Nvidia P2000 on linux docker issues (500-800ms)

[techflyer]

Running a P2000 on linux with docker. Followed the instructions from the deepstack website. Deepstack 2022, Nvidia 515 drivers with CUDA 11.7. I can see via "nvidia-smi" that the service is running. But seeing detections around 500-800ms. Looking around other people are usually closer to 100ms. Is anyone finding higher times due to 2022 or the CUDA version or the driver version?

 

[sebastiantombs]

I'm running on Win10 but would suggest dropping back to the 2021 version of DeepStack. I, and many others, have had lots of problems with the 2022 version for some reason.

 

[techflyer]

I'm running on Win10 but would suggest dropping back to the 2021 version of DeepStack. I, and many others, have had lots of problems with the 2022 version for some reason.

I will try that this weekend and report back. Thanks for the reply

 

[aadje93]

Hi, we spoke a bit in PM,

I'm running the P2000 in a windows host now, not ubuntu vm -> docker -> deepstack anymore as its now in the purpose built BI machine (which is way overspecced, but need room to grow that 8MP pool )

Things to check;

• What mode do you run, low, normal or high
• Do you use sub-streams or main stream (turn off "use main stream if available"
• What driver, cuda and DS version do you use
• What model do you use? As the the default model checks a whole load of items not relevant for CCTV, so our user @MikeLud1 has made a great custom model, with the latest being only "person/vehicle" so Deepstack doesn't have to look through a whole list of possible matches, just 2. Something with legs or wheels

 

[aadje93]

to add;

I'm using DS 2022.01.1, cuda 11.7.0.516.01 and driver 516.01 so all seems about newest.

My settings;





all cams sync the trigger to this camera (right tab = artificial intelligence). As stated, now the P2000 is running int he BI dedicated box, only because installing custom models is faster.

I'm now testing General 3.1 and animals 2.0 together and seems pretty dang fast, but need to see if the cats are detected. Would be awsome if BI would get a option for animal category for alerts next to persons and vehicles

Main points to note;

no "use main stream if available" High mode for deepstack, disable default object detection due to using a CCTV special custom model, so we don't want the slow default models. Thats also the objects:0 part to 100% make sure to disable default object detection.

 

[MikeLud1]

to add;

I'm using DS 2022.01.1, cuda 11.7.0.516.01 and driver 516.01 so all seems about newest.

My settings;

View attachment 130983

View attachment 130984

all cams sync the trigger to this camera (right tab = artificial intelligence). As stated, now the P2000 is running int he BI dedicated box, only because installing custom models is faster.

I'm now testing General 3.1 and animals 2.0 together and seems pretty dang fast, but need to see if the cats are detected. Would be awsome if BI would get a option for animal category for alerts next to persons and vehicles

Main points to note;

no "use main stream if available" High mode for deepstack, disable default object detection due to using a CCTV special custom model, so we don't want the slow default models. Thats also the objects:0 part to 100% make sure to disable default object detection.

If you uncheck Default object detection you do not need to use objects:0

 

[techflyer]

I assume you have default object detection off because you have a custom model folder? If I turn my off, deepstack drops down to 75ms but alerts are canceled because nothing is found. I am running the latest drivers according to NVidia.

 

[MikeLud1]

I assume you have default object detection off because you have a custom model folder?

Yes, if you are only using custom models on all your cameras then you can uncheck. If any of your cameras are using the default DeepStack model the you need to have it checked.