Awesome. Been meaning to play around with Turi Create. Seems pretty easy to get started with. One of my colleagues wrote a post on building the hotdog/not hotdog app with it in an afternoon. https://heartbeat.fritz.ai/building-not-hotdog-with-turi-cre...
Yeah that is interesting and something that I wondered about while I wrote this post. There's only so many exercises you can model with a watch on your left wrist (at least with the accelerometer and gyroscope data). Another example that's been on the back of my mind was if I tried to go a step beyond classification to count reps for an exercise like single arm raises, would I be able to tell when I do a rep with my right arm? Definitely something I want to go a little deeper on in subsequent posts to find the limit of what we can and cannot classify. Like you said, I reckon static exercises would be pretty difficult and there would be a pretty high miss on those unless we had access to additional info.
I didn't get a chance to profile the app for battery consumption since my recordings were only 5-10s each but I'd assume that while the watch is recording the battery life would be significantly less if you had it on for an extended period of time.
I think part of it was that until the series 3, you still needed the phone nearby for wifi connectivity so developers were less interested in building a watch-specific app. Most of the ones out there were just widgets or notification apps that played nicely with the phone's.
As each series improves and Apple opens up new libraries so your watch becomes more and more independent of the phone, I believe we'll get more watch developers to up their game.
You don’t actually need the phone nearby with the previous watches either, they will connect directly to WiFi if they need to - they just prefer to use the phone for battery saving.
I was kinda surprised the first time I got an iMessage on my series 2 when my phone was in airplane mode!
Yes I would love to get the CoreML model onto the watch itself. I think the interesting part will be tweaking the model and exploring how we can build it with space, battery consumption, and resource utilization in mind. The nice thing is that we can possibly deploy 2 models (one to the phone and one on the watch): the one on the phone runs when the watch is nearby and is more robust. The other runs directly on the watch itself with some tradeoffs). We might discover that we need better hardware on the watch but that would be interesting to explore too.
