How novwaul/sr3 works
Project type and purpose
• This is a research-oriented Python project implementing the SR3 super-resolution diffusion model (4× upsampling) from “Image Super-Resolution via Iterative Refinement” (arXiv:2104.07636).
• It provides data loading, UNet-based diffusion training, sampling routines, and evaluation metrics.
File structure
Top-level files:
• main.py — CLI entrypoint for training or testing
• train.py — high-level training loop
• diffusion.py — defines forward/backward diffusion processes
• data.py — dataset loading, augmentation, batching
• metrics.py — IS/FID/PSNR/SSIM computation
• utils.py — helper functions (logging, checkpoint I/O, config parsing)
• LICENSE, README.md
UNet/
• model.py — definition of the SR3 UNet architecture with attention and group-norm tweaks
README for additional insights
• Specifies PyTorch-style hyperparameter settings and CLI flags in main.py
• Describes architectural differences vs. vanilla DDPM (attention at two depths, group-norm groups=8, embedding scale change, gamma sampling)
• Outlines train/test commands and sample/metric evaluation
a. Components:
1. User CLI (main.py)
2. Data Pipeline (data.py)
3. Diffusion Module (diffusion.py)
4. Model Component (UNet/model.py)
5. Training Loop (train.py)
6. Sampling Loop (main.py + diffusion.py + UNet)
7. Metrics & Logging (metrics.py + utils.py)
8. Checkpoint & Config Management (utils.py)
b. Relationships & Data Flow:
• CLI reads config → invokes Training or Testing
• Data Pipeline → produces (LR, HR) batches → feeds Diffusion Module
• Diffusion Module calls UNet at each time step to predict noise or denoised output
• Training Loop updates UNet weights via optimizer & schedules (lr, β schedule)
• Checkpoints saved/loaded via Utils
• Sampling Loop uses learned UNet + reverse diffusion to generate SR images
• Metrics Module computes FID/IS/PSNR/SSIM on outputs → logs via Utils
c. Architectural patterns & principles:
• Modular separation of concerns: data / model / diffusion logic / training / metrics
• Explicit scheduling components (learning rate, beta, gamma)
• Pluggable UNet architecture for easy experimentation
d. Key technologies & frameworks:
• Python 3.x
• PyTorch (tensors, nn.Module, DataLoader, optim)
• NumPy, PIL/OpenCV (data augmentation)
• Logging (TensorBoard or custom via utils)