Back to Projects

Personal — 2024

Gradient-Based Image Synthesis via Score Matching

Jupyter PyTorch Deep Learning
ML Homework 4: Model Training and Evaluation

Overview

Implemented a Noise-Conditional Score Network (NCSN) for generative modeling, following Song & Ermon (2019).

Trained a U-Net to approximate the score function of MNIST data with denoising score matching, and generated images via annealed Langevin dynamics.

Resources