Hell Yeah! Direct Deepstack Integration - 5.4.0 - March 31, 2021

[Pentagano]

Quite a few of these cards do, my 1660 is the same, fan comes on when required, normally during a heavy Deepstack event, then goes off again, card temp without fan is normally around 50 which I'm not too worried about. You could in theory run the fans all the time on low with a custom curve. Don't forget that the fans will most likely be the first thing to fail, so consider that also

How are the processing times looking if you increase the settings, try the ExDark model maybe.

Just tried the Dark model on some previous snapshots. My custom model made with real dark photos works better.
The dark even on high mode did not capture the cats I trained my model for.
Thanks for the tip though.
I guess I could go through the training again with my photos using the dark api when I have a weekend to spare as it takes hours and hours.

 

[Pentagano]

VRAM showing 0% on asus gpu tweak software. Is this just a bug as it was showing fine yesterday.
How do I see it on the afterburner?

Come to think of it it may be because I'm on remote desktop ? I think on the monitor directly last night it was showing.

 

[jaydeel]

Well, adding :0 after any model seems to have no effect,

It is my understanding that the :0 disable option is for the default objects and faces model options only.

Excerpt from the help pdf…


I've just posted a utility that you might find useful for toggling custom models...

Added Dark Model... process times slower?

BI is checking the image every two minutes or so to confirm static objects. So it should go there, analyze, and release it

ipcamtalk.com

 

[Pentagano]

GPU-Z has a port server - integrated it into my homeassistant card - nice to see.
Also on the badges





telegram alert from nodered if fan hits a certain speed:

 

[CamCrazy]

Maybe not the issue, but it is dark:0 not exdark:0

I know, my typo sorry, but no, it doesn't work

 

[CamCrazy]

Just tried the Dark model on some previous snapshots. My custom model made with real dark photos works better.
The dark even on high mode did not capture the cats I trained my model for.
Thanks for the tip though.
I guess I could go through the training again with my photos using the dark api when I have a weekend to spare as it takes hours and hours.

Understand, I trained my own model with around 50 images and to be honest it took less time than I expected, works well. For my situation the ExDark was not much good either.

 

[wittaj]

A custom model trained to your field of view from your images will always produce the best and fastest responses.

 

[CamCrazy]

A custom model trained to your field of view from your images will always produce the best and fastest responses.

No doubt from my brief testing, although it seems footage in zero light with just IR causes longer processing times for Deepstack using custom model, a lot longer! my times can be 3000-4000 instead of 100-200 during the day with the default model. Works well regardless, just annoys me a bit that the times are so long

 

[Pentagano]

No doubt from my brief testing, although it seems footage in zero light with just IR causes longer processing times for Deepstack using custom model, a lot longer! my times can be 3000-4000 instead of 100-200 during the day with the default model. Works well regardless, just annoys me a bit that the times are so long

3-4 seconds?? oh boy.
I run 3 cams with a custom model and standard model. Both take about 100ms. Custom only has dark images and is on Mode hight also. But I don't use IR. I did before but still wasn't much difference in time.

 

[wittaj]

No doubt from my brief testing, although it seems footage in zero light with just IR causes longer processing times for Deepstack using custom model, a lot longer! my times can be 3000-4000 instead of 100-200 during the day with the default model. Works well regardless, just annoys me a bit that the times are so long

WOW - was your model trained with the dark B/W and IR images?

 

[CamCrazy]

3-4 seconds?? oh boy.
I run 3 cams with a custom model and standard model. Both take about 100ms. Custom only has dark images and is on Mode hight also. But I don't use IR. I did before but still wasn't much difference in time.

Annoying isn't it, I believe it is simply due to the scene, these cameras are in pitch black with average IR, big challenge. I'm running 10 cameras so far. Try running a few cameras in the dark with IR and watch those times go up

Daytime processing times are low, ie 60-150ms, I dream of streetlights

 

[CamCrazy]

WOW - was you model trained with the dark B/W and IR images?

I trained my night model with essentially vehicles headlights, taillights, side view etc, works well

 

[IReallyLikePizza2]

Active cooling for the P4 arrived today. Can't wait to put it in, its been thermally throttled 24/7 since I put it in (No surprise since its a small heatsink with no fan!)

 

[CamCrazy]

Active cooling for the P4 arrived today. Can't wait to put it in, its been thermally throttled 24/7 since I put it in (No surprise since its a small heatsink with no fan!)



View attachment 106494
View attachment 106495

View attachment 106496

I like the colour. Don't understand why such an expensive card when new would be designed thermally inefficient? doesn't make sense. If they thought the cooler wasn't enough then make it bigger or add a fan(s). Seems like a classic case of make the customer fix the problems we couldn't be bothered to

Am I missing the point, were these cards designed for another purpose which wouldn't cause this issue?

 

[IReallyLikePizza2]

I like the colour. Don't understand why such an expensive card when new would be designed thermally inefficient? doesn't make sense. If they thought the cooler wasn't enough then make it bigger or add a fan(s). Seems like a classic case of make the customer fix the problems we couldn't be bothered to

Am I missing the point, were these cards designed for another purpose which wouldn't cause this issue?

These cards are designed to be installed in servers, which will have high airflow

 

[CamCrazy]

These cards are designed to be installed in servers, which will have high airflow

With a relatively low load, if it's throttling now then how much airflow does it need from case fans to stop that. I get it that these are rack mount server cards but regardless, seems like poor design. See how the new blower works out

 

[IReallyLikePizza2]

The airflow in servers with these cards is designed to blow a lot of air in that location. My server closet is always very warm, and there is minimal airflow in the case

The general idea is that you want ALL of your fans to be hotswap, and if you put one on the card, its over

 

[IReallyLikePizza2]

I've upped the amount of work the GPU is doing a lot, I've added deepstack to all my cameras. Usage is higher now, I've seen up to 20%

Temperature is now stable

 

[IReallyLikePizza2]

The i350-T4 NIC will have to take one for the team on airflow

 

[sebastiantombs]

See if upping the number of instances of DS in the AI tab of the console. I upped mine to two and cut detection times in half or even more.

Air flow is why I'll never fool with a SFF. Well, one reason anyway.