5.4.9 - June 25, 2021

[sebastiantombs]

I removed the CPU version of DS and installed the GPU version. Moved all the .dll, .lib and other files for cuDNN to their appropriate directories and made sure the environmental variables were set correctly. I also found out that I have DS set to high for detection.

Bottom line is, at least for now with about a half hour of running, is that detection is WAY up, as in back to 95% or greater. Detection is happening a lot faster, roughly ten times faster. A typical detection was taking 500-1000ms with the CPU and quite frequently ran into 1500+ms and occasionally even longer. The GPU version is detecting in as little as 55ms and typically stays under 200ms from what I've seen so far from the .dat files.

I did try using dark.pt for a few minutes, two different detections and got the -1 error. Tying that made me think that I may need to setup a models directory in the new install and put the model in it before it will work. DUUUHHHHH.

 

[kc8tmv]

I also found out that I have DS set to high for detection.

Do you still have it set on HIGH or did you change that?

 

[beepsilver]

If I remember correctly you can only use the GPU version if you have a separate video card? Nvidia?

Edit: found it, lazy me should have looked before asking.


Install DeepStack GPU
To install and use DeepStack GPU version on your Windows machine, follow the steps below

• Ensure your machine has an NVIDIA GPU
• Download and Install CUDA 10.1 from NVIDIA
• Download and Install cuDNN from NVIDIA

Once the above are installed, download and run DeepStack GPU version for windows via the link below.

 

[sebastiantombs]

@kc8tmv (I keep typing your call sign as kc8MTV ) Yes, it's still set on high.

@beepsilver Yes, and it has to be an NVidia card for DS to utilize it. It uses a developer platform from NVidia along with some NVidia developer tools added to that platform. It's a little complicated compared to the normal "point and click" installs we're all used to, but not really much of a big deal. That is where the reference for the cuDNN files cam in.

 

[kc8tmv]

(I keep typing your call sign as kc8MTV ) Yes, it's still set on high.

"Video Killed The Radio Star!"

 

[sebastiantombs]

dah dit dah

 

[sebastiantombs]

The models directory was still there complete with dark.pt in it. I guess the uninstall wasn't all that complete. Using the additional model doesn't work. On th bright side, DS is still running at 95% or better, more like 100%, detection rate as the scene is getting darker and darker outside, evening and heavy clouds with, maybe, some T&L.

 

[Vettester]

Thanks to all you bleeding edge diehards.

Somebody’s got to do it.

 

[Vettester]

I've been keeping up with the updates always running the latest version without using any custom models. Some of you have reported that dark.pt is working in 5.4.9.6 so I decided to give it a shot. After re-enabling the custom model I noticed that every canceled alert had an error and my alerts weren't working at all.



To resolve this I tried disabling the custom models, but that didn't work and the errors were still there. I ended up stopping the BI service in Windows and deleting the Perspective entry in the Windows registry. After restarting BI the DS errors were gone and everything started working again.

 

[beepsilver]

I've been keeping up with the updates alway running the latest version without using any custom models. Some of you have reported that dark.pt is working in 5.4.9.6 so I decided to give it a shot. After re-enabling the custom model I noticed that every canceled alert had an error and my alerts weren't working at all.

View attachment 94793

To resolve this I tried disabling the custom models, but that didn't work and the errors were still there. I ended up stopping the BI service in Windows and deleting the Perspective entry in the Windows registry. After restarting BI the DS errors were gone and everything started working again.

View attachment 94794 View attachment 94795

Thanks for the report...curious if you get the errors again after your restart.

 

[beepsilver]

I walked around the house to trigger all my cameras this evening. Have dark.pt and default running with DS on medium. No errors. But on my second pass while in the backyard, my wife drove up and into the garage. 4 out of 5 cameras reported error 100 and no alerts were saved. I suspect headlights are the problem, but I can't be sure.

 

[Vettester]

curious if you get the errors again after your restart.

I don't have custom models enabled so I'm not sure. If I get some time this weekend I'll re-enable the dark.pt and see what happens.

 

[wittaj]

I walked around the house to trigger all my cameras this evening. Have dark.pt and default running with DS on medium. No errors. But on my second pass while in the backyard, my wife drove up and into the garage. 4 out of 5 cameras reported error 100 and no alerts were saved. I suspect headlights are the problem, but I can't be sure.

Do you have the object detection and reset at a certain % enabled on those cameras with the headlights - wouldn't explain the error 100 but maybe who knows with some of these LOL

 

[beepsilver]

Do you have the object detection and reset at a certain % enabled on those cameras with the headlights - wouldn't explain the error 100 but maybe who knows with some of these LOL

Just the zone A under object detection but nothing checked for resets for size or object travels

 

[kklee]

Just reporting on my findings with DS set to Low and custom mode dark enabled. It seems to be working well, no DS errors and detection seems to be the same as Medium. I'll leave it like this for a while longer and see how it goes.

Just as a test, I tried to set up DS on another PC that runs 24/7 and is mostly idle, hoping that moving the DS load off the BI system would help. I pointed the BI AI config to the other PC and BI talks to it, but something's not right. DS on the other PC shows detection activity from BI, but BI returns errors I run a clip through the analyzer. Probably missing something obvious...

 

[sebastiantombs]

Overnight, DS running on the stock model, I got about 97% detection on one camera and 5% on another. I attribute that to headlight bloom and lack of illumination on the 5% camera. The 97% has light from a street lamp helping and is a 5442 versus the other one being a 3241. I did raise the number if images to process on both so I'll see what happens tonight.

@kklee I tried the same thing using another machine. The other machine is an older quad core but has a gig connection. It would detect but never got things back to the BI machine quickly enough to result in a detection. Timing is everything.

 

[stev067]

Hey guys, I'm new and have spent the past week fine tuning my system, and was pointed to this thread. I'm not sure if this is relevant or common knowledge, but is it normal/expected that the break time will cut the real time images short if it isn't set long enough? For example, I had my break time set for 4 seconds, and had set 20 real time images, twice per second. So the real time images should take 10 seconds to complete, but the break time cut it short at 4 seconds, for a total of 8 images. Once I realized that, I upped my break time to 10 seconds, and got all 20 images.
Also, I have never set up custom models, and I have zero errors in my log, been running a week at the latest versions.

 

[kklee]

Overnight, DS running on the stock model, I got about 97% detection on one camera and 5% on another. I attribute that to headlight bloom and lack of illumination on the 5% camera. The 97% has light from a street lamp helping and is a 5442 versus the other one being a 3241. I did raise the number if images to process on both so I'll see what happens tonight.

@kklee I tried the same thing using another machine. The other machine is an older quad core but has a gig connection. It would detect but never got things back to the BI machine quickly enough to result in a detection. Timing is everything.

Checking out the overnight activity, I'm not getting any DS errors, but object detection tanked, so I think I'll leave it at medium without custom/dark, that seems to work the best for my system.

My other machine is an i7-7700K, 16GB RAM, 1Gbe, same as the BI system, so I was hoping to offload DS to it. Might play around with it some more today and try some packet captures to see what going on.

One other discovery I made is that after changing the DS settings in BI, I need to reboot before they take effect, toggling the DS Stop/Start in the AI settings tab doesn't help enable the changes. When I check Task Manager after fiddling around with settings, I notice that there are multiple instances of Python running, so I think the original instance of DS is still running and BI is starting new instances that are not working.

 

[sebastiantombs]

I've repeatedly turned DS on and off without rebooting, and have turned models on an off as well, and have never had a problem, other than the problems with detection I've noted in other posts. It just seems to work for me. On a brighter note, the GPU version is working quite well with really fast detection times, sometimes on a partial image as the target enters the frame. I have said for a while that the amount of time to make a detection is critical and I suspected that was my problem with an older i7. Switching to the GPU version seems to have proved it in my case at least.

 

[kklee]

Unfortunately, I gave away the only spare GPU I had quite a while ago, otherwise I'd switch to the GPU version. Guess I'll stick with my current config until I can get my hands on another nVidia card.