Today I completed training my first image classification model.
It’s based on ResNet50 and built to recognize food images.
Was it perfect? Nah. Was it mine? Hell yes. 😎

🔧 What I Did Today

• Connected a pre-trained ResNet50 as a feature extractor
• Added custom GlobalAveragePooling + Dense layers on top
• Froze the base model (no fine-tuning yet)
• Used categorical_crossentropy for loss and Adam for optimization
• Trained the model for 10 epochs with ImageDataGenerator
• Watched accuracy crawl up slowly like a lazy snail 🐌

📈 Results & Evaluation

• Plotted accuracy and loss graphs — they’re moving in the right direction (thankfully)
• Built a confusion matrix — most classes were okay, but paprika and pomegranate confused the heck out of the model
• Visualized sample predictions — some hits, some misses, but it’s working!

💾 Model Saved!

model.save('models/food_classifier_resnet50.h5')

That file now holds all my effort, frustration, and “why is my accuracy still 0.1” moments.

🧠 Reflection

I finally feel like I’m understanding how these pieces fit together:
generators, model layers, predictions, evaluation.
It’s not just running code anymore — I get it now.
And honestly? That feels better than any accuracy score.

Tomorrow? Maybe fine-tuning.
Today? We celebrate 🥳