Knowledge Discovery Homework 2

Welcome to the second homework of the Knowledge Discovery course! You have the opportunity to test what you have learned so far on generative models. The homework can be done in groups of two students.

Task Description

You will have about three weeks to create a generative model that is able to synthesize images (on any domain) at a minimum resolution of (64x64) and suitably conditioned through image labels. Once the model is trained, its evaluation should be carried out in terms of Inception Score and FID.

Moreover, the student is also asked to use the generated synthetic dataset as a form of data augmentation for training a classifier on the labels used for conditioning and assess whether it provides a contribution to classification performance or not.

As for HMW1, you’ll have to provide both code (github link to the repo) and summarization report of your work. The report (max 4 pages) should be organized as follows:

• Model description + figure.
• Dataset description (classes, number of images per class, images size and augmentation) including training, validation and test splits + sample figure.
• Training procedure (optimizer, hyperparams, loss, epochs, stopping criteria, batch size).
• Performance analysis (in terms of both generation and classification).

Submission Modality

To submit your homework you’ll have to create a GitHub repository (using your account) where you upload your notebok, dataset (zip folder with the images you’ve used) and a pdf report. The submission has to be carried by filling out the provided Google form.