MxNet for Embedded device inference

Question

Is there a way to take MxNet model adn deploy it to embedded device directly? As "embedded", objective is to have super lightweight, optionally optimized for ARM/neon.

score 0 · Answer 1 · answered Mar 08 '18 at 18:42

0

Yes, here is the tutorial which explains step by step how to run MxNet inference ARM-device (Raspberry Pi in that case) - https://mxnet.incubator.apache.org/tutorials/embedded/wine_detector.html

answered Mar 08 '18 at 18:42

Sergei

1,617
15
31

Thanks. But I was really hoping for something embedded i.e. C++ code examples whcih are optimized for ARM/Neon platform. – krishnakamathk Mar 09 '18 at 22:19
I don't think there is any specific guidance on how to write high performance code for ARM/ Neon platform. But as for being lightweight, take a look into that project - https://github.com/apache/incubator-mxnet/tree/master/amalgamation It combines all prediction api into a single file, making it more lightweight – Sergei Mar 10 '18 at 01:00
But the concern is shared object file is huge! Last time I checked it was about 100 MB and not a good thing for Embedded devices. – krishnakamathk Mar 29 '18 at 17:47

MxNet for Embedded device inference

1 Answers1