Hi team, I am using MTCNN in pytorch, and it looks like pure non-max suppression implementation in pytorch(cuda) is waaay slower than numpy implementation on cpu. I ended up running just the nms part on cpu to get decent frame rates. Could someone please correct any obvious mistakes, here is the co…

NMS is one of those operations where it is common to write custom kernels because, as you note, implementing this in PyTorch directly is not so fast. In the MaskRCNN-Benchmark-Implementation there is are CPU and GPU kernels for NMS, e.g.: facebookresearch/maskrcnn-benchmark/blob/master/ma…

NMS implementation slower in pytorch compared to numpy

vision

tom (Thomas V) July 25, 2019, 1:05pm 10

See here for continuation: