How khyeongkyun/zeroflood works
ZeroFlood supports flood-segmentation dataset preparation and a Pangaea-based fine-tuning, evaluation, and inference workflow using geospatial foundation encoders and task decoders. Dataset creation downloads TerraMesh/LISFLOOD inputs and preprocesses them; fine-tuning consumes the resulting data through dataset/preprocessing abstractions, then trains and evaluates a configured model. The README documents prediction-saving inference and multiple encoder families. Some subsystem links reflect README workflows or SOURCE INDEX references; unsampled implementations and their internal wiring are not inferred.
Main entry points: run.py Run Entry · download.py Source Downloader · encoders Foundation Encoders · preprocess.py Dataset Preprocessor
Components 20
Flood Data
- Source Downloader
data/download.py - Dataset Preprocessor
data/preprocess.py - TerraMesh Utilities
data/terramesh.py
Benchmark Runtime
- Raw Dataset Adapters
pangaea/datasets - Dataset Wrapper
pangaea/datasets/base.py - Input Preprocessing
pangaea/engine/data_preprocessor.py - Run Entry
pangaea/run.py - Training Loop
pangaea/engine/trainer.py - Evaluation
pangaea/engine/evaluator.py - Saved Predictions
pangaea/engine/evaluator.py
Model Components
- Foundation Encoders
pangaea/encoders - Task Decoders
pangaea/decoders - TerraMind Model
pangaea/encoders/terramind/model/tm_utils.py - TerraMind Tokenizers
pangaea/encoders/terramind/tokenizer/vqvae.py - Diffusion Pipeline
pangaea/encoders/terramind/tokenizer/scheduling/diffusion_pipeline.py
Training Support
- Experiment Logging
pangaea/utils/logger.py - Label Subsampling
pangaea/utils/subset_sampler.py
Other components
- Researcher
- Source Datasets
- ZeroFlood HF Dataset
Connections 20
- Researcher → Source Downloader: requests download
README.md - Source Datasets → Source Downloader: provides inputs
README.md - Source Downloader → Dataset Preprocessor: downloads data for
README.md - Dataset Preprocessor → ZeroFlood HF Dataset: reproduces dataset
README.md - Researcher → Run Entry: runs benchmark
README.md - ZeroFlood HF Dataset → Raw Dataset Adapters: supplies fine-tuning data
- Run Entry → Dataset Wrapper: constructs datasets
pangaea/run.py - Dataset Wrapper → Input Preprocessing: uses preprocessing
pangaea/datasets/base.py - Run Entry → Foundation Encoders: instantiates encoder
pangaea/run.py - Run Entry → Task Decoders: instantiates decoder
pangaea/run.py - Task Decoders → Foundation Encoders: holds encoder
pangaea/decoders/base.py - Run Entry → Training Loop: runs fine-tuning
pangaea/run.py - Run Entry → Evaluation: runs evaluation
pangaea/run.py - Run Entry → Label Subsampling: selects label subsets
pangaea/run.py - Training Loop → Experiment Logging: uses logging
pangaea/engine/trainer.py - Evaluation → Saved Predictions: saves predictions
README.md - Researcher → Run Entry: runs inference
README.md - Foundation Encoders → TerraMind Model: includes TerraMind
pangaea/encoders/terramind/model/encoder_embeddings.py - TerraMind Model → TerraMind Tokenizers: uses tokenizer components
- TerraMind Tokenizers → Diffusion Pipeline: supports generation