Today, the food image classifier got its biggest meal yet: real data.
After days of tuning layers and adjusting learning rates, I finally asked the real question:

“What if the model isn’t the problem?” 🤔

Turns out, it wasn’t.

🦾 What I Did Today

• Switched focus from model tuning to data expansion
• Built a Google Image crawler using Selenium
• Scraped and organized 36 classes worth of food images
• Cleaned corrupted images & restructured the dataset
• Trained a fresh MobileNetV2 on the newly built dataset

📈 Results

• Accuracy jumped from ~14% to 41% 🚀
• Confusion matrix actually shows learning — not just guessing anymore
• Class balance still needs work, but now we’re at least in the game

❗ Takeaway: It’s Not Just the Model

Changing architectures is like swapping out shoes.
But if you’re walking on broken ground (aka messy data), it doesn’t matter what you wear.

Data quality and quantity beat fancy layers. Every. Single. Time.

🧭 Next Steps

• Fine-tune MobileNetV2 now that the base is strong
• Try EfficientNetB0 or ConvNeXt for a comparison
• Experiment with Grad-CAM to visualize attention
• Deploy a simple UI (maybe with Streamlit?) for demo

🧠 Reflection

This was the first time I collected and cleaned my own dataset.
Watching the crawler “hunt” images and seeing them fill up folders in real-time was…
kind of thrilling. 😅

It made me realize:
Modeling is just the tip of the iceberg.
Real work (and real growth) starts when you build your own data pipeline.

Today, I stopped tuning a broken violin.
Instead, I learned to string it myself.