This repository contains exercises and implementations from the Deep Learning course at Heinrich Heine University (HHU), taught in the Winter Semester 2024/25. The course progresses from foundational topics like automatic differentiation to advanced deep learning systems such as transformers and challenge-driven applications.
The course consists of 13 programming assignments, each with focused objectives ranging from low-level tensor operations to high-level sequence modeling. Each folder includes the necessary notebooks, data, and helper code to reproduce the solutions.
Exercises
├── Assignment 01 → Autodiff puzzles and Jacobians
├── Assignment 02 → Mini-batch MLP regressor (GPA → IQ)
├── Assignment 03 → GeLU, Leaky-ReLU, einsum with backprop
├── Assignment 04 → Random/Grid CV with Fashion-MNIST
├── Assignment 05 → Custom CNN layers (Conv2D, ConvT)
├── Assignment 06 → Inception modules (GoogLeNet-style)
├── Assignment 07 → Normalization, Focal Loss
├── Assignment 08 → ResNet with stochastic depth & augmentation
├── Assignment 09 → CIFAR-10 competition challenge
├── Assignment 10 → Character-level next-token prediction
├── Assignment 11 → BPE tokenizer implementation
├── Assignment 12 → Transformer for hate speech detection
└── Assignment 13 → GPTrump: leaderboard challenge submission
pip install -r requirements.txt| Assignment | Metric | Achieved |
|---|---|---|
| 07 – Normalization | Fashion‑MNIST val acc | 86.8 % |
| 08 – ResNet + Aug | Fashion‑MNIST val acc | 76.5 % |
| 09 – CIFAR‑10 Challenge | Test accuracy | 67.3 % |
| 12 – Hate Speech | Test accuracy | 79.6 % |
| 13 – GPTrump | Final perplexity | 2.48 |