Description
Winter 2025
Problem Description: Image Classification with Deep Neural Networks
Your goal in this assignment is to train Fully Connected Deep Neural Networks (FCNs / DNNs) to classify
images in FashionMNIST data set. FashionMNIST is similar to MNIST that you worked with in assignment
3. As in MNIST, each sample is a grayscale 28 × 28 image and the training set consists of 60K images and
the testing set consists of 10K images. Images depict articles of clothing that belong to a class from one of
10 classes. It is a fundamental benchmark dataset in machine learning as well. In comparison to MNIST,
FashionMNIST classification is more challenging. This warrants more powerful and nonlinear models. Here
you will implement FCNs toward this aim.
Figure 1 First 64 Images in FashionMNIST Dataset.
Some comments and hints
Here are some useful comments and facts to guide you along the way.
1. You are provided a code template (HW4_HelperTemplate.ipynb) for FCN for FashionMNIST. Your
goal will be to complete the template into a working model and decide on the architecture of your
model (i.e. number of layers and neurons in each layer). Keep in mind that the more layers you will be
adding the more weights you will have to train which will result into a higher computational complexity
which is not necessarily beneficial for classification accuracy and for your experiments. Once you have
completed this you will perform hyperparameter tuning and analyze the classification for this dataset.
2. In the provided template, the dataset is loaded with torchvision, which contains multiple datasets
such as MNIST, FashionMNIST, and has useful transformations (e.g. normalization) and loads data
as Pytorch tensors. The training data is split into Training and Validation datasets and loaded into
DataLoader as batches. DataLoader allows for easy iteration over the batches and samples within it.
Batches are useful for setting when weights updates will occur and also if you are using GPU to speed
up computation. You will need to define the batch size you want for your training and test sets.
3. The input data (features) are saved as 28 × 28 images. You will need to reshape images into a vector
of 784 which will be input into your network.
4. For implementation of different optimizers, regularization, initialization and normalization, refer to
example codes and pytorch library tutorials (https://pytorch.org/tutorials/) for examples.
Tasks
1. Design your FCN model to have an adjustable number of hidden layers and neurons in each layer with
relu activation. The first (input) layer should be of dimension 784 and the last (output) layer of 10
corresponding to 10 classes. Use Cross Entropy loss to perform the classification and SGD optimizer
with learning rate as an adjustable parameter. Set the number of epochs as an adjustable parameter
and train the network. Inspect the training loss curve and include validation of accuracy, on the
validation set, at each epoch. Perform testing at the end of training. In summary, your parameters
should allow you to vary the number of training batches, number of layers, number of neurons in each
layer, learning rate, number of training epochs. You can add more parameters as you see fit.
2. Find a configuration of the FCN (by varying the adjustable parameters) that trains in reasonable time
(several minutes is reasonable) and results in reasonable training loss curve and testing accuracy (at
least above 85% testing accuracy). These parameters will be your baseline configuration/parameters.
3. Perform hyperparameter tuning from the baseline, trying the following configurations. Report your
results in terms of tables, plots of loss curves and accuracy curves.
(a) Consider different optimizers including RMSProp, Adam and SGD with different learning rates.
You can find the corresponding functions in the torch.optim library. Log your training loss,
validation and test accuracy. Compare these optimizers, and observe which optimizer is most
suitable for your FCN. Try to explain why.
(b) Analyze the overfitting/underfitting situation of your model. Include Dropout regularization and
discuss whether this improves performance.
(c) Consider different Initializations, such Random Normal, Xavier Normal, Kaiming (He) Uniform.
Discuss how these initialization affect the training process and the accuracy.
(d) Include normalization such as Batch Normalization. Discuss the effect of the normalization on the
training process or testing accuracy.
4. Bonus (+2 points): Perform hyperparameter tuning of FCN to achieve testing accuracy of > 90% for
FashionMNIST and > 98% for MNIST. Report your configurations and try to explain them. Apply
the most successful classifier from assignment 3 to FashionMNIST (on full dimension of 784 features)
and compare with results above.
