Insightface: retinaface inference time
All 4 comments
@yja1 Why did you close the issue? did you solve the problem?
Neltherion
on 16 May 2019
detector.detect contain imread time and resize time
so I m thinking maybe should calc time of self.model.forward(db, is_train=False)
but the self.model.forward time always 0.3 ms
yja1
on 17 May 2019
@yja1 That's it. The Model at least takes 0.3 ms to forward the data (which is pretty short), all the rest of processing which contains processing anchors and detecting landmarks takes more time than the model forward itself.
@nttstar Do you think there is a way to decreases the anchor and landmarks processing time (whatever that happens after model.forward)? I did some benchmarks and the CPU NMS is fast enough but the processing of proposals, scores and landmarks after it is a bit slow, especially if you want to process the input data in batches (and not one image per detect call).
Neltherion
on 17 May 2019
@Neltherion Have you found a solution to process in batch?
yxchng
on 25 Dec 2019
Related issues
Alloshee
路
3Comments
lzg188
路
5Comments
WIll-Xu35
路
3Comments
mdv3101
路
5Comments
nttstar
路
6Comments
Most helpful comment
@yja1 That's it. The Model at least takes 0.3 ms to forward the data (which is pretty short), all the rest of processing which contains processing anchors and detecting landmarks takes more time than the model forward itself.
@nttstar Do you think there is a way to decreases the anchor and landmarks processing time (whatever that happens after
model.forward)? I did some benchmarks and the CPU NMS is fast enough but the processing ofproposals,scoresandlandmarksafter it is a bit slow, especially if you want to process the input data in batches (and not one image perdetectcall).