3 ms·
One can run astonishingly accurate near-SotA ML models on even the rad hardened space rated systems with scant resources. For example after 8-bit quantization,
by datameta 6y ago
One can run astonishingly accurate near-SotA ML models on even the rad hardened space rated systems with scant resources. For example after 8-bit quantization, pruning, and compression a wake-word NN can take just 14KB. An RPi tier SBC can do quite a bit of inferencing!
- d_silin 6y agoHey, do you have any sources for that to read?
- datameta 6y agoSure. I saw talks by NXP and Renesas at the ARM Dev Summit last week that demonstrated the use of the Glow ML compiler [0] and CMSIS-NN Library [1] to squeeze a Wake Word NN [2] from 20KB to 14KB (with much greater reduction possible on larger ML models). Both Tensorflow Lite and Pytorch is supported in this toolchain. [0] https://github.com/pytorch/glow https://github.com/pytorch/glow [1] https://arm-software.github.io/CMSIS_5/NN/html/index.html https://arm-software.github.io/CMSIS_5/NN/html/index.html [2] https://github.com/tensorflow/tensorflow/tree/master/tensorflow/lite/micro/examples/micro_speech https://github.com/tensorflow/tensorflow/tree/master/tensorf...