ARCHITECTURE.md

LoveDA Benchmarking Project Architecture

Overview

This document explains the project structure and component relationships. We use a modular architecture following PyTorch Lightning conventions.

graph TD
    CLI[CLI Interface] --> Data[Data Module]
    CLI --> Models[Model Registry]
    CLI --> Lightning[Lightning Module]
    Data -->|Provides| Lightning
    Models -->|Provides| Lightning
    Lightning -->|Generates| Metrics[Metrics & Reports]
    Metrics -->|Updates| README[README.md]

Loading

Key Components

1. Command Line Interface (cli.py)

• Purpose: Entry point for benchmark execution
• Key Features:
  • Argument parsing for model selection, data paths, and hyperparameters
  • Coordinates data loading, model initialization, and training workflow
  • Handles output generation and logging
• Implementation Choices:
  • Uses Click library for robust CLI handling
  • Separates configuration from core logic for flexibility

2. Data Module (data.py)

• Responsibility: LoveDA dataset handling
• Key Functions:
  • load_data(): TorchGeo dataset integration
  • get_transform(): Image preprocessing pipeline
  • get_dataloader(): Batch data loading
• Design Notes:
  • Custom transforms for satellite imagery
  • Automatic dataset validation checks

3. Model Registry (models.py)

• Purpose: Model initialization and configuration
• Contents:
  • Pretrained segmentation model definitions
  • Backbone selection logic
  • Custom head implementations
• Supported Models:
  • DeepLabV3
  • FCN-ResNet50
  • (Extensible to new architectures)

4. Lightning Module (lightning_module.py)

• Core Logic: Training/validation workflow
• Key Aspects:
  • Metric tracking (mIoU, accuracy, inference speed)
  • Memory usage monitoring
  • Checkpoint management
  • Automated README updates
• Integration:
  • Connects data loaders and models
  • Implements benchmark protocols

Support Files

setup.py

• Purpose: Package configuration
• Handles:
  • Dependency management
  • CLI command registration
  • Python package metadata

.egg-info

• Note: Auto-generated package metadata (do not edit manually)
• Contains:
  • Installed files list
  • Dependency information
  • Entry points configuration

Workflow Sequence

1. User runs CLI command with parameters
2. System loads data through TorchGeo integration
3. Initializes selected model architecture
4. Executes Lightning training/validation loop
5. Records metrics and updates documentation
6. Generates final benchmark report