Transfer Learning with Pre-trained Models

Transfer learning is one of the most powerful techniques in modern deep learning.

Instead of training a model from scratch on your dataset, you:

1. Take a model pre-trained on a huge dataset (ImageNet)
2. Freeze its convolutional base
3. Add your own classifier layer(s)
4. Train for a short time
5. (Optional) fine-tune deeper layers

Benefits:

• Much higher accuracy
• Works with small datasets
• Faster training
• Lower compute cost

In this section we apply transfer learning to Cats vs Dogs using TensorFlow Datasets

Load the Cats vs Dogs Dataset

TensorFlow Datasets (TFDS) provides cats_vs_dogs fully prepared for use.

Install TFDS if needed:

Load dataset:

Labels:

• 0 = cat
• 1 = dog

Build Input Pipeline

Transfer learning models expect consistent image size, so resize all images.

This pipeline:

• Resizes images
• Normalises pixels
• Batches data
• Prefetches for speed

Load a Pre-trained Base Model (MobileNetV2)

We use MobileNetV2 (fast, accurate, lightweight):

Freeze it:

This makes it a fixed feature extractor.

Add Your Own Classification Head

Typical transfer-learning head:

Why this design?

• GlobalAveragePooling2D → reduces feature map dimensions without dense layers
• Dropout → reduces overfitting
• Sigmoid → output probability of "dog"

Compile the Model

Use a small learning rate:

Train the Frozen Model (Feature Extraction Phase)

Typical accuracy: 93–96% after just a few epochs, this is the power of transfer learning.

Fine-Tuning for Extra Accuracy

Once the head is trained, unfreeze part of the base model:

Unfreeze only the newest convolutional layers (recommended):

Recompile with a very low learning rate:

Train again:

Fine-tuning often boosts accuracy by 1–3%.

Data Augmentation (Highly Recommended)

Adding augmentation improves generalisation:

Use it at the start of your model:

Using Other Pre-trained Keras Models

You can replace MobileNetV2 with any Keras Application model:

All follow the same pattern:

1. include_top=False
2. weights="imagenet"
3. Freeze
4. Add classifier head
5. Train
6. Optionally fine-tune

Full Working Transfer Learning Script