Skip to content

Changelog

v0.1.1 (Unreleased)

New Features:

  • Unified data resolution: all standard datasets open through the hydrodataset/hydrodatasource resolver (open_dataset / resolve_data_path / open_dataset_from_source_cfgs). source_cfgs now accepts dataset_id; paths are resolved centrally from hydro_setting.yml.
  • Data sources: support for caravan, grdc_caravan, camels_aus, camels_br, camels_cl, camels_gb, camels_col; self-made readers (selfmadehydrodataset, selfmadeforecastdataset, longtermdataset, floodeventdatasource, stationhydrodataset, tghydrodatasource) come from hydrodatasource.
  • Cloud-Zarr lazy loading: CloudZarrLazyDataset + CloudZarrChunkBatchSampler for joint training on remote Zarr stores; dask-backed cloud eager-load config keys (cloud_eager_load, combine_relevant_target_read, fast_dapeng_scaler, ...).
  • Monthly time units (ME/MS/M) in addition to hourly/daily.
  • New models: WDNE adaptive-routing flood net, Diffusion, FNO, MoE family (Dense/Sparse/Switch/HydrologyTopK/GlobalTopK), MTSLSTM, DPLMC reservoir model.
  • New datasets: TgHydroDataset, MultiSourceDataset, MultiSourceAEFDataset, FNODataset, CloudZarrLazyDataset.
  • New trainers/components: streaming_evaluation, performance_monitor, cuda_prefetcher, dplmc_rsvr_trainer.

Improvement:

  • Split-range validation is now warn-only: adjacent train/valid/test periods (left-closed, right-open) do not overlap; strict overlap emits an advisory warning instead of raising (allow_split_overlap suppresses it).
  • Rolling-forecast result recovery corrected (_recover_samples_to_4d_by_basins / _by_forecast, _rolling_preds_for_once_eval); by-basins evaluation places each basin's last forecast step correctly and accumulates across basins.
  • data_sources.py reduced to a thin registry layer; torchhydro no longer re-implements GAGES-II / MOPEX / reservoir readers (moved to hydrodatasource).
  • Removed duplicate definitions in fno.py (bad-merge artifacts); full unused-import / undefined-name sweep (flake8 F401/F821/F405/F811/F841).
  • Dependency upgrades: hydrodataset>=0.3.0, hydrodatasource>=0.3.0, hydroutils>=0.2.0; added torch-geometric-temporal, diffusers, accelerate, mpi4py, h5py; PyG prebuilt wheels via uv index.

Bug Fixes:

  • DapengScaler reads both mean_prcp and p_mean variable names.
  • Legacy static-variable names migrated to standard CAMELS names (DRAIN_SQKM→area, PPTAVG_BASIN→p_mean, SLOPE_PCT→slope_mean).
  • resulter.py handles missing shap gracefully and uses hydroutils streamflow unit conversion.
  • Mixed-precision AMP scaler is owned by the train_utils worker (not deep_hydro).

v0.1.0

Previous stable release. Data sources were still provided in-repo; the unified resolver migration, cloud-Zarr lazy loading, and the new model families above were introduced on dev and are tracked in the v0.1.1 entry.