This repository is part of my coursework for Deep Learning (ML2), a university course covering modern deep learning. Topics include neural-network training and optimization, convolutional architectures, sequence models (RNN/LSTM/Transformers), generative models (VAEs/GANs/diffusion), adversarial robustness, and self-supervised/contrastive learning.
HW4 is the final project of the course: train a deep generative model from scratch and actually get it to sample convincing images. The chosen model is a DCGAN — a generator built from ConvTranspose2d / BatchNorm2d / ReLU blocks ending in Tanh, and a matching convolutional discriminator. The generator is trained on the 102-category Oxford Flowers dataset at 64×64, with the report covering training stability, mode-collapse symptoms, and a latent-space walk. The generation script loads the trained weights and produces image grids for the write-up.
hw4_code.py— training and evaluation codehw4_generation.py—Generatornetwork definition and sampling utilitieshw4_model.pkl— trained generator weightshw4_report.pdf— write-up of the DCGAN design, training curves and sampleshw4_code_IPYNB.pdf— PDF export of the code notebookHW4.pdf— the assignment sheetrequirements.txt— Python dependencies used during training
Suheil Khourieh