2

I just picked up a Google Coral USB Accelerator to learn ML without spending a lot of money on a new PC/GPU (I still have a Macbook Air from 2014 and it's laughably slow)

I'd like to try to build on the work done on this video shot detector model and train something for a related use case: https://arxiv.org/pdf/1705.08214.pdf

That model is a lot different from any of the Coral CV examples that I've been playing around with so far. Those that I have run all take single frame as input, but this approach requires a group of frames to be passed to the model together.

I'm trying to figure out if this sort of thing is even supported on the EdgeTPU - I found this page of operations to compare with the table on pg 3 of the PDF:

https://coral.ai/docs/edgetpu/models-intro/#supported-operations

The "Conv2D" listed seems pretty explicit that it's 2D, and there is no equivalent 3D operation listed. So does that mean I'm out of luck here?

I wonder if anybody has any other ideas / prior art I should look into on this sort of video analysis that would take advantage of the Coral Edge TPU?

Thanks!

Carini
  • 21
  • 1

1 Answers1

0

Good question, thanks for sharing that paper also, very nice. The problem is that edgetpu is made for edge devices and have very limited memory. This make simple ops involving too many inputs much more complicated than a normal cpu model. I guess this could eventually be supported in the future with both SW+HW side, but for now the constraint is still on the supported ops list, unfortunately.

Nam Vu
  • 1,727
  • 1
  • 11
  • 24