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.