Changelog

All notable changes to U-RNN are documented here.

[1.1.0] — 2026-03

Multi-dataset support: train on Futian (Shenzhen, 20 m/5 min, 400×560) and UKEA (UK, 8 m/5 min, 52×120) via LarFNO benchmark data
configs/location2_scratch.yaml: ready-made config for UrbanFlood24 location2 lightweight training (125×125, 200 epochs)
pyproject.toml: installable package (pip install -e .); adds urnn-train and urnn-test CLI entry points
Visualization upgrade: test.py now produces a 3-row Reference / U-RNN / Absolute Error comparison figure with fixed colorbars (0–2 m depth, 0–0.3 m error)
Ablation results: prewarming=false with 200 epochs achieves R²=0.989 on the lightweight dataset (vs 0.943 for prewarming=true at 300 epochs)
Metrics comparison tool: tools/compare_metrics.py — U-RNN vs LarFNO comparison with LaTeX table output
Dataset conversion tool: tools/convert_larfno_data.py — converts LarFNO Futian/UKEA data to U-RNN format
GitHub infrastructure: issue templates (bug report, feature request), CI workflow (import + config check across Python 3.8–3.10)
CONTRIBUTING.md: contribution guidelines and new-dataset workflow
requirements_tensorrt.txt: TensorRT-only dependencies split from core requirements.txt
Colab notebook: notebooks/quickstart.ipynb — end-to-end demo in < 5 minutes, no local GPU required
🤗 Hugging Face Hub: model page at holmescao/U-RNN (coming soon)

configs/lite.yaml: epochs corrected from 300 → 200 (200 epochs with prewarming=false achieves R²=0.989, outperforming 300-epoch baseline); added explicit location: location1 and event list files for reproducibility
configs/location2_scratch.yaml: input_height/input_width corrected from 128 → 125 (matches downsample_dataset.py --spatial_factor 4 output: 500÷4=125)
configs/ root cleaned up: ablation / fine-tuning / transfer-learning experiment YAMLs moved to configs/experiments/
README: restructured to 15 sections (removed duplicate Performance section and Scenario D); added Colab callout, HF Hub badge, pip install -e . note, and LarFNO dataset news
requirements.txt: removed TensorRT and Apex deps (now in requirements_tensorrt.txt); pinned openpyxl>=3.0.10
Default training recommendation: 200 epochs, prewarming=false (previously 300 epochs)

main.py scheduler_update(): added WarmUpCosineAnneal_v2 to the step-update branch — the v2 scheduler was created but .step() was never called, causing the learning rate to remain constant at the initial value throughout training
Dynamic2DFlood._prepare_input: spatial rainfall (T,H,W) now correctly handled (previously treated as scalar)
Dynamic2DFlood._prepare_target: flood (T,1,H,W) → (T,H,W) squeeze
test.py compute_metrics: added torch→numpy conversion guard
net_params.py: decoder GRU input_channels formula corrected
model.py: reg_output_t indexing fixed ([:, :, 0] not [:, :, 0:1])
main.py WarmUpCosineAnneal: fixed warmup/cosine formula (removed spurious /0.1 division)