Implement Attention Augmented Conv2d for improving object detection

[dhananjaisharma10]

🚀 Feature

I would like to add support for the Attention Augmented Conv2d layer from the paper Attention Augmented Convolutional Networks.

Motivation & Examples

This feature is useful for the following key reasons:

• Attention mechanism from NLP is increasingly being proved to be beneficial for computer vision problems as it helps get a global context in an image.
• The paper proves that the relative, multi-head attention is better than 4 other attention mechanisms including Squeeze and Excitation attention.
• It also shows that this layer yields a 1.4% mAP improvement over a strong RetinaNet baseline; this is the main reason for adding it inside detectron2.

The feature would simply be a class inheriting torch.nn.Module.

I look forward to hearing from the team. Thank you!

[TylerMcCarthyGitHub]

Any idea wether this was ever implemented?