CodeProject.AI Version 2.0

[mailseth]

I saw that M.2 adapter, but decided that I’d fill my PCIe slots first if I wanted to max my machine out. It’s just a HP EliteDesk G4 800 I got cheap on eBay.

Close. I put the pads under the Dual card to serve as extra heat dissipation. Then on top of the Dual card I put copper heat spreaders, which also add some space. And then I put the heat sink on top of that. Similar to the pdf at the end of my comment.

Yep to what you said, except four holes. I had not heat problems with the single TPU boards.

 

[AlwaysSomething]

Yeah I read about the 2nd pad underneath but they said the top is more important. But hey if you have it, mind as well use it. Also thanks for the pics. That will help others as well if they decide to get one.

I also bit the bullet and ordered everything late last night. Should have everything in a few weeks.

Ironically, I also bought a cheap Elitedesk G4 not too long ago and plan on migrating everything to that for the extra hard drive and slightly newer CPU. Debating getting a 2nd BI license so I can keep a "stable" version and "current" version. Or in the software development world it would be "Development" and "Production". LOL. Would also be some redundancy in case of Murphy's Law which always hits me.

Thanks again.

 

[mailseth]

Yeah, pulling heat off the bottom isn't nearly as important as a heatsink on the top. I figured with the big heatsink, it wouldn't hurt to add some structure on the bottom, too. If anything, I wish I used a slightly thicker thermal pad for better support.

Also, of note, you may want to invest in thermal grease or double-sided sticky pad for the heat spreader. I ended up using the latter.

Also, FWIW, the creator of the PCIe adaptor cards went for a full fan setup with the heatsink, but I haven't found that necessary.

PCIe x1 adapter cooling options · Issue #7 · magic-blue-smoke/Dual-Edge-TPU-Adapter

Option 1 Dual Edge TPU cards are distributed by Mouser, so I went through their catalog looking for heat sinks. This one fits, with few thoughts in mind: Distance between mounting holes on PCB is 4...

github.com

 

[Corvus85]

Hey all,

I've been out of the loop over the past year on all the updates and new features of BI/CPAI.

I've recently added a GPU to my system and updated BI/CPAI to the latest versions, and I'm noticing that in the AI analysis that CPAI is continuing to analyze every snapshot asked of it despite getting a positive result on the first snapshot.

This seems a little inefficient to me. Is there a better way to do this?

It also seems to look for static object too for some reason. I've never enabled this option, but it seems to have enabled itself. Is there a good reason as to why anyone would want this turned on?

Here's how I've currently got it set up on all my cams.



And here's what the AI analysis says:

 

[silencery]

Does anyone know how a Coral Dual Edge might perform against a 3 year old i7 (10700K) for CPAI? It seems things are moving pretty quickly in terms of adding Coral support, despite the old hardware. Thinking of picking one up to test out.

Thanks!

 

[mailseth]

An answer to that question is going entirely depend on your setup. How many models, cameras, & FPS for starters. What sort of accuracy are you looking for and what hardware are you plugging it into and how cost sensitive are you. Personally, I’d say go for it and see if it works for you.

 

[silencery]

good point. I'll post back with results after it comes in after a few weeks

 

[AlwaysSomething]

My current CPU is even older, i5-8500, and I needed low power so I went with a TPU vs GPU. (just ordered an additional dual TPU).

As Seth mentioned a few posts up, depending on the model size you want to use you may need more than 1 TPU. For me, the Small found almost everything I needed but the Medium was better. It also tagged/labeled more accurately. For example, instead of just classifying a truck it would classify it as a bus (hope that makes sense). A better example would be cat vs a dog. I also tested the Large model but found I didn't need it.

FYI - A couple of my cameras are also monitoring a long distance (I think up to 100-120 feet away) so the objects may be small in the image which makes it even harder. If monitoring closer/larger objects then the Small and maybe even the Tiny model will work for you. I've seen people swear by Frigate and the Tiny model but then their examples are photos of there 20 x 20 deck or people standing right in front of their doorbell camera. If that is the case then Tiny is probably all you need.

I think it was also Seth that also mentioned somewhere that the original models were not targeted towards this audience (BI) and things like books, desks, couch, chair, etc. were being searched for. I remember seeing these when first testing my indoor cameras and was like HUH??? LOL

Hopefully as things progress we'll get better models geared towards this audience (BI), as well as the TPU, and the models will be more efficient.

If only there was a way to just remove the data we don't want from a model. I know we can build our own models now with the latest version of CPAI but I just haven't had time to look into. Maybe the answer is in there. Maybe the data the model is built from is readily available (like a library) and we just need to copy what we want, leave what we don't want out, and then build the models ourselves. IDK

That being said, thank you to MikeLud, Seth, and all the others (who I know there are more but drawing blanks on the names right now) that are taking the time to build these custom models for us.

 

[mailseth]

Yeah, the default YOLO models are looking for irrelevant things ranging from a bench to broccoli. My ideal model would be something based on YOLOv8 medium that does everything (all IPcam labels + plate labels + fire detection, etc). Running a medium model is roughly the compute cost of running a small model twice. As far as I can tell, there isn’t much compute cost in additional labels. Extra labels can be confusing, though.

MikeLud has been busy lately, but I’m looking forward to future model releases. I’ve considered getting some better hardware for my own training /experimentation, but that would also require time I don’t have. This TPU optimization is entirely enough of a distraction.

 

[mike2246]

Why is it my wyze cam alerts are WAY more reliable then my BI/Coral alerts? Same detection zone, same camera. Wyze alerts are always instant and spot on whereas BI is hit or miss with alerts. At this point I almost want to use BI as an archive to have justvin case.
In BI I have person confidence set to 55%

 

[wittaj]

Why is it my wyze cam alerts are WAY more reliable then my BI/Coral alerts? Same detection zone, same camera. Wyze alerts are always instant and spot on whereas BI is hit or miss with alerts. At this point I almost want to use BI as an archive to have justvin case.
In BI I have person confidence set to 55%

Many have found Coral to not be all that.

Keep in mind that CodeProject is open source being developed mainly by people for free and folks contributing their knowledge and time to develop the program. Maybe someone gets paid by virtue of being an owner of the site and ads, but no idea if that is the case. But I suspect most are donating their time, just like members here donate their time for the overall good of this community.

So you are trying to compare the AI of a paid for-profit company versus the grassroots efforts of a collective group of people donating their time to develop a free open source system.

I have found the Dahua AI to be way more accurate than DeepStack or CodeProject. As it should, Dahua is a $5B company compared to a community based CodeProject development team doing it on their free-time.

But CodeProject is improving all the time. And has way more granular options. And will eliminate the need of paid subscription services to have access to camera AI that some vendors are doing.

Will it ever get as good as the AI of some of these for profit companies? Who knows. But will it get good for most people - yes. And folks not trying to do too much with one field of view and camera are having success with it.

 

[mike2246]

I get it, just makes me wonder how much I missed. Only noticed because I got a new v4 that came with a trial subscription.

 

[wittaj]

That is why many of us have BI show the cancelled alerts so that we can make sure upon review that we don't miss anything.

But to keep getting the Wyze AI, you will need to continue with that subscription.

 

[mailseth]

Why is it my wyze cam alerts are WAY more reliable then my BI/Coral alerts? Same detection zone, same camera. Wyze alerts are always instant and spot on whereas BI is hit or miss with alerts. At this point I almost want to use BI as an archive to have justvin case.
In BI I have person confidence set to 55%

What model are you using? Have you tried a different/newer/larger model? Ideally we will have the YOLOv8 IPcam models soon, which will probably be the best option.

 

[mike2246]

Using the coral tpu one I think there was only 1 option for that device.

 

[mike2246]

Off topic, I have an extra coral m2 tpu I ordered and didn't fit. Mouser refunded and said keep it, if anyone could use it just pay shipping and it's yours.
Model G650-04528-01

 

[mailseth]

The default model for Coral is very small and fast. It’s designed to fit entirely on the TPU with little additional computing. It’s great if you want to analyze at 130 FPS. But it’s not all that accurate.

 

[Skinny1]

Need some help, I just bought a new computer, and I am having a heck of a time with Codeproject Al installing LPR. I have installed it and uninstalled it twice using "Do not use DOWNLOAD Cache" and I am still getting the following error messages.

17:09:58:ALPR_adapter.py: Could not locate cudnn_ops_infer64_8.dll. Please make sure it is in your library path!

Any Ideas?

Also, I am using a Nvidia GeForce RTX 4060 Ti 8gb video card. Which Object Detection should I use? (YOLOv5 .NET) or (YOLOv5. 62) or (YOLOv8 .NET?

Thank you,

Rick