how accurately can we predict power use from a tflite graph? I know it's a bit of vague it-depends question & ignores important things like actual sensor draw, but wondering what options there are... cc @petewarden @dansitu
-
We've always had to run on actual devices for good predictions, especially for projects like MNASNet. Estimations from FLOP count were up to 5x off actual latency.
Makes sense that rule-based estimates would be way off. But I'd be very optimistic about the prospect of training a DL model to estimate the performance profile of a model given its graph, trained on an artificially generated dataset.
-
-
my sim to real challenges continue! it used to be robot grasping imgs that kept me up at night, now it's going to be microcontroller emulators vs real devices :( :( :(
-
In
@EdgeImpulse we give reasonably accurate latency estimates for 9 different devices using a lookup table! We classify models depending on overall architecture & have benchmarks for MACCs/ms for each class. We then just multiply out to get an estimate—it's usually quite close. - Show replies
New conversation -
Loading seems to be taking a while.
Twitter may be over capacity or experiencing a momentary hiccup. Try again or visit Twitter Status for more information.