[franciscop]

Is this available for https://coral.ai/ somehow (USB accelerator)? Would it be difficult to convert it? I've played with the USB accelerator and it's cool, but would love to use some of these better algorithms since I found the default available ones were lacking.

[joshvm]

You can use yolov5 - here's a repo I made for an Ultralytics competition: https://github.com/jveitchmichaelis/edgetpu-yolo

Note: typical constraint is RAM and changes to the EdgeTPU compiler which now fails to convert larger models. Previously (version 15?) it would delegate layers to the CPU, but now it just doesn't work at all for large input sizes.

Also while it works, I think it's unlikely to be much better than a well trained mobilenet SSD. The advantage is you can train in pytorch and go from there, training quantised/edge models in Tensorflow is tricky.

[codeinassembly]

It is available via ONNX, which can convert to tensorflow weights. From there, it's possible to perform post-training quantization and it should finally be available for use with Coral. However, there's a good chance operations aren't yet supported by their chip, and accuracy will surely take a hit.

[eis]

I see you are working at "dagshub". Maybe you can let the people know that it's not a good show to create fake accounts here to push the story and leave useless praising comments.