Coral TPU

[Daniel15]

Do you have any models that can be used with the new CodeProject.AI Coral TPU module?

[kittenmcnuggets]

I think they are working it, or looking at least

[MikeLud]

I will try converting some of my models sometime this week. I do not have a Coral so I will need Chris to test them.

[jshridha]

Let me know if you need access to a tpu. I can make a machine with one available for you to test on.

[8666]

It would be nice to have the same IPcam models for Coral TPU too. So it will load balance the detection between yolo CPU and Coral while getting the same expected results

[kittenmcnuggets]

Are they maybe packed with the module now? Maybe not, but I've seen a few BI folks claiming they were using coral on windows and declaring the same custom models. Also could be not.realizing...

[pbohannon]

Hmm, does ONNX->PB->TFLITE not work for IPcam to Coral? I was able to convert the IPcam-combined.onnx to a .tflite file but haven't had time to install/test. It looks like 30m: -rw-r--r-- 1 xxxx xxxx 30127028 Oct 19 16:14 IPcam-combinedlite

[kars85]

I will try converting some of my models sometime this week. I do not have a Coral so I will need Chris to test them.

@MikeLud - I have a Coral USB I would send you if you're interested to help with your development.

[pbohannon]

I do as well and would contribute to either shipping it or giving remote access. Having your models ported would be -amazing- I have both the m.2 (although not in service because won’t fit my form factor) and USB TPU. Let us all know how to help and you can count on it!

On Tue, Nov 7, 2023 at 9:55 AM Karson @.***> wrote:

I will try converting some of my models sometime this week. I do not have a Coral so I will need Chris to test them.

I have a Coral USB I would send you if you're interested to help with your development.

— Reply to this email directly, view it on GitHub https://github.com/MikeLud/CodeProject.AI-Custom-IPcam-Models/issues/5#issuecomment-1799346142, or unsubscribe https://github.com/notifications/unsubscribe-auth/AKNN2S4OCAJZ75XPBGNBQSDYDJYZRAVCNFSM6AAAAAA2Q3N5YWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMYTOOJZGM2DMMJUGI . You are receiving this because you commented.Message ID: @.*** com>

[MikeLud]

@kars85 & @pbohannon Thanks for the offer, I was looking at buying a Coral so I just did.

[smhc]

Thanks for looking into this @MikeLud - the currently available models are abysmal and require using the larger, slower models for half decent detection. They also have extremely poor 'dark' detection capability.

Even if there was something available, AFAIK coral/codeproject.ai doesn't support multiple custom models so you can't alternate day + night detection. Hopefully the 'ipcam-general' that includes dark model images will help.

I did try experiment with: https://github.com/jveitchmichaelis/edgetpu-yolo - but it didn't go very well. Currently YOLO on CPU is a significantly better option than using coral - so will be watching this development closely to see if it can be salvaged, cheers.

[MikeLud]

I started digging in on making my custom models work with Coral. It look like I will need to make a new module ObjectDetection (YOLOv5 Coral)

[kars85]

I started digging in on making my custom models work with Coral. It look like I will need to make a new module ObjectDetection (YOLOv5 Coral)

Thanks @MikeLud! How is it going? Will that model be bundled in a future update?

[raptordemon]

i see we have the models, has the YOLOv5 Coral been made too? if so where do we find it?

[MikeLud]

i see we have the models, has the YOLOv5 Coral been made too? if so where do we find it?

I am still working on the module that works with YOLOv5 models

[raptordemon]

that is awesome, thanks for working on all of this

[minigranis]

Any hints as to when this will be available ? I have seen the models here on Github but I dont understand if I can add them to CP manually (running docker version of CP)

[xm4rcell0x]

Any hints as to when this will be available ? I have seen the models here on Github but I dont understand if I can add them to CP manually (running docker version of CP)

Same here