How lukashaas/pigeon works
Main Components
a) Data Processing Pipeline:
• Input Layer: Raw images
• Preprocessing Module (from /preprocessing directory)
• Dataset preprocessing
• Geo augmentation
• Embedding generation
b) Core Model Architecture:
• CLIP Embedder (models/clip_embedder.py)
• Super Guessr Model (models/super_guessr.py)
• Proto Refiner (models/proto_refiner.py)
• Custom Layers
• HEDGE layer
• Positional Encoder
c) Dataset Management:
• Multiple dataset types:
• Pretraining datasets
• Finetuning datasets
• Benchmark datasets
• Geocell System (from dataset_creation/geocell)
d) Training System:
• Training loops
• Evaluation metrics
• Model saving/loading
Specific Design Elements
Use these visual elements:
• Rounded rectangles: Processing modules
• Hexagons: Model components
• Blue: Data processing components
• Green: Model components
• Yellow: Training components
• Gray: Storage/Data
• Red: Evaluation components
Important Details
• Configuration system (config.py) influence on all components
• Multiple dataset paths for different purposes
• Model component interactions
• Training feedback loops
• Evaluation metrics flow