Using VGG16 network trained on ImageNet for transfer learning and accuracy comparison

The same task has been undertaken using three different approaches in order to compare them.

Task: Image classification
Dataset: Dogs vs Cats dataset from Kaggle.
Dataset size: 1000 x 2 training images. 400 x 2 validation images.

Keras provides ImageDataGenerator class for performing data augmentation. We will use this to augment the data via random transformations. Rotation, scaling, shearing, flipping, zooming are some possible ways.

Training a small convnet from scratch.
Architecture: A very small convnet with few layers and few filters per layer, alongside data augmentation and dropout. Dropout and augmentation help avoid overfitting.

Using the bottleneck features of a pre-trained network.
I have used the VGG16 architecture which was pretrained on the ImageNet dataset.
The strategy is to only instantiate the convolutional part of the model, everything up to the fully-connected layers. Then run the model on training and validation data, record the output (bottleneck features), then train a small FC model on top of those features.

Fine-tuning the top layers of a a pre-trained network
Fine-tuning is when you start with a trained network, then retrain it on a new dataset with very minor weight updates. This can be done in three steps:

• instantiate the convolutional base of VGG16 and load its weights
• add our previously defined fully-connected model on top, and load its weights
• freeze the layers of the VGG16 model up to the last convolutional block

The first and second approaches were run for 50 epochs each. The third was much more computationally intensive and hence was run only for one epoch and yet displayed a very high accuracy.