5.4.9 - June 25, 2021

Interesting... What are the specs of your PC? I'm using an i7-4790 CPU with 16 GB of RAM.
 
So I uninstalled Deepstack and reinstalled it. Re-started BI, took the new update (5.4.9.10), and alas, the same issue of deepstack essentially doing nothing when I have custom models or facial recognition enabled. I decided to manually call deepstack through powershell, with my own settings, which should mirror what BI is giving it. And boom, it works. Seems like maybe the "auto start/stop with Blue Iris" is causing Deepstack to launch improperly. Can anyone else verify? Here is the specific powershell command I used to start Deepstack, and like I said, the settings are the same as I plugged into the BI settings AI tab:

deepstack --VISION-DETECTION True --MODELSTORE-DETECTION "C:\DeepStack\MyModels" --VISION-FACE True --PORT 5000
 
  • Like
Reactions: sebastiantombs
@stev067 Are you checking to make sure DS is running when you turn it on in BI? My "standard" procedure when I'm testing is to enable models, point to the models directory, stop DS, wait a minute, re-start DS and then double check to make sure it's actually running. I used that when I was running the CPU version and still do the same with the GPU version.
 
@stev067 Are you checking to make sure DS is running when you turn it on in BI? My "standard" procedure when I'm testing is to enable models, point to the models directory, stop DS, wait a minute, re-start DS and then double check to make sure it's actually running. I used that when I was running the CPU version and still do the same with the GPU version.
Good call. The "auto start/stop" is not starting DS unless I have facial recognition and custom models disabled.
 
  • Like
Reactions: sebastiantombs
I ended up shooting myself in the foot with this, too. If it wouldn't start, I'd shut down BI, reboot, then set BI back to run as a service. That always got it all going, for me anyway.
 
I ended up shooting myself in the foot with this, too. If it wouldn't start, I'd shut down BI, reboot, then set BI back to run as a service. That always got it all going, for me anyway.
Well I thought I was on to something, but no matter how I do it now, I'm getting error-100's on something, whether it's just the faces portion, or everything. Inconsistent problems are the hardest to figure out.
 
  • Like
Reactions: looney2ns
It seems that faces is the most difficult one. I believe I read somewhere that faces and dark do not work well together yet, but I think it has worked for some. One of the updates when I was testing it did but it didn't last.
 
  • Like
Reactions: looney2ns
5.4.9.x series has not been happy whenever I turn on the dark model, lots of DS errors, evident when running clips through the real-time analyser. 5.4.9.10 seems to work better, just updated to it and ran a few clips through the analyser, no errors at all. Of note, the idle CPU percentage with this version also seems to be lower, so I think that the extra CPU overhead is allowing DS to work better (running on an i7-7700K). I've got a Quadro P400 on the way, hopefully that will help with getting the most out of DS by offloading it to the GPU.

Update: After running 5.4.9.10 for a while now, DS seems to be working well with the dark model enabled, no DS errors being logged.
 
Last edited:
  • Like
Reactions: looney2ns
I think I got custom models and facial recognition working now... Here are some pointers that weren't really obvious to me:
Once custom models are enabled on the main AI tab, all custom models are enabled by default for all profiles on all cameras. In my case, I wanted to enable dark.pt for my night profile on my camera, but not have it apply to my day profile. So on my camera's trigger/AI tab for the day profile, I tricked it by putting a semi-colon in the "Custom Models" tab, so it will not run dark.pt during the day. And then on my night profile, that field can be left blank, or specified as "dark.pt". I believe I was just putting "dark" before, and that was causing errors.
I do still have an issue with static object detection working inconsistently, but I made another thread on that.
 
Has anybody had issues with recordings not switching to main-streams on 5.4.9.10? I've just noticed on 2 x cameras where continuous + alerts is selected, they will trigger and send to deepstack, confirm objects and execute alert actions but don't appear to start recording the main stream.

EDIT: Also just tried continuous + triggered - same results, only recording sub-stream. Strange...
 
Last edited:
I haven't noticed that, yet, but I will double check later. One thing I have noticed about 5.4.9.10 is that memory management has seemed to have gotten much better. I've been tracking it for quite a while, since February, and it has normally ranged between 2 and 4.5GB over a day. Now it's been stable and under 3GB since I updated yesterday.
 
Last edited:
I haven't noticed that, yet, but I will double heck later. One thing I have noticed about 5.4.9.10 is that memory management has seemed to have gotten much better. I've been tracking it for quite a while, since February, and it has normally ranged between 2 and 4.5GB over a day. Now it's been stable and under 3GB since I updated yesterday.

Thanks - I'm going to do a full re-install later just to see if that fixes. Agreed on memory usage, mine too is much more stable on this latest release. My mem range throughout 5.4.9.x hasn't fluctuated as much as yours however.
 
@hikky_b I just went back and checked all my cameras. Only two don't run sub streams because they are not capable of that, but all the others switch fine both when going full screen and during playback of clips.
 
@hikky_b I just went back and checked all my cameras. Only two don't run sub streams because they are not capable of that, but all the others switch fine both when going full screen and during playback of clips.

Thanks for checking - it must be something that's gone wrong through updating - it has been working previously and does work when going full-screen, just not recordings. I'll proceed with re-install and update here once finished.
 
  • Like
Reactions: sebastiantombs