Closed ShoufaChen closed 4 years ago
I think the inference time should find the final detections, so I think the nms should be included, note that for two stage framework, the anchor generation time should not be included, what I say may be not right, just a reference.
Checkout https://gluon-cv.mxnet.io/model_zoo/detection.html, it's based on V100 GPU, but you may get some ideas of the FPS
Hi, @zhreshold . Do you mean your training and inference are both based on V100 GPUs? I check out your training log on VOC and on coco, and find that you can get 12 sample/s
training speed for res101 on 8 GPUs, and also 12 sample/s
for res50 on 4 GPUs. But I can only get 2.6 samples/s
for res101 on 4 k40 and 3.6 samples/s
for res50 on 4 k40. I wonder if this is a device problem? Or anything goes run with my configurations?
@FishYuLi Correct, it's GPU's capability. K40 is pretty dwarfed comparing with Volta GPUs.
4- 8 GPU scaling seems pretty bad though? Maybe CPU bounded?
Извините что по русски, меня кто-нибудь понимает? Оцените мой сайт http://videosektor.net/
Well Hacking whatsapp should be done like instruction please
Hi, @zhreshold
Do you have a plan to provide the object detection speed at the inference phase, like
fps
. I am not sure how to determine the inference time, e.g. before nms or not?