Closed TsingWei closed 1 year ago
We did not optimize the inference speed, especially the KDA attention that needs to be implemented with CUDA implementation (currently in PyTorch) for fast inference speed.
Is KDA a plug-and-play alternative to Deformable Attention?
Yes. You can safely use the original deformable attention, and the performance will not be influenced by a large margin.
Just wonder why there is no FPS compared with other methods in the paper.