Better CMSIS-NN Support
Hi folks - time for a short update,
First, thanks to everyone who's backed our OpenMV Cam H7 Kickstarter! We've raised 70K for the OpenMV Cam H7 now! Awesome! Anyway, If you haven't backed us yet please do! We've still got a few days left on the Kickstarter. https://www.kickstarter.com/projects/1798207217/openmv-cam-h7-machine-vision-w-micropython
Next, I spent some time updating the CMSIS-NN examples on the OpenMV Cam Github. We now have a README that walks you through how to use the library with exact command line values to run:
With this new guide and a deep-learning machine you can now actually train networks. You can run all networks on the OpenMV Cam H7. For the OpenMV M7 only the smile and cifar10_fast networks are small enough to be runnable. In particular, networks need to be no more than 30KB or so. Anyway, if you want to create your own custom CNN now you can do so by following our README walk through on how we made our smile detection CNN. Once you've got a deep-learning rig and caffe installed then training a new network is very easy.
Finally, there was a bug in the CMSIS-NN code from ARM that has now been fixed which was previously causing issues with running your own CNN. It has now been fixed on the master of the OpenMV Cam GitHub.
Anyway, we're going to try to get an IDE release out with all these fixes along with new CNN examples now that we've documented how to do things.
Thanks for reading,