For these experiments, we use two sets of datasets: MNIST-USPS (https://www.kaggle.com/datasets/bistaumanga/usps-dataset), Clipart and Realworld categories of the Office-Home dataset (https://www.hemanthdv.org/officeHomeDataset.html). The former is the case where both the datastes contain images of handwritten digits and the latter contains images of 65 categories from four domains (of which you are supposed to work on only Clipart and Real-world categories). The following are the tasks:
A. Implement Adversarial Discriminative Domain Adaptation (ADDA) with Resent-50 as the base classifier on the source data. Use the Wasserstein metric for adversarial feature learning.
B. Implement a Cycle-GAN for the pair of MNIST-USPS datasets. Use the output of the converted target in the source classifier and report the result on adaptation.