Anyone running DeepStack 2022.01.1?

[cybernetics1d]

Does anyone currently use this latest version Deepstack 2022.01.01 GPU docker in linux container or linux server/VM? I'm currently deepquestai/deepstack:gpu-2021.09.1 docker in a LXC and it works great for me, but whenever i tried to use the latest image, it always come back with status 500 1m timed out response.

 

[digger11]

It's been almost a week now and the Jetson Nano is working fine with 2022-01-1 and BI 5.5.4.5. I am not running any custom models. I am only feeding it 2 LPR cameras during the day nothing at night. Response times are low to mid 200ms.

I've been playing with a 2Gb Nano to see whether it would be viable as a DeepStack server. I'm currently running DeepStack 2022-01-1 with the combined.pt custom model and without starting the default object detection API. (The combined model seems to be a little lighter weight than the default model.)

Under "normal" activity (I wish I had watched it during the snow storm a couple of days ago) it's processing an average of 14 images/minute with an average processing time of 356ms. This is from 5x4mp, 1x3mp, and 5x2mp cameras, all sending mainstream images to DeepStack. It takes ~430ms to process 4mp images, and ~300ms to process 2mp images.

Sending the D1 (704x480) substream images instead was showing sub to low 200ms processing times.

The 2Gb system doesn't seem adequate to handle large models, or multiple models. The dark model by itself barely runs, and trying to run combined and dark at the same time pretty much brought the system to its knees. I haven't been able to get the Nano to run the openlogo model at all. Even bumping up the timeout to 5 minutes never returned a result prior to timeout.

I suspect the 4Gb Nano would be noticeably better, but have to wonder whether 4Gb would be enough to make a robust DeepStack platform capable of running all of the APIs and multiple custom models.

In a niche case where someone only wants to run default object detection, or just a single lightweight custom model, and configures things to just process substream images, the 2Gb Nano seems like a viable choice, and an improvement over DeepStack CPU. It's definitely a lower power consumption alternative to DeepStack GPU on a graphics card.

On the other hand, the Nvidia GT 1030 I picked up at Microcenter for $115 and put in my daughter's BlueIris server is quite handily running DeepStack GPU with the combined model and turning in a 257ms average processing time on the mainstream images of 8 Dauha 4mp cameras, and it's rated at a pretty thrifty 30W power consumption.

 

[sebastiantombs]

So now I'm really scratching my head.

Over the weekend I reverted back from 2022.01.01 to the previous version. I uninstalled all CUDA and CUDnn files as well as DeepStack. Dropped back versions on CUDA and CUDnn as well as DeepStack. Rolled back the NVidia driver as well. Everything is back to normal, CPU around 20% and memory under 40%. No more timeouts or "100" errors. Enabled detection back on about a dozen cameras during all profiles. With 2022.01.01 I had cut it down to three at night and five during the day.

The only thing I can think of is that I've got a trial version of HomeSeer running on the machine. That will end tomorrow or Wednesday when the HomeTroller gets here and I move it off of the BI machine. I may try 2022.01.01 again once that's up an running properly.

The thing that really bothers me is that even f I shut down HomeSeer 2022.01.01 was a disaster, even with the original CUDA, CUDnn and NVidia driver.

 

[bethzur]

Does anyone currently use this latest version Deepstack 2022.01.01 GPU docker in linux container or linux server/VM? I'm currently deepquestai/deepstack:gpu-2021.09.1 docker in a LXC and it works great for me, but whenever i tried to use the latest image, it always come back with status 500 1m timed out response.

Yes, I'm using it on Ubuntu in the Docker container they note on their installation page. It worked out of the box for me. The previous installation required a ton of updating since I have a newer GPU. This one just worked.

Make sure you installed the GPU version and run it with GPUs enabled and that you have the NVIDIA runtime docker installed. Also the version of CUDA that you install much match what's in the container (11.4) and with the same NVIDIA libraries. I installed 470.

I installed all of this on the host like this (after saying yes to Ubuntu's question about using non-free drivers during setup):

sudo ubuntu-drivers autoinstall

I then ran it like this:

sudo docker run -d --restart unless-stopped --gpus all -e VISION-DETECTION=True -v localstorage:/datastore -p 15000:5000 deepquestai/deepstack:gpu-2022.01.1

 

[JL-F1]

So now I'm really scratching my head.

Over the weekend I reverted back from 2022.01.01 to the previous version. I uninstalled all CUDA and CUDnn files as well as DeepStack. Dropped back versions on CUDA and CUDnn as well as DeepStack. Rolled back the NVidia driver as well. Everything is back to normal, CPU around 20% and memory under 40%. No more timeouts or "100" errors. Enabled detection back on about a dozen cameras during all profiles. With 2022.01.01 I had cut it down to three at night and five during the day.

The only thing I can think of is that I've got a trial version of HomeSeer running on the machine. That will end tomorrow or Wednesday when the HomeTroller gets here and I move it off of the BI machine. I may try 2022.01.01 again once that's up an running properly.

The thing that really bothers me is that even f I shut down HomeSeer 2022.01.01 was a disaster, even with the original CUDA, CUDnn and NVidia driver.

Same for me. 2022 constant deepstack restarts. went back to 2021.9 and all is fine again. I didn't even uninstall and re-install the cuda/cudnn, just kept them at the ones I installed for ds 2022

 

[cybernetics1d]

Yes, I'm using it on Ubuntu in the Docker container they note on their installation page. It worked out of the box for me. The previous installation required a ton of updating since I have a newer GPU. This one just worked.

Make sure you installed the GPU version and run it with GPUs enabled and that you have the NVIDIA runtime docker installed. Also the version of CUDA that you install much match what's in the container (11.4) and with the same NVIDIA libraries. I installed 470.

That may be my issue, but when I googled the issue, I saw someone reported that upgrading the driver didn't fix the issue for them so I didn't want to try it yet. I'm running Proxmox 7 host with Debian 10 lxc and is running stable 460 driver with 11.2 CUDA version. Are you noticing better performance with the latest version compare to the previous 2019-09-01 version?

 

[bethzur]

That may be my issue, but when I googled the issue, I saw someone reported that upgrading the driver didn't fix the issue for them so I didn't want to try it yet. I'm running Proxmox 7 host with Debian 10 lxc and is running stable 460 driver with 11.2 CUDA version. Are you noticing better performance with the latest version compare to the previous 2019-09-01 version?

All those things I noted have to be done or it definitely won't work.

I wasn't running the old version very long before upgrading. I upgraded my CPU and motherboard so I needed to reinstall everything (since I also upgraded the OS.) It was pretty easy to get this new version of DS working. I followed the instructions and it just worked. It seems pretty good and fast. I'm using an RTX 3060 which is likely overkill, but that's what I could get.

I felt like my old version was such a hack since I manually updated the container to work with the RTX 30x0-series cards. It did work, but it was fairly bloated in terms of disk space.

 

[cybernetics1d]

Finally decided to look into this and since nvidia driver 470 is now available on Debian 11 via backport so I went ahead and upgraded to this later driver. It's finally working now with the latest docker 2022.01.1 and not throwing any 500 timed out status.

 

[Nunofya]

So now I'm really scratching my head.

Over the weekend I reverted back from 2022.01.01 to the previous version. I uninstalled all CUDA and CUDnn files as well as DeepStack. Dropped back versions on CUDA and CUDnn as well as DeepStack. Rolled back the NVidia driver as well. Everything is back to normal, CPU around 20% and memory under 40%. No more timeouts or "100" errors. Enabled detection back on about a dozen cameras during all profiles. With 2022.01.01 I had cut it down to three at night and five during the day.

The only thing I can think of is that I've got a trial version of HomeSeer running on the machine. That will end tomorrow or Wednesday when the HomeTroller gets here and I move it off of the BI machine. I may try 2022.01.01 again once that's up an running properly.

The thing that really bothers me is that even f I shut down HomeSeer 2022.01.01 was a disaster, even with the original CUDA, CUDnn and NVidia driver.

Have you head of people using this newer version with CPU only working, with no problems?

 

[slabbel]

Have you head of people using this newer version with CPU only working, with no problems?

I'm using the new CPU version without any issues

 

[gouthamravee]

Have you head of people using this newer version with CPU only working, with no problems?

I was using CPU 2022.01.01 without issues, but the GPU version did not work. As others have noted it runs, returns 200, but nothing is detected.
I'm currently on the previous GPU version - gpu-2021.09.1
That is working fine on a Nvidia T400 with the latest production nvidia GPU drivers.