ivy.deform_conv2d

[themaigod]

Why should this be implemented?
As discussion in #20138, deformable convolution is a popular and common operation. Here is the reason why I think it should be added as an experimental api:

• 3+ of the native frameworks have this function (Pytorch (in torchvision), mindspore, paddlepaddle)
• it's an NN-related function and should be as efficient as possible (official implementations in native frameworks have advantage in speed)
• it's needed for a complex/long frontend function implementation

Links to native framework implementations

• PyTorch (in torchvision)
• Paddlepaddle:paddle.vision.ops.deform_conv2d and paddle.static.nn.deform_conv2d. Notice that deformable_conv is old API and connot be used.
• Mindspore
• MXNet v2 MXNet v1
• Though Tensorflow does not have official implementation, you can check multiple related PRs for that. It seems that the implementations were not merged due to some outer issues, such as feature freeze. PR1 PR2 PR Sggestion

Link to Third Party Implementations only by Python which are complex and low efficiency
-PyTorch
-TensorFlow

Link to the paper which released this op

• Deformable Convolution
• Deformable Convolution v2

[themaigod]

• [deformable_conv2d ] deformable_conv2d #20945

[themaigod]

• Create the API in ivy functional api
• Create the backend implementation in PyTorch for it
• Create the backend implementation in Paddlepaddle for it
• Create the backend implementation in MXNet for it
• Write tests for it

[AnnaTz]

Thanks for the suggestion! I have added this to our ToDo list and linked your PR to the new issue!

[themaigod]

Thanks for the suggestion! I have added this to our ToDo list and linked your PR to the new issue!

Thanks.