TDDE70 Deep Learning Coursework

This repository contains my solutions for the TDDE70 Deep Learning course (Linköping University, Spring 2024), including an intro notebook and four labs.

---

Contents

• Intro: PyTorch basics: tensors, GPU, autograd, linear regression
• Lab 0: PyTorch & NN fundamentals
• Lab 1: Autoencoders & U‑Net
• Lab 2: Denoising Diffusion Probabilistic Models (DDPM)
• Lab 3: Graph Neural Networks (CGCNN)

---

Intro: PyTorch Basics

Get started with:

• Tensors & GPU support
• Autograd & computational graphs
• Building & training a linear regression model

---

Lab 0: PyTorch & Neural Network Fundamentals

• Custom Modules: Define nn.Module and fully‑connected layers
• Data Loading: Convert MNIST to tensors, use DataLoader
• Simple CNN: Conv layers with batch‑norm & dropout
• Training & Eval: Optimizers, training loops, accuracy metrics
• Robustness: Test on rotated MNIST digits

---

Lab 1: Autoencoders & U‑Net for Image‑to‑Image Tasks

• Data Prep: Custom Dataset for denoising & segmentation (GTAV)
• Model Design: DoubleConv, Down, Up, UpSkip blocks; Autoencoder & U‑Net
• Training: Denoising (MSE) & segmentation (weighted CE) with Trainer classes
• Enhancements: Skip connections & EMA weight averaging

---

Lab 2: Diffusion Models

Task 1: 2D Two‑Moon Generation

• Theory: DDPM forward/backward processes, noise schedule
• Implementation: MLP denoiser with positional embeddings & noise utilities
• Training & Sampling: T=50 steps, L₂ loss, visualize samples

Task 2: Class‑Conditional MNIST Generation

• Architecture: U‑Net with timestep & label embeddings, self‑attention
• cDDPM: T=1000 diffusion steps on 32×32 MNIST
• Sampling: Generate digits 0–9 conditioned on labels

---

Lab 3: Graph Neural Networks (CGCNN)

• PyG Basics: Data & DataLoader for graphs
• MPNN Equations: Derive CGCNN’s message & update functions
• CGCNNLayer: Gated message passing with BatchNorm
• Full Model: Stack layers, global mean pooling, MLP head
• Training: Compare invariant (distance) vs non‑invariant (vector) edge features