Comment by franciscop
2 days ago
I have a couple of these, unfortunately I've been waiting for the ecosystem to get better and run newer/improve models to no avail. I attempted some YOLO ports (since Coral uses a specific architecture) and not sure if I'm just bad at this or it's actually hard, but beyond the basic examples with Google's own ecosystem I wasn't able to run anything else on these. I was hoping an upgrade from seeing this on HN, but it seems to be the same old one.
Well, a small shoutout for my project to facilitate YOLO on the EdgeTPU without any dependencies on Ultralytics. Some contributors helped a bit with more recent versions, but it seems like all you need is to twiddle a few of the input shapes. I say facilitate rather than port because someone else did the hard work actually implementing the converter to tflite.
https://github.com/jveitchmichaelis/edgetpu-yolo
You really do need the model to be defined in Tensorflow though. Channel ordering screws things up in PyTorch and it's not trivial to get round it, otherwise the compiler does all sorts of weird gymnastics to permute for you.
See, this looks great, so now I have to give this a try! Just a quick Q to avoid the pain of trying if it's known not to work, is there a major reason this might not work on a raspberry pi? Thanks for sharing
If you can get the libraries installed then the code should work. The only risk is that Coral is not particularly well maintained and you might need to downgrade to e.g. an older version of Debian. Feel free to post an issue if you run into trouble.
The only repeating use I've found for these so far is object detection in Frigate (NVR software).
Same. I have a few Corals I was trying to build some prototypes with. Gave up and found Hailo. Much more powerful. Compatible with far more models. The documentation isn’t great, but its far better than Coral.
which Hailo product are you using and for what?
Similar experience; broken software.