Performance of this model in real time
bogatovam opened this issue · 2 comments
Hello!
I want to run model predictions on webcam data, and I am faced with the problem of poor performance: predictions for one frame take about two seconds. It's horrible! Perhaps this is a problem with the Tensorflow, or my hardware...
But I want to know if you managed to launch this model in real time? If you succeed, please tell me the FPS (or other metric) and what hardware(CPU, GPU, they model) you ran it on.
I really hope for an answer. Thanks in advance!
Hi @bogatovam
The performance you get maybe slightly lower than what I got, but it's of the same order of magnitude, about 1 frame per second.
The neural network involved is quite large, and when running only on CPU, I doubt it could run any faster. It should run faster on a proper GPU (should be a decent one).
I don't have a GPU which supports TF, so I never tried it.
but you are welcome to use it.
the only change I can think of is importing the correct TF gpu version (tensorflow-gpu) instead of the regular one.
In my case, the performance issue was resolved by upgrading Tensorflow to 2.2.0.
Now predict takes 0.1 sec on Nvidia GeForce GTX 1650