geoprior.utils.split

geoprior.utils.split

Group-holdout split for sequence data.

Exports:

train_windows_T{T}_H{H}.npz val_windows_T{T}_H{H}.npz test_windows_T{T}_H{H}.npz future_inputs_T{T}_H{H}.npz splits_groups.json

Leakage fix (Zhongshan 2 windows/pixel):

split by group_id first, then window inside split.

Functions

build_future_inputs_npz(*, df_scaled, ...[, ...])
build_group_holdout_npzs(*, df_train, ...[, ...])	Build train/val/test windows using group holdout.
pack_xy_npz(x, y)
split_group_keys(keys, *[, cfg])
subset_by_keys(df, *, group_cols, keys[, ...])
write_splits_json(path, *, group_cols, ...)

Classes

SplitCfg([seed, ratios, decimals])

class geoprior.utils.split.SplitCfg(seed: 'int' = 42, ratios: 'tuple[float, float, float]' = (0.7, 0.15, 0.15), decimals: 'int' = 8)

Bases: object

Parameters:

• seed (int)

• ratios (tuple[float, float, float])

• decimals (int)

seed: int = 42

ratios: tuple[float, float, float] = (0.7, 0.15, 0.15)

decimals: int = 8

__init__(seed=42, ratios=(0.7, 0.15, 0.15), decimals=8)

Parameters:

• seed (int)

• ratios (tuple[float, float, float])

• decimals (int)

Return type:

None

geoprior.utils.split.split_group_keys(keys, *, cfg=SplitCfg(seed=42, ratios=(0.7, 0.15, 0.15), decimals=8))

Parameters:

• keys (ndarray)

• cfg (SplitCfg)

Return type:

dict[str, ndarray]

geoprior.utils.split.subset_by_keys(df, *, group_cols, keys, decimals=8)

Parameters:

• df (DataFrame)

• group_cols (Sequence[str])

• keys (ndarray)

• decimals (int)

Return type:

DataFrame

geoprior.utils.split.write_splits_json(path, *, group_cols, time_steps, horizon, train_end, cfg, splits)

Parameters:

• path (str)

• group_cols (Sequence[str])

• time_steps (int)

• horizon (int)

• train_end (float | None)

• cfg (SplitCfg)

• splits (dict[str, ndarray])

Return type:

str

geoprior.utils.split.pack_xy_npz(x, y)

Parameters:

• x (dict[str, Any])

• y (dict[str, Any] | None)

Return type:

dict[str, ndarray]

geoprior.utils.split.build_group_holdout_npzs(*, df_train, artifacts_dir, group_cols, time_col_used, x_col_used, y_col_used, subs_col, gwl_target_col, gwl_dyn_col, h_field_col, static_cols, dynamic_cols, future_cols, time_steps, horizon, mode, model_name, train_end, keys_ok, cfg=SplitCfg(seed=42, ratios=(0.7, 0.15, 0.15), decimals=8), normalize_coords=True)

Build train/val/test windows using group holdout.

Returns dict containing paths and coord_scaler.

Parameters:

• df_train (DataFrame)

• artifacts_dir (str)

• group_cols (Sequence[str])

• time_col_used (str)

• x_col_used (str)

• y_col_used (str)

• subs_col (str)

• gwl_target_col (str)

• gwl_dyn_col (str)

• h_field_col (str)

• static_cols (Sequence[str])

• dynamic_cols (Sequence[str])

• future_cols (Sequence[str])

• time_steps (int)

• horizon (int)

• mode (str)

• model_name (str)

• train_end (float | None)

• keys_ok (ndarray)

• cfg (SplitCfg)

• normalize_coords (bool)

Return type:

dict[str, Any]

geoprior.utils.split.build_future_inputs_npz(*, df_scaled, artifacts_dir, time_col, time_col_num, lon_col, lat_col, subs_col, gwl_col, h_field_col, static_features, dynamic_features, future_features, group_cols, train_end_time, forecast_start_time, horizon, time_steps, mode, model_name, normalize_coords, coord_scaler=None)

Parameters:

• df_scaled (DataFrame)

• artifacts_dir (str)

• time_col (str)

• time_col_num (str | None)

• lon_col (str)

• lat_col (str)

• subs_col (str)

• gwl_col (str)

• h_field_col (str)

• static_features (Sequence[str])

• dynamic_features (Sequence[str])

• future_features (Sequence[str])

• group_cols (Sequence[str])

• train_end_time (Any)

• forecast_start_time (Any)

• horizon (int)

• time_steps (int)

• mode (str)

• model_name (str)

• normalize_coords (bool)

• coord_scaler (Any)

Return type:

str