3_ImageClassification

Example: MLP for Image Classification - MNIST Dataset

The MNIST dataset is the Iris dataset of Neural Networks when it comes to images. It contains 60,000 training and 10,000 testing examples of labaled hand-written grayscale digits from 0 to 9. Each image is 28x28 pixels, but these will be flattened into 768 dimensions, which will serve as features for the MLP.

example_MNIST.py trains a MLP with 3 layers, dropout, batch normalization and hidden relu activations. $R = 768$, $S_M = 10$. The output function is a Log-Softmax, and the MLP is trained to minimized a Categorical Cross-Entropy performance index.

New Code

• "Reproducibility".
• Loading data from keras.datasets.
• Dropouts and Batch Normalization.
• Using model.add on a sequential model to add an arbitraty number of hidden layers in a for loop.
• Mini-batching and evaluating on validation set during training on model.fit.

Exercise: MLP for Image Classification - FashionMNIST Dataset

exercise_FashionMNIST.py asks to train a MLP on the FashionMNIST dataset, which is exactly the same as the MNIST dataset but replacing the digits with images of different types of clothing. It was made to have a harder benchmark for image classification models. The adaptation from example_MNIST.py is thus trivial, although the MLP accuracy is 10% lower than on the MNIST dataset.

New Code

• Early Stopping
• Loading and Saving Models
• Plotting images