README.md

metadata

title: ExoPrompt Greenhouse Climate Demo
emoji: 🌱
colorFrom: green
colorTo: blue
sdk: docker
app_port: 8501
tags:
  - streamlit
  - time-series-forecasting
  - greenhouse
  - exoprompt
pinned: false
license: mit
short_description: ExoPrompt paper demo — greenhouse climate forecasting

ExoPrompt — Greenhouse Climate Forecasting Demo

📄 Paper  ·  💻 Original paper repo  ·  🛠️ Demo source  ·  🤗 Checkpoints  ·  🚀 Live demo

Interactive Streamlit demo for ExoPrompt: transformer-based greenhouse climate forecasting with structured exogenous prompts and physics-based simulation.

Two models are exposed:

• ExoPrompt Transformer — Transformer backbone conditioned on 254 structural, environmental, and crop parameters via learnable soft prompts.
• Vanilla Transformer — same Transformer backbone, no exogenous conditioning (baseline for comparison).

Both models are loaded from the published paper checkpoints (200k-sample HPS pretraining).

Local development

# install deps
uv sync

# run streamlit locally
uv run streamlit run app.py

Checkpoints

Checkpoints are hosted in a sibling Hugging Face model repo and downloaded on first launch via huggingface_hub.hf_hub_download (cached on the Space's persistent disk).

Project structure

exoprompt-inference/
├── app.py                              ← Streamlit entry (loaded by HF Spaces)
├── pyproject.toml
└── src/exoprompt_inference/
    ├── data/                           ← inference dataset + loaders
    ├── inference/                      ← model loader + Predictor
    ├── streamlit/                      ← Streamlit app + UI components
    └── _vendor/                        ← carved-out training code
        ├── models/                     ← lit module, abstract base, physi-net wrapper
        ├── utils/                      ← scaler, time features, pickle helper
        └── time_series_library/        ← Transformer backbone (only)

Citation

@article{SOYKAN2026111673,
  title   = {ExoPrompt: Transformer-based greenhouse climate forecasting with structured conditioning and physics-based simulation},
  journal = {Computers and Electronics in Agriculture},
  volume  = {246},
  pages   = {111673},
  year    = {2026},
  author  = {Soykan, G{\"u}rkan and Babur, {\"O}nder and Liu, Qingzhi and Tekinerdogan, Bedir}
}