Views
No views yet
Statut : baseline méthodologique, contre-performante vs Lot B (statistical lapse + QDM + RBF Sencrop). Conservée pour traçabilité et starting point d'expériences futures.
base_ch=32, n_levels=3, use_film=True| Métrique | Lot B (statistical) | Lot C (this model) |
|---|---|---|
| Bias moyen | -0,06 °C | +2,63 °C |
| RMSE | 1,40 °C | 5,87 °C |
| POD global | 44 % | 9 % |
| FAR global | 34 % | 81 % |
| CSI global | 0,36 | 0,07 |
2022-best.ckpt (Lightning, ~5 MB)2023-best.ckpt2024-best.ckpt2025-best.ckpt1pip install 'git+https://github.com/maurinl26/downscaling@main'
2
3# Inference on year 2023 (S3 output)
4uv run python -m downscaling.scripts.recalibrate_dl_film \
5 --year 2023 \
6 --cerra-atm cerra_atm_2023.nc \
7 --dem dem_attributes.nc \
8 --sencrop /path/to/sencrop_bulk \
9 --out s3://your-bucket/recalibrated/dl_film \
10 --device auto --epochs 20 --early-stopping-patience 5karpos26/lot-c-dl-film-baseline-v1/analyses/ and the EPIC C5 annex.1@misc{karpos_lot_c_baseline_2026,
2 author = {Maurin, Loïc},
3 title = {Karpos Lot C DL FiLM Baseline v1},
4 year = {2026},
5 publisher = {HuggingFace},
6 url = {https://huggingface.co/karpos26/lot-c-dl-film-baseline-v1},
7 note = {Baseline méthodologique, à dépasser via multi-year + gel-aware loss}
8}