README.md

---
license: mit
library_name: pytorch
tags:
  - time-series-forecasting
  - greenhouse
  - transformer
  - exoprompt
  - physics-informed
  - lightning
language:
  - en
pipeline_tag: time-series-forecasting
---

# ExoPrompt — Greenhouse Climate Checkpoints

PyTorch Lightning checkpoints for the **ExoPrompt** paper:

> **ExoPrompt: Transformer-based greenhouse climate forecasting with structured
> conditioning and physics-based simulation.**
> Gürkan Soykan, Önder Babur, Qingzhi Liu, Bedir Tekinerdogan.
> *Computers and Electronics in Agriculture*, vol. 246, p. 111673, 2026.
> DOI: [10.1016/j.compag.2026.111673](https://doi.org/10.1016/j.compag.2026.111673)

These are the pretraining checkpoints used to produce the paper's quantitative
results on the 200k-sample GreenLight subsets. Each checkpoint corresponds to the
best validation epoch of a Transformer backbone trained on one lighting scenario,
in either the vanilla or ExoPrompt-conditioned configuration.

## Files

| Filename | Variant | Lighting scenario | Architecture |
|---|---|---|---|
| `200k_exo_hps.ckpt` | ExoPrompt | HPS | Transformer + exo-prompt projector (254-d) |
| `200k_exo_led.ckpt` | ExoPrompt | LED | Transformer + exo-prompt projector (254-d) |
| `200k_exo_mixed.ckpt` | ExoPrompt | Mixed (HPS + LED) | Transformer + exo-prompt projector (254-d) |
| `200k_vanilla_hps.ckpt` | Vanilla | HPS | Transformer (no exogenous conditioning) |
| `200k_vanilla_led.ckpt` | Vanilla | LED | Transformer (no exogenous conditioning) |
| `200k_vanilla_mixed.ckpt` | Vanilla | Mixed (HPS + LED) | Transformer (no exogenous conditioning) |

All six share the same Transformer backbone hyperparameters and the same
forecasting head (`pred_len = 96`, `label_len = 48`, 3-feature indoor-climate
output: `tAir`, `vpAir`, `co2Air`).

## Usage

These checkpoints are designed to load with the
[`exoprompt-inference`](https://github.com/gsoykan/exoprompt-inference) Streamlit
demo, or directly via the
[official ExoPrompt repo](https://github.com/gsoykan/ExoPrompt). Minimal example:

```python
from huggingface_hub import hf_hub_download
from exoprompt_inference.inference.model_loader import (
    load_model_from_checkpoint,
    ModelType,
)

ckpt_path = hf_hub_download(
    repo_id="gsoykan/exoprompt-checkpoints",
    filename="200k_exo_hps.ckpt",
)
model = load_model_from_checkpoint(
    ckpt_path,
    model_type=ModelType.TIME_SERIES_LIB_MODEL,
    device="cpu",
)
```

## License

[MIT](https://opensource.org/licenses/MIT) — same as the source code in the
[official ExoPrompt repo](https://github.com/gsoykan/ExoPrompt).

## Citation

```bibtex
@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}
}
```