After reading about tinyML, I thought it might be interesting to try this for myself. The easiest way seemed to be via the SparkFun Edge. Turned out to be fairly fiddly, especially getting the Mac serial port to work correctly at 912600 bps so that new code could be flashed to the board. After researching this in the forum, installing a new driver fixed that problem.
While SparkFun do have a guide to building for the Edge, I found it much easier just to use the instructions here as these commands also downloaded and installed the tools required. It just takes a few copies and pastes of the commands to get the code built.
In the end, I was able to get the person detection example working. It seemed to work ok but each inference takes about 8 seconds. My overall reaction was that I am not sure it was worth the effort – there’s little doubt that, having got this demo to work, the board will be condemned to live in a box with lots of other no-longer used dev boards. I think that there probably are use cases for really low power inference in some consumer devices but you really have to want to do this to bother with the process! I am going to stick to GPUs, Neural Compute Sticks and Edge TPUs for the moment as they are easier to use and work very well for my applications. Still, if someone comes up with an interesting and easy to use model implemented for the SparkFun Edge, I might dig it out again to try it out.