CodeProject.AI Version 2.0
- Thread starter MikeLud1
- Start date
Is there a good current document on how to setup and tune CPAI? I am looking at this from moving to substreams, recommended size/quality of those streams, which models to use, etc. I see so much conflicting information between this forum, BI's forum, reddit and CPAI website. It would be great to build a site somewhere with "best practices" to make this all work easier so we can all spend time on other projects. I am game to document this but I dont have 200 hours to try every combination of settings and dig through all the various sources of information as some of them seem very dated at this point as well. Thanks!
Iv been using codeproject ai with a google coral. I tried the 3 models and just ended up sticking with the small model size mobilenetssd one . Its the default one enabled for the coral plugin. Not really much tweaking you can do with the program really. Its been working great for a few months.
Are you running bare metal blueiris? I recommend getting a coral device as it takes a lot of stress of your cpu for processing ai. I have a few of them and my favorite one is the pci express route. Get a dual edge tpu and there is a guy that make a pci express adapter for it.
Dual Edge TPU Adapter - PCIe x1 Low Profile
Dual Edge TPU Adapter is designed for Coral m.2 Accelerator with Dual Edge TPU to be used on a system with PCIe x1 slot available.
www.makerfabs.com
I'm using a Minipc for my Unraid BI server. If I get a USB coral device how do I configure it to work?
Or instead of USB do they make a coral that fits in the internal wifi chip slot since I'm not using it.
Or instead of USB do they make a coral that fits in the internal wifi chip slot since I'm not using it.
Last edited:
They do make one that fits in that Wi-Fi chip slot. I recommend going that route. I have a USB one and I had issues with it. Some people don't but I personally did
This is the WiFi tpu I used in my lenovo thinkcenter. Works faster than the USB.
I've also just ordered a m2 WiFi format tpu and m2 b+m tpu format to have a selection.
I have an a+e format that I bought by error.
Need a pcie adapter for that which I haven't bought yet.
May see if I can sell that one. Bit of an odd format unfortunately
Also if you've got a sff/tff you don't want usb devices flopping about. Just makes it tidier
Last edited:
Than
Thanks, I have the SER5 mini
Thanks, I have the SER5 mini. So WiFi adapter probably only option for me. Does the Coral AI do anything for photo or video editing? Or just for codeproject?
True. I had trouble with the usb on my small lenovo as it would sometimes change usb device number so had to update the proxmox lxc config.
But now it's been on my unraid server solid for 1 year.
But yeah I've been testing the WiFi tpu and very impressed with it so ordered more. One for the unraid and another m2 for a 2nd pc.
I too was impressed with them. Ended up ordering a few of each they had. I currently run a few with the pci express adapter and another with a raspberry pi5 with the ai hat. The pi5 is flawless and rock solid.
Hat AI! for Raspberry Pi 5 - Coral TPU Hat Ai - Pineboards
Introducing the Hat! Ai! - your gateway to integrating the Google Coral Edge TPU into your Raspberry Pi 5 projects. Our Hat Ai! does not block GPIO access.
pineberrypi.com
ordered the wrong one, got the mini pc apart and the wifi slot says M.2 Wifi 2230
will this card work below? the only difference is it says M.2 2230 E if not they have another that is M.2 2230-A-E-S3
Hey guys,
I haven't really checked these forums in over a year or so. I've got a spare USB Coral adapter from another project and I see now that CPAI now supports Coral TPUs.
At the moment, my BI server is using a GPU which is connected to another PC running CPAI in a docker container.
So I've installed CPAI on the BI machine and also installed the Coral USB drivers.
When I go into the CPAI Explorer and run some benchmarks, I am extremely underwhelmed. I'm getting 0.5 operations per second when using the same test image as the remote CPAI server with a 1050ti getting 8.8 operations per second.
I was prepared to take some performance hit, knowing that this is just a USB TPU, but I wasn't expecting it to be THIS bad.
Is this level of performance expected? Or am I forgetting to do something?
I haven't really checked these forums in over a year or so. I've got a spare USB Coral adapter from another project and I see now that CPAI now supports Coral TPUs.
At the moment, my BI server is using a GPU which is connected to another PC running CPAI in a docker container.
So I've installed CPAI on the BI machine and also installed the Coral USB drivers.
When I go into the CPAI Explorer and run some benchmarks, I am extremely underwhelmed. I'm getting 0.5 operations per second when using the same test image as the remote CPAI server with a 1050ti getting 8.8 operations per second.
I was prepared to take some performance hit, knowing that this is just a USB TPU, but I wasn't expecting it to be THIS bad.
Is this level of performance expected? Or am I forgetting to do something?
Lot's of posts here of people experiencing poor performance from Coral even in a M2 slot, so a USB will be even worse. Coral just isn't ready for prime-time yet.
Ah, I see. I'll check back in another year or so.
I guess I'll rely on my remote GPU server solution until then.
Send me the test image and ill use it on my and tell you the results. I got the the pci cards one running
Personally I love the coral devices and they have been working very well for me. Not the usb one though. Dont trust that thing.
I downloaded an 8k image and processed it
| # | Label | Confidence |
|---|---|---|
| 0 | bird | 64% |
| 1 | bird | 61% |
| 2 | bird | 57% |
| 3 | bird | 55% |
| 4 | person | 54% |
| 5 | bird | 54% |
| 6 | person | 54% |
| 7 | bird | 54% |
| 8 | person | 52% |
| 9 | bird | 51% |
| 10 | bird | 51% |
| 11 | bird | 51% |
| 12 | person | 50% |
| Processed by | ObjectDetectionCoral |
| Processed on | localhost |
| Analysis round trip | 745 ms |
| Processing | 189 ms |
| Inference | 9 ms |
| Timestamp (UTC) | Tue, 02 Apr 2024 03:14:39 GMT |
Well, the inference time sounds right. Not sure what it’s doing the rest of the time. Resizing a 4K image to the input tensor size takes roughly 10 ms. (Should be even faster in the latest version.)