swissrivernetwork.benchmark.dataset¶
swissrivernetwork.benchmark.dataset ¶
Dataset primitives for the Swiss River Network Benchmark.
This module exposes:
- :func:
read_graph/ :func:read_csv_train/ :func:read_csv_test— low-level graph + CSV readers for the three datasets (swiss-1990,swiss-2010,zurich). - :func:
select_isolated_station— pure column-select used by isolated LSTM / Transformer training; does not drop rows. - :class:
SequenceDataset— sliding-window dataset over per-station time series.short_subsequence_method="drop"silently drops runs shorter thanwindow_len(this is why isolatedwt_hatdumps at large eval window lengths can cover fewer days than the raw test CSV — a fact that motivates the inner-join in :func:util.merge_graphlet_dfs). - :class:
SequenceWindowedDataset— graphlet-style dataset that stacks neighbor features alongside the target station.
The module is import-cheap (no torch CUDA calls at import time) so it can be imported by CPU-only tooling such as the CLI or the documentation builder.
SequenceWindowedDataset ¶
Bases: SequenceDataset
SequenceFullDataset ¶
Bases: SequenceDataset
Returns the full available sequence (no windowing)
read_csv_prediction_train ¶
read_csv_prediction_train(
graph_name: str,
method: str,
station,
predict_dump_dir: str | Path | None = None,
base_dir: str | Path = PROJ_DIR,
)
read_csv_prediction_test ¶
read_csv_prediction_test(
graph_name: str,
method: str,
station,
predict_dump_dir: str | Path | None = None,
based_dir: str | Path = PROJ_DIR,
)