Transfer learning with MobileNetV2: Alpaca vs non-alpaca classification

We do transfer learning on a pre-trained CNN to build an alpaca vs non-alpaca classifier.

A pre-trained model is a network that’s already been trained on a large dataset and saved, which allows you to use it to customize your own model cheaply and efficiently. We use MobileNetV2 that was designed to provide fast and computationally efficient performance. It has been pre-trained on ImageNet, a dataset containing over 14 million images and 1000 classes.

In this project we perform the following tasks:

• Create a dataset from a directory
• Preprocess and augment data using the Sequential API
• Adapt a pretrained model to new data and train a classifier using the Functional API and MobileNet
• Fine-tune a classifier’s final layers to improve accuracy

I did this project in the Convolutional Neural Networks course as part of the Deep Learning Specialization.

Datasets

We have 327 image files belonging to 2 classes in ‘alpaca’ and ‘not alpaca’ subdirectories. We use Keras preprocessing with image_data_set_from_directory() to create training and validation datasets directly from these files. We use 80% for training and the remaining 20% for validation, resulting in 262 training examples and 65 validation examples. Some examples from the training dataset:

MobileNetV2 architecture

The base model used for the transfer learning (MobileNetV2) has the following architecture:

With the base model, we deleted the top layer (the 1000-class classification layer) and added a binary classification layer, while setting the rest of the base model untrainable. After only 5 epochs of training, the training accuracy is 0.79 and the validation accuracy is 0.77.