Repo for the project of Jinghua, Junhao, and Kaiwen about using diffusion models for NN parameter generation to accelerate neural architecture design for edge AI devices. Some starter code implementations are taken from https://github.com/NUS-HPC-AI-Lab/Neural-Network-Parameter-Diffusion and https://github.com/facebookresearch/DiT.
Green ResNet NN architectures are seen during diffusion model training, while red ResNet NN architectures are not seen during diffusion model training.
Gray bars: NN parameters randomly drawn from uniform distribution; blue bars: NN parameters generated by a trained diffusion model; olive bars: NN parameters obtained through 50 epochs training on CIFAR-10 training dataset.