Today I focused on improving the food image classifier I started yesterday.
The first model was like a sleepy turtle 🐢 — cute but very, very slow (and kind of blind).
Time to wake it up with some fine-tuning magic and proper evaluation.

🔧 What I Did Today

• Enabled fine-tuning for the top 100 layers of ResNet50
• Adjusted the ImageDataGenerator to be less aggressive (because maybe the model couldn’t handle spicy augmentations 🌶️)
• Added ReduceLROnPlateau to dynamically lower learning rate
• Introduced EarlyStopping to avoid wasting time when the model gets lazy
• Extended training to 50 epochs (with early exit)
• Added proper evaluation: confusion_matrix, classification_report, and accuracy_score

📉 Results & Observations

• Accuracy was still around 14% — ouch
• Confusion matrix revealed prediction bias: the model kept guessing the same few classes
• Class distribution in y_pred showed heavy imbalance — not a good sign
• Turns out: I was training on CPU the entire time 😭
  (1 epoch ≈ 90 seconds. Now it makes sense.)

💾 Takeaway & Action Plan

Despite the low score, I learned a lot today:

• Fine-tuning needs time (and power)
• Evaluation tells a bigger story than accuracy alone
• And most importantly: don’t forget to check if you’re using your GPU

Tomorrow?
GPU setup. No excuses.
It’s time to make this model sprint, not crawl.

🧠 Reflection

This project is starting to feel real.
I’m tuning layers, inspecting predictions, fixing data flow —
and slowly turning this from a hobby into a discipline.

Today wasn’t a big win. But it was a true step forward.
And that’s more than enough.