🧠 [2025-06-04] MobileNetV2 Transfer Learning + Theory Review

Today was a combination of experimenting with MobileNetV2 and reviewing CNN-related concepts.
I resumed training on the real-time activity recognizer project using RGB input and pretrained weights, while brushing up on some Keras API logic and CNN structures.

📚 1. What I Studied

I revisited core training logic and neural network components to better understand model behavior and callbacks.

✅ Topics Reviewed

• model.fit(...) arguments and how steps_per_epoch works
• Callback functions like ModelCheckpoint, ReduceLROnPlateau, EarlyStopping
• Simple Dense model definition using Sequential()
• Covariate shift: difference between training/test input distribution and its impact on generalization
  → Explored strategies like data diversity & domain adaptation

🔁 2. Real-Time Activity Recognizer — Experiment_C

After previous grayscale trials failed to improve performance, I returned to RGB format and ran a MobileNetV2-based experiment with pretrained ImageNet weights.

🧪 Experiment C – MobileNetV2 RGB Baseline

• Backbone: MobileNetV2 (frozen base)
• Input: RGB, 128x128, preprocessed with mobilenet_v2.preprocess_input
• Top Layers:
  • GlobalAveragePooling2D()
  • Dense(8, activation='softmax')
• Optimizer: Adam(learning_rate=0.0001)
• Callbacks:
  • ModelCheckpoint
  • ReduceLROnPlateau
  • EarlyStopping
• Training: 125 epochs, batch size = 128
• Result:
  • Accuracy and loss improved steadily
  • RGB clearly outperformed grayscale
  • Graphs auto-saved to:

    C:\Users\ghwns\HJ_git\CV-Projects\real-time-daily-activity-recognizer\figures
    

🧠 Key Takeaways

• MobileNetV2 shows stable performance even with minimal tuning
• Grayscale input brought no benefits in this case
• Revisiting foundational concepts helped clarify model behaviors and training structure

🎯 Next Steps

• ⏫ Try deeper models or increase filter sizes
• 🔁 Experiment with different batch/image sizes
• 🧪 Explore EfficientNetB0, ResNet variants
• 📊 Compare this result with older grayscale & custom CNN baselines

Not a breakthrough day — but a stable, meaningful iteration.
The real-time recognizer pipeline is shaping up well.