examples/pytorch_dataloader.py

import argparse
from dataclasses import dataclass
from typing import Tuple

import numpy as np
import torch
import xarray as xr
from torch.utils.data import DataLoader, Dataset


@dataclass(frozen=True)
class Window:
    lookback_hours: int = 168
    horizon_hours: int = 24
    stride_hours: int = 24


class NwmBasinForecastDataset(Dataset):
    """
    Returns:
      x: (basin, lookback_hours, 2) float32  -> [streamflow, precipitation_rate]
      y: (basin, horizon_hours) float32      -> streamflow
    """

    def __init__(self, zarr_path: str, *, split: str, window: Window = Window()) -> None:
        ds = xr.open_zarr(zarr_path, consolidated=True)
        self.ds = ds
        self.window = window

        # (time, basin)
        sf = ds["streamflow"]
        pr = ds["precipitation_rate"]

        # Basic time splits for 2018–2019 hourly data.
        # Train: 2018
        # Val:   2019-01 .. 2019-06
        # Test:  2019-07 .. 2019-12
        t = ds["time"].values
        train_end = np.datetime64("2018-12-31T23:00:00")
        val_end = np.datetime64("2019-06-30T23:00:00")

        if split == "train":
            t0 = 0
            t1 = int(np.searchsorted(t, train_end, side="right")) - 1
        elif split == "val":
            t0 = int(np.searchsorted(t, train_end, side="right"))
            t1 = int(np.searchsorted(t, val_end, side="right")) - 1
        elif split == "test":
            t0 = int(np.searchsorted(t, val_end, side="right"))
            t1 = len(t) - 1
        else:
            raise ValueError("split must be one of: train, val, test")

        n = window.lookback_hours + window.horizon_hours
        self.start_indices = list(range(t0, t1 - n + 2, window.stride_hours))

        self.sf = sf
        self.pr = pr

    def __len__(self) -> int:
        return len(self.start_indices)

    def __getitem__(self, idx: int) -> Tuple[torch.Tensor, torch.Tensor]:
        i = self.start_indices[idx]
        lb = self.window.lookback_hours
        hz = self.window.horizon_hours

        sf_x = np.asarray(self.sf.isel(time=slice(i, i + lb)).values, dtype=np.float32)  # (lb, basin)
        pr_x = np.asarray(self.pr.isel(time=slice(i, i + lb)).values, dtype=np.float32)  # (lb, basin)
        y = np.asarray(self.sf.isel(time=slice(i + lb, i + lb + hz)).values, dtype=np.float32)  # (hz, basin)

        # Return per-basin sequences.
        x = np.stack([sf_x.T, pr_x.T], axis=-1)  # (basin, lb, 2)
        y = y.T  # (basin, hz)

        return torch.from_numpy(x), torch.from_numpy(y)


def main() -> None:
    p = argparse.ArgumentParser()
    p.add_argument("--zarr", default="nwm_hydrology_benchmark.zarr")
    p.add_argument("--split", default="train", choices=["train", "val", "test"])
    p.add_argument("--batch-size", type=int, default=4)
    p.add_argument("--num-workers", type=int, default=0)
    args = p.parse_args()

    ds = NwmBasinForecastDataset(args.zarr, split=args.split)
    dl = DataLoader(ds, batch_size=args.batch_size, shuffle=(args.split == "train"), num_workers=args.num_workers)

    x, y = next(iter(dl))
    print("x:", tuple(x.shape), x.dtype, "y:", tuple(y.shape), y.dtype)


if __name__ == "__main__":
    main()