Windows GPU - Long detection times over long run times

I’ve been running DeepStack on a Windows GPU install for over 6 months now and have noticed one issue, the longer the service runs the longer the detection delay. It starts around 200ms for a response but slowly grows over the course of a couple of days to over 2s. Stopping the service and reopening resolves the issue, it does seem to be related to GPU memory getting full and closing it will clear memory.

I currently “solved” this issue by running a batch file every night that’ll close all instances of Deepstack and open them again. This is an alright solution except for if images are currently processing when it does this which causes errors in the log. If anyone has a more elegant solutions I would love to try it, ie flushing GPU memory without reloading Deepstack?

The second issue with this is that the first detection request always takes quite a while, up to 10 seconds in some cases. Considering this is for security cameras and I am running 3 instances it can be a considerable issue and can even trigger timeout errors as well. Is it possible to send a dummy request through the Command Line/PowerShell to each instance? Even if I need to point to a static image on the drive, I am fine with that, as long as it gets the initial request delay dealt with. I couldn’t find any examples if sending a request via command line.

1 Like

@charredchar thanks for posting on this issue and apologies for giving a late response. Please create an issue on this on the DeepStack Github repository.