How ignf/flair-hub works
FLAIR-HUB provides patch-level multimodal, multitask semantic-segmentation training, prediction, and metrics, plus a separate large-area zonal inference entry point. The main workflow loads configured datasets, constructs a data module, optionally trains, then predicts or computes metrics and writes outputs. README documentation also describes multimodal inputs, pre-trained model initialization, and zonal rasters. Source excerpts establish the principal entry-point and model/data behavior; unsampled subsystems are represented only where documented, without speculative wiring.
Main entry points: main.py Patch CLI · inference.py Area inference · stages.py Training stage · flair_model.py Multimodal fusion
Components 24
Patch workflow
- Patch CLI
src/flair_hub/main.py - Config setup
src/flair_hub/utils/config_io.py - Dataset paths
src/flair_hub/data/utils_data/paths.py - Data module
src/flair_hub/tasks/module_setup.py - Patch loader
src/flair_hub/data/dataloader.py - Raster prep
src/flair_hub/data/utils_data - Training stage
src/flair_hub/tasks/stages.py - Task module
src/flair_hub/tasks/tasks_module.py - Trainer setup
src/flair_hub/tasks/trainers.py - Checkpointing
src/flair_hub/models/checkpoint.py - Prediction stage
src/flair_hub/tasks/stages.py - Prediction writer
src/flair_hub/writer/prediction_writer.py - Metrics
src/flair_hub/writer/metrics_core.py
Segmentation models
- Multimodal fusion
src/flair_hub/models/flair_model.py - Image encoders
src/flair_hub/models/monotemp_model.py - Time-series encoder
src/flair_hub/models/multitemp_model.py
Zonal inference
- Zonal CLI
src/flair_zonal_detection/main.py - Area inference
src/flair_zonal_detection/inference.py - Area tiling
src/flair_zonal_detection/dataset.py - Tile slicing
src/flair_zonal_detection/slicing.py - Model utilities
src/flair_zonal_detection/model_utils.py - Raster postprocess
src/flair_zonal_detection/postprocess.py - GeoTIFF outputs
Other components
- Researcher
Connections 24
- Researcher → Patch CLI: runs config
- Patch CLI → Config setup: loads settings
- Patch CLI → Dataset paths: resolves datasets
- Patch CLI → Data module: builds loaders
- Data module → Patch loader: creates batches
- Patch loader → Raster prep: preprocesses rasters
- Patch CLI → Training stage: starts training
- Training stage → Trainer setup: configures run
- Training stage → Task module: trains tasks
- Task module → Multimodal fusion: forwards batches
- Multimodal fusion → Image encoders: encodes imagery
- Multimodal fusion → Time-series encoder: encodes time series
- Training stage → Checkpointing: saves weights
- Patch CLI → Prediction stage: starts prediction
- Prediction stage → Task module: runs inference
- Prediction stage → Prediction writer: writes predictions
- Prediction writer → Metrics: computes metrics
- Researcher → Zonal CLI: runs config
- Zonal CLI → Area inference: dispatches
- Area inference → Area tiling: loads area tiles
- Area inference → Tile slicing: slices outputs
- Area inference → Model utilities: loads model
- Area inference → Raster postprocess: processes rasters
- Area inference → GeoTIFF outputs: writes GeoTIFFs