Kernel Dead Issue

[sivaji123256]

Hi @BloodAxe @ofrimasad @jonathan-sha @spsancti @yurkovak ,
Thanks for the cool work. I was trying with the quick notebook on 5 min video.But the kernel was dead issue was rising.I couldn't run the inference. Any idea on why is happening ?

[dagshub]

Join the discussion on DagsHub!

[BloodAxe]

My guess that it is OOM error. In theory if the video has high resolution frames that we accumulate during processing may sum up and take all system memory. Do you mind sharing a video (or at least it's dimensions) you're trying to process?

[Lifeguard-alex]

i get the same problem on
yolo_nas_l.to(device).predict(input_video_path).save(output_video_path)
memory get all 125gb i have and crash + 2 gb of swap

i get this working ok on the same video on yolo 7 and 8 and yolox

i am infer on yolo_nas_l

[Louis-Dupont]

Hi @sivaji123256 and @Lifeguard-alex ,
Please check out #976, does this solve your issue ?

[BloodAxe]

We have implemented memory-efficient predict for video in super-gradients 3.5.0.
So you are welcome to upgrade to a new version of SG which should not have this issue.