zxy844288792 commented on issue #4404: [AutoTVM] select model with the most tuned schedules URL: https://github.com/apache/incubator-tvm/pull/4404#issuecomment-558302963 > > A general concern from me is that whether the current strategy or the proposed strategy are pretty heursitic. I believe both of them benefit the performance of some workloads but hurt some others. > > To the question about whether to accept this strategy, would you share more performance comparisons of some workloads from commonly used models with different strategies? At least we can have some confidence that the proposed strategy is empirically better. > > I agree that it is better to have some comparisons of some commonly used models between these two strategies. I will post here later today Results are attached here: | | -- | -- | -- | current strategy | proposed strategy densenet201 | 42.08180189 | 17.25377083 mobilnet1.0 | 0.460467339 | 0.459010601 resnet50_v1 | 2.447092533 | 2.366178036 alexnet | 6.949570179 | 1.42213583 yolo3_darknet53_coco | 538.1925201 | 33.61185551 ssd_512_resnet50_v1_voc | 601.3593554 | 25.04912853 In general, current strategy is much worse than the proposed strategy for untuned models.
---------------------------------------------------------------- This is an automated message from the Apache Git Service. To respond to the message, please log on to GitHub and use the URL above to go to the specific comment. For queries about this service, please contact Infrastructure at: [email protected] With regards, Apache Git Services
