Tools

Transfer Learning with Deep CNNs

Procedure

Model Selection

Two pretrained deep convolutional neural network models are used for transfer learning: VGG19 and MobileNetV2. Select one model.

1. VGG19: Deep CNN architecture with 19 layers that uses small convolution filters to learn rich and detailed image features.

2. MobileNetV2: Lightweight and efficient CNN model designed for faster training and low-resource environments. It uses depth-wise separable convolutions and inverted residual blocks.

Freezing/Unfreezing Options

Based on the selected model, the following freezing options are available:

MobileNetV2 Options:

  • Base Freeze: Freeze all backbone layers (0% unfreeze)
  • 10% Unfreeze: Unfreeze last 10% layers
  • 20% Unfreeze: Unfreeze last 20% layers

VGG19 Options:

  • Base Freeze: Freeze all backbone layers (0% unfreeze)
  • 20% Unfreeze: Unfreeze last 20% layers
  • 30% Unfreeze: Unfreeze last 30% layers

Steps

Step 1: Start the Experiment Setup Initialize the Python environment and import the required libraries such as TensorFlow/Keras, NumPy, Matplotlib, and other supporting packages for training and visualization.

Step 2: Load the Dataset

  • Load the Oxford Flowers dataset and organize it into training and validation sets.
  • Perform required preprocessing such as resizing images to a fixed input size supported by pretrained models (e.g., 224×224) and normalizing pixel values.

Step 3: Apply Data Augmentation (If Used) Apply image augmentation techniques such as rotation, zoom, flipping, and shifting to improve model generalization and reduce overfitting on limited data.

Step 4: Model Selection (Choose Backbone) Select one pretrained model backbone from the following options:

  • Option 1: VGG19
  • Option 2: MobileNetV2

Load the selected model using pretrained weights (commonly ImageNet weights) and exclude the original top classification layers (include_top=False). This is necessary because the original top layers are specific to ImageNet classes, which differ from the target dataset. Removing them allows the model to use pretrained features while adding a new classification head suitable for the current task.

Step 5: Build the Classification Head Add new layers on top of the pretrained backbone to match the dataset output classes. Generally, the classification head includes:

  • Global Average Pooling / Flatten layer
  • Fully Connected (Dense) layer(s)
  • Dropout (optional)
  • Final Dense layer with Softmax activation for multi-class classification

Step 6: Set Freezing / Unfreezing Strategy (Interactive Toggle) After selecting the model, choose the layer freezing option based on the model:

For MobileNetV2:

  • Base Freeze: Freeze all backbone layers (feature extractor only)
  • 10% Unfreeze: Unfreeze only the last 10% layers of the backbone for fine-tuning
  • 20% Unfreeze: Unfreeze only the last 20% layers of the backbone for fine-tuning

For VGG19:

  • Base Freeze: Freeze all backbone layers (feature extractor only)
  • 20% Unfreeze: Unfreeze only the last 20% layers of the backbone for fine-tuning
  • 30% Unfreeze: Unfreeze only the last 30% layers of the backbone for fine-tuning

Step 7: Compile the Model Compile the model with appropriate settings such as:

  • Optimizer: Adam / SGD
  • Loss function: Categorical Cross-Entropy / Sparse Categorical Cross-Entropy
  • Metrics: Accuracy

Note: Use a lower learning rate when fine-tuning to avoid disturbing pretrained weights too aggressively.

Step 8: Train the Model with Limited Steps Train the model for a limited number of epochs/iterations as per experiment constraints. During training, monitor:

  • Training Accuracy
  • Validation Accuracy
  • Training Loss
  • Validation Loss

Step 9: Plot Accuracy and Loss Snapshots After training, plot graphs for quick comparison:

  • Accuracy vs Epochs
  • Loss vs Epochs

This provides a clear understanding of model learning behavior under different freezing/unfreezing configurations.

Step 10: Compare Results Across Configurations Compare all experimental results for:

  • MobileNetV2: Base Freeze vs 10% vs 20% unfreeze
  • VGG19: Base Freeze vs 20% vs 30% unfreeze

Analyze which combination gives better accuracy and stable loss.

Step 11: Final Conclusion Conclude which pretrained backbone and which unfreeze percentage performs best for the Oxford Flowers dataset, and summarize the observations based on training curves and validation performance.