Trying to Learn AI the Hard Way

 Today was supposed to be my big "let's learn by doing" day.

I've been planning this for a while. I wanted to test something specific – forking existing AI models and training them with my own photos. I know it can be done. I've read about it. But reading and doing are two different things. I wanted to actually try it myself.

My goal was simple: learn more about AI and AI coding by getting my hands dirty. I also wanted to see how many models you can train at once, and whether training them with my own images actually makes them better.

So I picked two models from Roboflow Universe. Both are for electronics components, but they work differently:

  1. Model A – This is an object detection model. It finds components in an image and draws boxes around them. Like "here's a resistor, here's a capacitor."

  2. Model B – This is a classification model. It looks at an image and tells you what's in it, but doesn't show you exactly where. Like "this image contains a resistor."

Here's where I messed up.

I wanted to train Model B – the classification one. That was the plan. But in my excitement to get started, I grabbed the wrong one. I forked and started training Model A – the object detector.

I didn't even realize until after I uploaded my 17 images and hit the button.

Anyway. Mistakes happen. At least I'm training something, right? That's still learning. I uploaded my own photos – messy ones with weird lighting and my tweezers in the frame – and started the training.

Then the message popped up: "7 hours remaining."

Seven hours.

So I sat there. Waiting. Wondering if my 17 photos would make any difference. Wondering if I should have double-checked which model I picked. Wondering if tomorrow I'd wake up to something useful or just a big waste of time.

But then I checked back after about 4 and a half hours – earlier than expected – and it was already done. The model was ready to go. I went to the deploy section, grabbed the API, and just like that, I had my own working model.

roboflow

Even though I trained the "wrong" one by accident, I still ended up with something useful. In fact, I liked it so much that I built a simple web app around it. Now there's a free online tool where anyone can use this model to detect electronics components, right from their phone or computer.

👉click here:  Electronics Components Detection AI Web App

You just point your camera at a PCB, and it draws boxes around the components it finds. Pretty cool for a "mistake," right?

This whole thing taught me something – sometimes you learn more by messing up than by following the plan perfectly. I wanted practical AI knowledge, and I definitely got it. Plus, I have a working app to show for it.

Not bad for a day that started with a 7-hour wait.




Post a Comment

Previous Post Next Post