PreemptFire / README.md
github-actions[bot]
Sync from GitHub commit 9480952
08fbaac
|
Raw History Blame Contribute Delete
2.75 kB
---
title: PreemptFire Iberia Wildfire Risk and Model Comparison
emoji: "🌍"
colorFrom: red
colorTo: yellow
sdk: gradio
sdk_version: "4.44.1"
app_file: app.py
pinned: false
---
# PreemptFire: Iberia Wildfire Risk and Model Comparison
Gradio application for Hugging Face Spaces. Runs data collection following the `approach_b` branch and evaluates models trained on the **[PreemptFire-100m Kaggle Benchmark Dataset](https://www.kaggle.com/datasets/hollyfrogvi/wildfire-iberia-100m)**.
## Repository structure
- `app.py`: Gradio web interface (Iberia risk map, model comparison, live 2D scene collection, mini-dataset browser).
- `make_map.py`: CLI script to compute risk maps for the Iberia grid (3,146 locations).
- `live.py`: CLI script for single-coordinate evaluation and 22-channel 256x256 patch collection.
- `config_and_features.py`: feature calculation routines for TAB_1D (39 features), HIST_TAB (21 features), CLIM_FEATS (9 features), SEQ and STATIC_1D for nn_ts, FT (202 features), and 33-channel tensor construction for SimCLR.
- `data_collectors.py`: batched Open-Meteo ERA5 weather fetcher, NASA FIRMS VIIRS reader, and Scene2DCollector (Sentinel-2, Sentinel-1, DEM, OSM road travel time, soil moisture).
- `model_registry.py`: model loader and inference engine for tab_ssl_full, tab_full, tab_clim_full, tab_hist, tab_1d, nn_ts, coords, persist, clim, and ensembles.
- `mini_dataset_store.py`: persistence manager for collected .npz scenes, daily feature matrices, model predictions, and zip export.
- `prepare_hf_bundle.py`: pre-deployment helper to verify and stage model files and caches.
## Deployment to Hugging Face Spaces
1. Stage model weights and supporting data:
```bash
python prepare_hf_bundle.py \
--output-dir . \
--models-src /path/to/preemptfire_full_output \
--firms-gz /path/to/firms_points_2017_2025.csv.gz \
--scenes-dir /path/to/scenes_npz
```
If `--scenes-dir` is provided, the script extracts 10 km and 50 km population values into `population_cache.csv`, removing the need for Google Earth Engine credentials.
2. Create a Space on Hugging Face (SDK: Gradio, Hardware: CPU basic).
3. Push files using Git LFS:
```bash
git lfs install
git clone https://huggingface.co/spaces/<USER>/<SPACE_NAME>
cp -r hf_space/* <SPACE_NAME>/
cd <SPACE_NAME>
git lfs track "*.cbm" "*.pt" "*.pkl" "*.gz" "*.npz"
git add .gitattributes
git add .
git commit -m "Deploy PreemptFire app"
git push
```
4. Configure environment secrets in Space settings if needed:
- `FIRMS_MAP_KEY`: NASA FIRMS API key (only required for dates after 2025).
- `GEE_PROJECT`, `GEE_SERVICE_ACCOUNT`, `GEE_KEY_FILE`: optional fallback if not using `population_cache.csv`.