Description
In class, we have built (or will build) a classifier for the popular MNIST dataset of handwritten
digits. In this assignment, you will work with an analogous data set called Fashion-MNIST. Instead
of handwritten digits, there are images of 10 different classes of fashion items. One image of each
type is shown below.
You can load the data using the following commands:
fashion_mnist = tf.keras.datasets.fashion_mnist
(X_train_full, y_train_full), (X_test, y_test) = fashion_mnist.load_data()
The Fashion-MNIST data has exactly the same structure as the MNIST data. There are 60,000
training images in the array X_train_full and 10,000 test images in the array X_test, each of which
is 28 × 28. The labels are contained in the vectors y_train_full and y_test. The values in these
vectors are numbers from 0 to 9, but they correspond to the 10 classes in the following way:
Label Description
0 T-shirt/top
1 Trouser
2 Pullover
3 Dress
4 Coat
5 Sandal
6 Shirt
7 Sneaker
8 Bag
9 Ankle boot
Before you begin, preprocess the data in the following way.
1. Remove 5,000 images from your training data to use as validation data. So you should end up
with 55,000 training examples in an array X_train and 5,000 validation examples in an array
X_valid.
2. The numbers in the arrays X_train, X_valid, and X_test are integers from 0 to 255. Convert
them to floating point numbers between 0 and 1 by dividing each by 255.0.
Part I
Train a fully-connected neural network to classify the images in the Fashion-MNIST data set.
You should try several different neural network architectures with different hyperparameters to try
to achieve the best accuracy you can on the validation data.
Some things you might try adjusting
are:
(a) The depth of the network (i.e. the number of layers)
(b) The width of the layers
(c) The learning rate
(d) The regularization parameters
(e) The activation functions
(f) The optimizer
(g) Anything else you can think of or find online
Can you beat 90% accuracy on the validation data? What about 95%? Once you have experimented
and found a network architecture and hyperparameter values that you think perform well, use that
architecture and those hyperparameter values to train one final model. Check the accuracy of your
final model on the test data and report your results.
You should also include the confusion matrix
for the test data. You should include in your report what things you tried while experimenting, but
you only need to include plots or results for the final model.
Part II
Repeat the same procedure as Part I, but now use a convolutional neural network. In addition
to the things listed in the previous part, you can now try adjusting:
(a) The number of filters for your convolutional layers
(b) The kernel sizes
(c) The strides
(d) The padding options
(e) The pool sizes for your pooling layers
Report your results and compare them to your fully-connected network from Part I.
