See: Presentation.pdf and Report.pdf for detailed information.
This work is focused at performing Neural Architecture Search (NAS) on a UNet neural network (SR3) to optimize for minimal FLOPs and inference latency; while working in a denoising diffusion generative framework (DDPM).
The task is image super-resolution on DIV2K dataset for 64x64->128x128 patches.
Broadly, this work mainly draws inspiration from the following papers,
- Image Super-Resolution via Iterative Refinement (UNet base model and forward diffusion)
- Denoising Diffusion Probabilistic Models (DDPM framework for reverce dissusion)
- AutoGAN-Distiller: Searching to Compress Generative Adversarial Networks (FLOPs and latency addtion to loss for performing NAS)
- DARTS: Differentiable Architecture Search