What is Fine-Tuning a Model?

The primary objective of fine-tuning is to modify specific features from a general dataset to suit a new dataset without discarding the foundational knowledge acquired from the generic learning process. In this instance, I have fine-tuned the top layers of the pre-trained MobileNet-V2 model.

Find the Jupyter Notebook containing the code [here](#).

Parameters to Adjust:

• Learning Rate: Selecting an optimal learning rate is crucial for enhancing accuracy. In this case, we utilized a learning rate of 0.0001.
• Number of Epochs: Training the model for a greater number of epochs can lead to improved accuracy. Here, we trained the model for an additional 10 epochs, totaling 20 epochs.
• Unfreezing Top Layers: Instead of fine-tuning the entire MobileNet model, we focus on a limited number of upper layers.

Unfreezing the Top Layers of the Model

The first step involves unfreezing the base model while keeping the initial layers untrainable. After recompiling the model to apply these changes, we resume training for an additional 10 epochs, using a significantly reduced learning rate of 0.0001.

Upon fine-tuning, the model achieves nearly 98% accuracy.

# Set the top layers as trainable

base_model.trainable = True

# Display the number of layers in the base model

print('Number of layers in the base model: ', len(base_model.layers))

# Determine the layer from which to start fine-tuning

fine_tune_at = 100

# Freeze all layers prior to the fine_tune_at layer

for layer in base_model.layers[:fine_tune_at]:

layer.trainable = False

# This step enhances model accuracy incrementally

model.compile(loss=tf.keras.losses.BinaryCrossentropy(from_logits=True),

optimizer=tf.keras.optimizers.RMSprop(lr=base_learning_rate / 10),

metrics=['accuracy'])

Fine-Tuning Parameter Weights

To enhance performance further, we can adapt the upper layers of the pre-trained models to our new dataset through fine-tuning. This method is particularly effective when the training dataset is substantial and closely resembles the original dataset that the pre-trained model was initially trained on.

Results Following Fine-Tuning

Before undergoing fine-tuning, the model achieved an accuracy of 93.25% with a loss of 0.14 on the test set. After fine-tuning, these metrics improved to an accuracy of 97.89% and a loss of 0.058 on the test set.

Accuracy and Loss on Test Set (Before Fine-Tuning):

• Accuracy: 93.25%
• Loss: 0.14

Accuracy and Loss on Test Set (After Fine-Tuning):

• Accuracy: 97.89%
• Loss: 0.058

Now, take the plunge and implement hyperparameter tuning in your own deep learning projects!