Is Deformable Convolution supported in TensorRT?

AI & Data Science Deep Learning (Training & Inference) TensorRT
1

Deformable Convlution: [1703.06211] Deformable Convolutional Networks

I don’t see it listed in Support Matrix(Support Matrix :: NVIDIA Deep Learning TensorRT Documentation), is it supported by TensorRT?

Thanks.

2

Unfortunately, Def Conv is not currently supported by TRT. We are always paying attention to community feedback. Thanks.

3

Would it be possible to write a custom plugin for deformable convolutions in TRT?

4

Dconv is always contained in recent solutions like detection and segmentation, and I convert this layer using custom layer. These solutions runs but the speedup ratio is not so good as usual convolution.The document says tensorrt can merge (conv+bn+relu) into one CBR layer and merge multiple 1x1 kernel CBR to one layer. I do think the speedup ratio can be increased by merging (dconv+bn+relu) and so on if you guys kindly update tensorrt because these part is not opened to public in github. Waiting for you update.