How mengxiangming/dmps works
Project type and purpose
• This is a research‐oriented Python toolkit implementing “Diffusion Model Based Posterior Sampling (DMPS)” for noisy linear inverse problems.
• It provides a command-line interface (main.py + shell scripts) to load a pretrained diffusion model, configure a task (denoising, deblurring, colorization, super-resolution), perform posterior sampling, and save results.
File structure
• Top-level directories:
• configs/ – YAML files describing model hyperparameters, diffusion settings, and task-specific measurement operators
• data/ – data loader and sample inputs for demo
• guided_diffusion/ – core diffusion model implementation (UNet, Gaussian diffusion process, sampling scheduler)
• util/ – helper modules for image I/O, preprocessing, metrics, logging
• scripts/ – shell wrappers that call main.py with predefined configs
• Key files:
• main.py – orchestrates end-to-end workflow: parse args, load configs, initialize model & data, run sampler
• requirements.txt – Python dependencies (PyTorch, mpi4py, scikit-image, etc.)
• Deployment artifacts: pretrained checkpoints in models/ (not in tree), outputs saved under user-specified save_dir.
Insights from README
• Stages: environment setup → download pretrained model → prepare data → run sampling
• Supported tasks driven entirely by config files: model_config*.yaml + diffusion_config.yaml + one task_config.yaml
• External dependencies: GPU/CUDA, mpi4py for parallel sampling, blobfile for data I/O.
Main components & interactions
A. Configuration Manager
• Loads YAML files (model, diffusion, task)
• Passes hyperparameters to downstream modules
B. Data Loader
• Reads sample images / dataset directory
• Applies preprocessing (util/data_preprocessing, resizer)
C. Diffusion Model Module (guided_diffusion/)
• UNet backbone (unet.py)
• Gaussian diffusion scheduler (gaussian_diffusion.py, posterior_mean_variance.py)
• Conditioning methods (condition_methods.py) for measurement operator A
• Utility functions (fp16_util, dist_util)
D. Posterior Sampler (main pipeline in main.py)
• At each diffusion timestep, compute score with model
• Incorporate measurement likelihood via pseudo-likelihood term
• Iterate sampling loop → produce final x samples
E. Utilities & Tools
• Logging (util/logger.py)
• Metrics computation (util/compute_metric.py)
• Image I/O and visualization (util/img_utils.py, figures/)
F. Shell Scripts & CLI
• scripts/run_*.sh launch main.py with predefined configs
G. External Services
• PyTorch + CUDA for GPU acceleration
• mpi4py for distributed sampling (optional)
• File system for loading/saving checkpoints, data, results
Architectural patterns & design principles
• Modular separation: configs ↔ core model ↔ data ↔ utilities
• Plugin style for new tasks: drop in new task_config.yaml
• Single orchestration script (main.py) using dependency injection from configs
• Lightweight wrapper scripts for reproducibility
• Clear separation between pure ML code (guided_diffusion) and infrastructure (data, util, logging)
Final checklist
• All major Python packages/modules represented
• Clear boundary between user interface (CLI/scripts) and core library
• Configurable pipeline stages visibly annotated