天空вℓσи∂ @skyblond@qoto.org

I'm ready for an 8 hour long water outage tomorrow:

30 litres of tap water (mainly use it to flush the toilet)
25 litres of drinkable (after boiling) tap water (kept sealed).
And 20 litres of bottled water that I can directly directly.

I should be fine.

@freeschool I don't think so. Yolo is the core method to detect birds in a video, and my laptop's 3070 is too weak. A powerful GPU instance will cost money, but will be much faster...

OK... So now basically every VPS provider requires additional verification in order to use their GPU instance.

Just start doing it with AWS. It said my account needed to be verified before I can create GPU instances. Then I got my account verified. After that they said I don't have the quota to create any GPU instance. Now I'm working on increasing that f*ck quota.

I don't know why. After all this processing, I still need to pay for my bill.

"Boost"

I switched to yolov8 L instead of X and I got 4.5 FPS instead of 3 FPS.

My current algorithm is to move the clip to the boring folder if there is no bird detected. If the model says there is a bird, then move to the bird folder.

I checked the results, the boring folder is pretty good. very video file is boring.

The bird folder, however, gives 147 results, but 33 of them are nothing. Maybe it detects some shadow of something. Maybe I should count how many frames have birds and set a minimum requirement before moving to the bird folder.

Well... Another problem: my RTX3070 mobile is too weak to handle the workload. It's overheating and can only process 3 frames per second, which I have already discarded 75% of the frames.

I probably need to rent a gpu server from vultr or something.

I mean, the heavy lift is done by onnx and yolo, I just invoke them and copy files based on the result.

But still, it's exciting to see my code actually working. BTW I'm using kotlin/jvm. So no python nonsense.

YOLOv8 detect birds in video clips.

I think the main issue of this toot is I use the default 640x480 resolution, which makes the picture too small to distinguish the details. Running on 1080P natively does cost a lot of RAM and time, but it can tell it's a bird very quickly and confidently.

I need to go sleep...

The best result I can get with my laptop RTX 3070 8GB is 1.85 FPS, processing 1920x1088 videos, yolov8x

Maybe I should try the smaller model?

Tonight I'm going to stay up late writing code and squeezing the last bit of performance out of my poor graphics card as much as I can. I DONT need to go to bed early, I DONT need to live a healthy life, I NEED to satisfy my curiosity, and I want to see my code run correctly. Healthy living is like God and angels, they only come out when you're dead; and my curiosity, it's the real Mephistopheles. He brings me the true satisfaction and joy.

*that me desperately trying to make my code work*