CloudAXV/nowcastnet-runs

NowcastNet North China Experiments

This dataset contains training results from multiple NowcastNet experiments for precipitation nowcasting over North China (3-hour lead time).

Dataset URL: https://huggingface.co/datasets/CloudAXV/nowcastnet-runs

Repository Structure

Each experiment directory contains:

experiment_name/
├── checkpoints/     # Model weights (.ckpt files)
├── reports/         # Evaluation metrics and visualizations
├── logs/            # Training and test logs
└── results.tar      # Prediction outputs (tar archive of PNG visualizations)

Experiments

Experiment	Description
north_china_3h	Base radar and PWV v2 models (3h)
north_china_3h_physical	Physical PWV v3 model (3h)
north_china_3h_physical_pwv_v3_leadtime	PWV v3 with leadtime comparison
north_china_3h_physical_pwv_v3_tendency	PWV v3 with tendency modeling
north_china_3h_physical_pwv_v3_tendency_facl	PWV v3 tendency + FACL loss
north_china_3h_physical_pwv_v4_attn	PWV v4 with attention mechanism
north_china_3h_physical_pwv_v4_tendency	PWV v4 with tendency modeling
north_china_3h_physical_pwv_v4_tendency_facl	PWV v4 tendency + FACL loss
north_china_3h_v3_object	PWV v3 with object-based evaluation
north_china_3h_v3_object_consistency	PWV v3 object consistency experiments
north_china_3h_v3_object_consistency_tendency_facl	Object consistency + tendency + FACL
recomputed_reports	Re-computed evaluation reports across experiments

Accessing Results

Results are stored as tar archives. To extract:

tar -xf results.tar

Checkpoint Files

Each checkpoints/ directory contains:

• best.ckpt / best_state_dict.ckpt — Best model weights during training
• latest.ckpt / latest_state_dict.ckpt — Latest model weights
• *_model.ckpt — Single-file model checkpoint
• train_args.json — Training configuration
• train_log.csv — Training metrics log

Evaluation Reports

Each reports/ directory contains:

• horizon_mae.png / horizon_rmse.png — Error metrics by forecast horizon
• lead_mae.png / lead_rmse.png — Error metrics by lead time
• psd_*.png — Power spectral density comparisons
• sample_*.png — Sample prediction visualizations
• threshold_metrics.png — Threshold-based skill scores
• summary_*.json — Numerical evaluation summary