How aspiaspace/earthpt works
Project type and purpose
• EarthPT is a domain-specific, full-stack (code-only) machine learning project for time-series Earth Observation (EO) modeling.
• Its primary goal is to train a GPT-style large observation model on satellite time-series data (EarthPT-700M).
File structure
Top-level:
• config/ – Python configuration modules (e.g., earthpt700m.py)
• data/ – data preparation scripts (prepare_data.py)
• src/ – core code: configurator, model definition, sampling, training
• requirements.txt – dependencies
• README.md – project overview
• COPYING – license
Patterns & Roles:
• config/earthpt700m.py – hyperparameters & paths for a specific model
• data/prepare_data.py – raw EO ingestion & tokenization
• src/configurator.py – unified config loader and experiment manager
• src/model.py – GPT model + MLP tokenizer + loss
• src/sample_eo.py – inference / sample generation module
• src/train.py – training loop (+ logging & checkpointing)
README for additional insights
• The README confirms the core workflow: data preparation → model definition → training → evaluation.
• Dependencies: standard ML ecosystem (PyTorch, numpy, etc.).
• External integrations: HuggingFace for pretrained weights, ArXiv / CCAI for citations.