as cited by @Squadrick

Thank you

[vlordier]

@Squadrick you mentioned you'd implement DCN v2 in TF 2 : is this something you're having progress with ?

[Squadrick]

@vlordier Nope, haven't found the time to get around to this. Feel free to open a PR if you decide to work on this.