External GPU providers - how to choose what is enough?

[autumn coral]

Dear everyone,

i just learned basics of deep learning (namingly YOLOv5 v. 6.2) to use in my ecology project.

Background: Basically, I filmed wild bees in front of a nesting aid on a scheduled script. So, I have 1 hour material with potentially a lot of "no bee" frames. Thus, we trained YOLOv5 to detect bees and throw out "no bee" frames (one class "bee", AP0@0.5:0.95=0.78).

Challenge: Now, we have the challenge to finally run YOLO on all the video material (around 2,100 hrs material with 60 fps). Of course, we do not have the GPU power to do so. A GeForce RTX 3070 needed around 4 hours for 1 video. I am looking now in external GPU providers and I have no clue, what service I possibly need and how much it costs. I looked into Amazon AWS (they have different services and I do not even understand the pricing [is the comma a thousand separator or a Euro-cent separator :D] and google colab.

Question: Do you have any advise which offer to choose and how long it would take to run YOLO on the 2,100 hrs of material?

Thank you very much for any comment! I am not a computer scientist and often, this shows when meddling with this.

[naive pasture]

@glossy rose @unique wren@timber cloak Do you guys have thoughts on this?

[glossy rose]

It sounds expensive 😅

[unique wren]

@nimble basalt did a study for prices of the different providers and building your own server, etc.! Might be the best to answer this question if he's still active in here

[timber cloak]

?

[timber cloak]

i heard AWS has good gpus for cheap price, also, maybe reducing ress and fps may help

[dull meadow]

Consider a faster network
Face detection + cropping + alignment for faces is doable in one millisecond per frame https://www.cv-foundation.org/openaccess/content_cvpr_2014/papers/Kazemi_One_Millisecond_Face_2014_CVPR_paper.pdf
But since you only need to do the detection part, you might want to start here to build your own model

[glossy rose]

Yeah, I definitely agree it would be a good idea to make sure the whole process is absolutely as optimized as possible before investing in a cloud run. It sounds surprisingly expensive for what it sounds like you are trying to do and I would try to avoid using brute force here.

[autumn coral]

Thank you for all your answers! The more I look into it, the more it dawns to me that my project was way to courageous for me. For YOLO, there were many tutorials out there, so it was doable with help of a computer science friend. But now I am stuck with limitations everywhere. Time to downsize and subset my videos - and give AWS a try, I guess. I also thought of background filtering as an alternative, as most visitors of the nesting aid will be my mason bees anyway. Unfortunately, there are not as many step-by-step tutorials for beginners out there as for YOLO, but well, time to learn something new 🙂