[Paper] Optimizing Deep Learning Models For Raspberry Pi. Custom CNN (on MNIST data) performance from 114ms to 3.75ms. ResNet50 (on "flowers" data): from 1.1s to 1.0s (lowest) or 1.6s (highest).
[Paper] Optimizing Deep Learning Models For Raspberry Pi. Custom CNN (on MNIST data) performance from 114ms to 3.75ms. ResNet50 (on "flowers" data): from 1.1s to 1.0s (lowest) or 1.6s (highest).
arxiv.org /abs/2304.13039
I'm a little unsure on if I interpreted the results correctly. It seems like some things that TF Lite natively supports (apparently, their custom CNN model trained on MNIST) get really fast, and other things are a little hit-or-miss.
0
comments