Search results

‘Best’ practices is going to be what works best for you. People tend to be pretty happy with the custom IPcam models, but they aren’t really available right now on the TPU. YOLOv8 is the best in theory, but is going to require more compute and so may not be the best for you. Similarly, there may...

Not YOLOv8? I trust your opinion, but am curious why?

I wonder how large of a photo/label pairing dataset you could scrape off of Craigslist or Edmunds? Would probably be more complete in the past ten years.

The easiest way to ‘fix’ the problem you’re seeing would just be to be running the multi-TPU code on your single TPU. Do you see any earlier problems that would show why that’s not working?

Yeah, I wouldn't expect to see much difference between 0.03 FPS and 1.0 FPS as far as analysis load goes. It's going to really only make a difference when everything is moving because it's windy or raining.

Thanks for the report. I haven’t seen that problem before. From the stack trace, it looks like it’s using the original single-TPU code which hasn’t been changed recently (as far as I know.) I’m on my phone now, I’ll take a closer look when I get a chance.

If I'm reading that correctly, that's ~700 frames analyzed (motion triggers) over 6 hours, for around 1 frame every 30 seconds? How many frames does a motion trigger cause to be analyzed?

Good to know. What sort of video stream load are you putting on it? (Average FPS?)

Sorry to hear that. I hope it’s better in the next version.

That would be an API change between BI and CPAI, I’d guess. BI would need to send the info to CPAI in order for it to appear in anything on the CPAI side, and I don’t think it does. I’m not familiar with the details of the API, but I haven’t seen anything like that in there.

The most important part of my calculus ended up being power consumption, which is pretty hard to measure. Running a GPU on a steady stream of video for a year is going to cost a bit. For me, for example, I calculated out that if the card ended up using an extra ~40 watts, I'd spend $100-$200 per...

I’m hoping that the guy finishes this board: https://github.com/magic-blue-smoke/Dual-Edge-TPU-Adapter/blob/main/PCIex4_2xDualTPU_s.jpeg And then we can drop a few duals into a single slot.

I do not have experience. It was something I looked into and it seemed interesting until I learned I needed bifurcation and didn’t have it. With that one in particular you may need 4-way bifurcation, which is even more rare, but as I understand it if you only have 2-way that just means that only...

There is also this one, but your motherboard has to be compatible with bifurcation. https://a.co/d/cEtN1G7

For the original models: https://coral.ai/models/object-detection/ There are really two models (MobileNet & EfficentDet) in your comment in different sizes and variations. Mainly think of the different sizes. Then, for models that don't fit well on one TPU, I've been playing with segmenting...

Here's a feature branch to base any USB work off of that will make it much easier to keep things in sync: https://github.com/seth-planet/CodeProject.AI-Server/tree/usb_debug I wonder if the training set from that link can be adapted to train a YOLOv8 model? You should look into the docs here...

If you’re starting from ‘scratch’ with an existing .tflite file, the first step would be to instal the TPU compiler and converting to that format: https://coral.ai/docs/edgetpu/compiler/ I think the YOLOv8 export takes care of that for you, but it may still be needed if you find your own...

Possibly. I found it has a CPU-only instruction midway through the model so it doesn’t run as well as you’d think it should. Everything after that instruction ends up also being CPU-only. (Unless you’re running multi-TPUs, because then you can run the remaining instructions on the next TPU segment.)

Most models can be used on a Coral TPU but require some extra steps to get them exported as the appropriate .tflite file and then more steps to get them segmented across multiple TPUs. YOLOv8 is relatively easy to export since they have the settings built into their export feature, but it’s not...

Also, the original module code mixed different models under the label of MobileNet, so one thing to check is what the model is that is actually being loaded.

The funny thing about accuracy measurements is there are so many false positives to get rid of also, and everyone has different opinions about the cost of false positives. It also just occurred to me that if anyone is setting up a system for development, it should be based off of a GitHub...

I didn’t write that code, but we are trying to use CPAI model configurations in a way that it hadn’t been originally intended. Both because we are wrapping up a number of different sizes and types of models in the Coral module and because we are trying to only download the models on demand. So...

lol. Once you have kids (two under two until recently!) you lose control of time you didn’t even know you had and in ways you were were able to lose. I swear I’ve lost twenty IQ points on top of everything else. Feel free to use whatever os/install is most useful for you. After all I’ll be...

I still have all the time limitations of two young daughters, but maybe I should take you up on that offer. Next time there’s a release of CPAI and you’re feeling like you have the time to get everything installed and set up on your end, we could look into it. I’d probably need a walk through of...

That's part of my debugging problem, I'm not running CPAI or have it setup right now. I'm just running the core TPU code outside of CPAI to work on and test the performance. For example, the two next things I'm planning on looking at are a reorg of the internals to reduce context switching...

Part of the problem is that I suspect the logs are enough. There tends not to be enough logged, so someone needs to get in there and start printing out all of the intermediate values. And the unstable parts are even harder. I've tried to put everything together in such a way that the system can...

Yeah, I believe you. “It works on my machine.” Sounds like a poor excuse, but really does make it hard to debug and fix issues. And, like I said, all my TPUs are busy right now working on other things and haven’t actually been running CPAI in a real-world way for months.

Yeah. Something like that should work in theory, but I don’t have a USB one to test with. I’m sure I made an assumption somewhere but don’t know where. Seems silly to spend > $60 on flakey and slow hardware just to see how it breaks. But I may do it anyway just to make it happen.

Alright. I know there are a number of issues kicking around that I may not be able to debug anytime soon since I’m not in a position to do so. I’ve been only in the core of the code and even then I’ve been testing with particular load patterns and lack of driver/hardware failures.

Honestly, I don’t really understand what would cause you to have to reinstall to change the model size, but I haven’t been running that part of the code so I can’t speak to it. It’s an area we’ve actively been working on, however, to try to reduce the CPAI bandwidth cost. So if you’re feeling...