7
Remember when we had to train AI models from scratch for every little project?
I was digging through some old drives last night and found a bunch of stuff from back in 2018 when I was messing around with image classification. Used to take me a week just to scrape and label 500 photos of trucks for a basic model. Now with transfer learning and stuff like ResNet50 you can get the same results in an afternoon. Anyone else feel like the pace of these tools has just exploded in the last 5 years?
3 comments
Log in to join the discussion
Log In3 Comments
the_morgan29d ago
Maybe it's just me but "back in my day we trained models on 200 blurry cat photos" hit home so hard. I remember trying to make a dog breed classifier with like 150 images of pugs that were all different sizes and lighting and wondering why it kept misclassifying everything as a cat. If you're still scraping data yourself instead of using something like torchvision or HuggingFace you're wasting a ton of time. There's pre-trained models now that handle the heavy lifting and you just need to fine tune the last few layers for your specific use case. That saved me at least a week when I was working on a vehicle detector.
3
valgibson29d ago
Oh yeah, real vintage stuff right there. Back in my day we trained models on 200 blurry cat photos and called it revolutionary. Now you just sneeze and GitHub has a pretrained model that does your entire job for you.
2
smith.anna1d ago
I remember spending 3 full days just labeling 400 photos of delivery trucks for a basic detection model back in 2018. Now you can pull a pretrained YOLO model and have it working in an afternoon with way better accuracy. It's wild how fast things moved from grinding through every step to having everything ready to go. The whole field feels completely different than it did just a few years back. Do you think the pace is slowing down at all or are we still in the middle of it?
2