How limuloo/pydiff works
Project Type and Purpose
• This is a research‐oriented machine‐learning framework for “low‐light image enhancement” using Pyramid Diffusion Models (PyDiff).
• It builds on a foundational image‐restoration library (“BasicSR‐light”) and adds a diffusion‐based enhancement pipeline.
High-Level File-Structure Analysis
Top-level directories:
• BasicSR-light/ – core image‐restoration library
– archs/: network building blocks (EDSR, RRDB, StyleGAN2, SwinIR, etc.)
– data/: dataset definitions, samplers, degradations, transforms
– losses/, metrics/: loss functions & evaluation metrics (PSNR, SSIM, FID, LPIPS…)
– models/: base and task‐specific model classes
– ops/: custom CUDA/C++ operations (deformable conv, fused activations, upfirdn2d)
– scripts/: data‐prep, model conversion, metric calculators, training/inference wrappers
– train.py, test.py: entry points for training and evaluation
• PyDiff/ – new diffusion model layers & orchestration
– archs/: diffusion U‐Net, Pyramid diffusion architectures
– data/: LOL dataset loader
– models/: PyDiffModel class wrapping diffusion process
– train.py: distributed training launcher
– options/: config YAMLs for train/infer (v1, v2, v3, infer)
• dataset/, pretrained_models/: placeholders for input data and weights
• images/: figures (e.g., framework.png)
Key Points from README
• Two workflows: training (single/multi‐GPU distributed) and inference.
• Relies on two pipelined modules: BasicSR-light for infrastructure + PyDiff for diffusion.
Main Components
a. Configuration layer
• YAML files under BasicSR-light/options and PyDiff/options
• Hyperparameters, dataset paths, model settings
b. Data layer
• Raw “LOLdataset” folder → data‐preparation scripts (BasicSR-light/scripts/data_preparation + any custom PyDiff scripts)
• Data loader modules (BasicSR-light/basicsr/data & PyDiff/pydiff/data) apply samplers & transforms
c. Core library (BasicSR-light)
• Neural‐network primitives (archs/) + model wrappers (models/)
• Loss functions (losses/) & metrics (metrics/)
• Custom ops loaded as Python extensions (ops/)
d. Diffusion module (PyDiff)
• PyDiffModel orchestrates multi‐scale diffusion steps using U-Net variants (pydiff/archs/)
• Schedules and noise‐removal routines
e. Training pipeline
• pydiff/train.py (torch.distributed) → GPU cluster
• Interacts with BasicSR’s optimizer & scheduler utilities
• Checkpoints saved to pretrained_models/
f. Inference pipeline
• BasicSR-light/inference/*.py and pydiff/train.py -opt options/infer.yaml
• Generates enhanced images → evaluation metrics / visual outputs
g. Evaluation & Metrics
• Batch metric calculators in BasicSR-light/scripts/metrics
• Logging utilities record PSNR, SSIM, FID, LPIPS
Relationships & Data-Flow
• From Checkpoint + Config → Inference Script → Enhanced Output → Metric Calculator → Log/Report
• External dependencies: PyTorch/CUDA toolkits, distributed‐launcher, file‐system mounts for dataset & weights
Architectural Patterns & Design Principles
• Modular layering: separation of general restoration tools (BasicSR) vs. research innovation (PyDiff)
• Extension pattern: custom CUDA ops wrapped as Python modules
• Pipeline pattern: clearly delineated data‐prep → data‐load → model → training/inference
• Configurable workflow via YAML: supports reproducibility & experiment management