Datapipes

pdearena.data.twod.datapipes.common

RandomTimeStepConditionedPDETrainData

Bases: IterDataPipe

Randomized data for training conditioned PDEs.

Parameters:

Name	Type	Description	Default
dp	IterDataPipe	Data pipe that returns individual PDE trajectories.	required
n_input_scalar_components	int	Number of input scalar components.	required
n_input_vector_components	int	Number of input vector components.	required
n_output_scalar_components	int	Number of output scalar components.	required
n_output_vector_components	int	Number of output vector components.	required
trajlen	int	Length of a trajectory in the dataset.	required
reweigh	bool	Whether to rebalance the dataset so that longer horizon predictions get equal weightage despite there being fewer actual such datapoints in a trajectory. Defaults to True.	True

Source code in pdearena/data/twod/datapipes/common.py

class RandomTimeStepConditionedPDETrainData(dp.iter.IterDataPipe):
    """Randomized data for training conditioned PDEs.

    Args:
        dp (IterDataPipe): Data pipe that returns individual PDE trajectories.
        n_input_scalar_components (int): Number of input scalar components.
        n_input_vector_components (int): Number of input vector components.
        n_output_scalar_components (int): Number of output scalar components.
        n_output_vector_components (int): Number of output vector components.
        trajlen (int): Length of a trajectory in the dataset.
        reweigh (bool, optional): Whether to rebalance the dataset so that longer horizon predictions get equal weightage despite there being fewer actual such datapoints in a trajectory. Defaults to True.
    """

    def __init__(
        self,
        dp,
        n_input_scalar_components: int,
        n_input_vector_components: int,
        n_output_scalar_components: int,
        n_output_vector_components: int,
        trajlen: int,
        reweigh=True,
    ) -> None:
        super().__init__()
        self.dp = dp
        self.n_input_scalar_components = n_input_scalar_components
        self.n_input_vector_components = n_input_vector_components
        self.n_output_scalar_components = n_output_scalar_components
        self.n_output_vector_components = n_output_vector_components

        self.trajlen = trajlen
        self.reweigh = reweigh

    def __iter__(self):
        time_resolution = self.trajlen

        for u, v, cond, grid in self.dp:
            if self.reweigh:
                end_time = random.choices(range(1, time_resolution), k=1)[0]
                start_time = random.choices(range(0, end_time), weights=1 / np.arange(1, end_time + 1), k=1)[0]
            else:
                end_time = torch.randint(low=1, high=time_resolution, size=(1,), dtype=torch.long).item()
                start_time = torch.randint(low=0, high=end_time.item(), size=(1,), dtype=torch.long).item()

            delta_t = end_time - start_time
            yield (
                *datautils.create_time_conditioned_data(
                    self.n_input_scalar_components,
                    self.n_input_vector_components,
                    self.n_output_scalar_components,
                    self.n_output_vector_components,
                    u,
                    v,
                    grid,
                    start_time,
                    end_time,
                    torch.tensor([delta_t]),
                ),
                cond,
            )

RandomizedPDETrainData

Bases: IterDataPipe

Randomized data for training PDEs.

Parameters:

Name	Type	Description	Default
dp	IterDataPipe	Data pipe that returns individual PDE trajectories.	required
n_input_scalar_components	int	Number of input scalar components.	required
n_input_vector_components	int	Number of input vector components.	required
n_output_scalar_components	int	Number of output scalar components.	required
n_output_vector_components	int	Number of output vector components.	required
trajlen	int	Length of a trajectory in the dataset.	required
time_history	int	Number of time steps of inputs.	required
time_future	int	Number of time steps of outputs.	required
time_gap	int	Number of time steps between inputs and outputs.	required

Source code in pdearena/data/twod/datapipes/common.py

class RandomizedPDETrainData(dp.iter.IterDataPipe):
    """Randomized data for training PDEs.

    Args:
        dp (IterDataPipe): Data pipe that returns individual PDE trajectories.
        n_input_scalar_components (int): Number of input scalar components.
        n_input_vector_components (int): Number of input vector components.
        n_output_scalar_components (int): Number of output scalar components.
        n_output_vector_components (int): Number of output vector components.
        trajlen (int): Length of a trajectory in the dataset.
        time_history (int): Number of time steps of inputs.
        time_future (int): Number of time steps of outputs.
        time_gap (int): Number of time steps between inputs and outputs.
    """

    def __init__(
        self,
        dp,
        n_input_scalar_components: int,
        n_input_vector_components: int,
        n_output_scalar_components: int,
        n_output_vector_components: int,
        trajlen: int,
        time_history: int,
        time_future: int,
        time_gap: int,
    ) -> None:
        super().__init__()
        self.dp = dp
        self.n_input_scalar_components = n_input_scalar_components
        self.n_input_vector_components = n_input_vector_components
        self.n_output_scalar_components = n_output_scalar_components
        self.n_output_vector_components = n_output_vector_components
        self.trajlen = trajlen
        self.time_history = time_history
        self.time_future = time_future
        self.time_gap = time_gap

    def __iter__(self):
        # Length of trajectory
        time_resolution = self.trajlen
        # Max number of previous points solver can eat
        reduced_time_resolution = time_resolution - self.time_history
        # Number of future points to predict
        max_start_time = reduced_time_resolution - self.time_future - self.time_gap

        for batch in self.dp:
            if len(batch) == 3:
                (u, v, grid) = batch
                cond = None
            elif len(batch) == 4:
                (u, v, cond, grid) = batch
            else:
                raise ValueError(f"Unknown batch length of {len(batch)}.")

            # Choose initial random time point at the PDE solution manifold
            start_time = random.choices([t for t in range(max_start_time + 1)], k=1)
            yield datautils.create_data2D(
                self.n_input_scalar_components,
                self.n_input_vector_components,
                self.n_output_scalar_components,
                self.n_output_vector_components,
                u,
                v,
                grid,
                start_time[0],
                self.time_history,
                self.time_future,
                self.time_gap,
            )

TimestepConditionedPDEEvalData

Bases: IterDataPipe

Data for evaluation of time conditioned PDEs

Parameters:

Name	Type	Description	Default
dp	IterDataPipe	Data pipe that returns individual PDE trajectories.	required
trajlen	int	Length of a trajectory in the dataset.	required
delta_t	int	Evaluates predictions conditioned at that delta_t.	required

Tip

Make sure delta_t is less than half of trajlen.

Source code in pdearena/data/twod/datapipes/common.py

class TimestepConditionedPDEEvalData(dp.iter.IterDataPipe):
    """Data for evaluation of time conditioned PDEs

    Args:
        dp (torchdata.datapipes.iter.IterDataPipe): Data pipe that returns individual PDE trajectories.
        trajlen (int): Length of a trajectory in the dataset.
        delta_t (int): Evaluates predictions conditioned at that delta_t.

    Tip:
        Make sure `delta_t` is less than half of `trajlen`.
    """

    def __init__(self, dp: dp.iter.IterDataPipe, trajlen: int, delta_t: int) -> None:
        super().__init__()
        self.dp = dp
        self.trajlen = trajlen
        if 2 * delta_t >= self.trajlen:
            raise ValueError("delta_t should be less than half the trajectory length")

        self.delta_t = delta_t

    def __iter__(self):
        for begin in range(self.trajlen - self.delta_t):
            for u, v, cond, grid in self.dp:
                newu = u[begin :: self.delta_t, ...]
                newv = v[begin :: self.delta_t, ...]
                max_start_time = newu.size(0)
                for start in range(max_start_time - 1):
                    end = start + 1
                    data = torch.cat((newu[start : start + 1], newv[start : start + 1]), dim=1).unsqueeze(0)
                    if grid is not None:
                        data = torch.cat((data, grid), dim=1)
                    label = torch.cat((newu[end : end + 1], newv[end : end + 1]), dim=1).unsqueeze(0)
                    if data.size(1) == 0:
                        raise ValueError("Data is empty. Likely indexing issue.")
                    if label.size(1) == 0:
                        raise ValueError("Label is empty. Likely indexing issue.")
                    yield data, label, torch.tensor([self.delta_t]), cond

ZarrLister

Bases: IterDataPipe

Customized lister for zarr files.

Parameters:

Name	Type	Description	Default
root	Union[str, Sequence[str], IterDataPipe]	Root directory. Defaults to ".".	'.'

Yields:

Type	Description
str	Path to the zarr file.

Source code in pdearena/data/twod/datapipes/common.py

class ZarrLister(dp.iter.IterDataPipe):
    """Customized lister for zarr files.

    Args:
        root (Union[str, Sequence[str], dp.iter.IterDataPipe], optional): Root directory. Defaults to ".".

    Yields:
        (str): Path to the zarr file.
    """

    def __init__(
        self,
        root: Union[str, Sequence[str], dp.iter.IterDataPipe] = ".",
    ) -> None:
        super().__init__()

        if isinstance(root, str):
            root = [root]
        if not isinstance(root, dp.iter.IterDataPipe):
            root = dp.iter.IterableWrapper(root)

        self.datapipe: dp.iter.IterDataPipe = root

    def __iter__(self):
        for path in self.datapipe:
            for dirname in os.listdir(path):
                if dirname.endswith(".zarr"):
                    yield os.path.join(path, dirname)

build_datapipes(pde, data_path, limit_trajectories, usegrid, dataset_opener, lister, sharder, filter_fn, mode, time_history=1, time_future=1, time_gap=0, onestep=False, conditioned=False, delta_t=None, conditioned_reweigh=True)

Build datapipes for training and evaluation.

Parameters:

Name	Type	Description	Default
pde	PDEDataConfig	PDE configuration.	required
data_path	str	Path to the data.	required
limit_trajectories	int	Number of trajectories to use.	required
usegrid	bool	Whether to use spatial grid as input.	required
dataset_opener	Callable[..., IterDataPipe]	Dataset opener.	required
lister	Callable[..., IterDataPipe]	List files.	required
sharder	Callable[..., IterDataPipe]	Shard files.	required
filter_fn	Callable[..., IterDataPipe]	Filter files.	required
mode	str	Mode of the data. ["train", "valid", "test"]	required
time_history	int	Number of time steps in the past. Defaults to 1.	1
time_future	int	Number of time steps in the future. Defaults to 1.	1
time_gap	int	Number of time steps between the past and the future to be skipped. Defaults to 0.	0
onestep	bool	Whether to use one-step prediction. Defaults to False.	False
conditioned	bool	Whether to use conditioned data. Defaults to False.	False
delta_t	Optional[int]	Time step size. Defaults to None. Only used for conditioned data.	None
conditioned_reweigh	bool	Whether to reweight conditioned data. Defaults to True.	True

Returns:

Name	Type	Description
dpipe	IterDataPipe	IterDataPipe for training and evaluation.

Source code in pdearena/data/twod/datapipes/common.py

def build_datapipes(
    pde: PDEDataConfig,
    data_path,
    limit_trajectories,
    usegrid: bool,
    dataset_opener: Callable[..., dp.iter.IterDataPipe],
    lister: Callable[..., dp.iter.IterDataPipe],
    sharder: Callable[..., dp.iter.IterDataPipe],
    filter_fn: Callable[..., dp.iter.IterDataPipe],
    mode: str,
    time_history=1,
    time_future=1,
    time_gap=0,
    onestep=False,
    conditioned=False,
    delta_t: Optional[int] = None,
    conditioned_reweigh: bool = True,
):
    """Build datapipes for training and evaluation.

    Args:
        pde (PDEDataConfig): PDE configuration.
        data_path (str): Path to the data.
        limit_trajectories (int): Number of trajectories to use.
        usegrid (bool): Whether to use spatial grid as input.
        dataset_opener (Callable[..., dp.iter.IterDataPipe]): Dataset opener.
        lister (Callable[..., dp.iter.IterDataPipe]): List files.
        sharder (Callable[..., dp.iter.IterDataPipe]): Shard files.
        filter_fn (Callable[..., dp.iter.IterDataPipe]): Filter files.
        mode (str): Mode of the data. ["train", "valid", "test"]
        time_history (int, optional): Number of time steps in the past. Defaults to 1.
        time_future (int, optional): Number of time steps in the future. Defaults to 1.
        time_gap (int, optional): Number of time steps between the past and the future to be skipped. Defaults to 0.
        onestep (bool, optional): Whether to use one-step prediction. Defaults to False.
        conditioned (bool, optional): Whether to use conditioned data. Defaults to False.
        delta_t (Optional[int], optional): Time step size. Defaults to None. Only used for conditioned data.
        conditioned_reweigh (bool, optional): Whether to reweight conditioned data. Defaults to True.

    Returns:
        dpipe (IterDataPipe): IterDataPipe for training and evaluation.
    """
    dpipe = lister(
        data_path,
    ).filter(filter_fn=filter_fn)
    if mode == "train":
        dpipe = dpipe.shuffle()

    dpipe = dataset_opener(
        sharder(dpipe),
        mode=mode,
        limit_trajectories=limit_trajectories,
        usegrid=usegrid,
    )
    if mode == "train":
        # Make sure that in expectation we have seen all the data despite randomization
        dpipe = dpipe.cycle(pde.trajlen)

    if mode == "train":
        # Training data is randomized
        if conditioned:
            dpipe = RandomTimeStepConditionedPDETrainData(
                dpipe,
                pde.n_scalar_components,
                pde.n_vector_components,
                pde.n_scalar_components,
                pde.n_vector_components,
                pde.trajlen,
                conditioned_reweigh,
            )
        else:
            dpipe = RandomizedPDETrainData(
                dpipe,
                pde.n_scalar_components,
                pde.n_vector_components,
                pde.n_scalar_components,
                pde.n_vector_components,
                pde.trajlen,
                time_history,
                time_future,
                time_gap,
            )
    else:
        # Evaluation data is not randomized.
        if conditioned and onestep:
            assert delta_t is not None
            dpipe = TimestepConditionedPDEEvalData(dpipe, pde.trajlen, delta_t)
        elif onestep:
            dpipe = PDEEvalTimeStepData(
                dpipe,
                pde.n_scalar_components,
                pde.n_vector_components,
                pde.n_scalar_components,
                pde.n_vector_components,
                pde.trajlen,
                time_history,
                time_future,
                time_gap,
            )

    return dpipe

---

pdearena.data.twod.datapipes.navierstokes2d

NavierStokesDatasetOpener

Bases: IterDataPipe

DataPipe to load Navier-Stokes dataset.

Parameters:

Name	Type	Description	Default
dp	IterDataPipe	List of hdf5 files containing Navier-Stokes data.	required
mode	str	Mode to load data from. Can be one of train, val, test.	required
limit_trajectories	int	Limit the number of trajectories to load from individual hdf5 file. Defaults to None.	None
usegrid	bool	Whether to output spatial grid or not. Defaults to False.	False

Yields:

Type	Description
Tuple[Tensor, Tensor, Optional[Tensor], Optional[Tensor]]	Tuple containing particle scalar field, velocity vector field, and optionally buoyancy force parameter value and spatial grid.

Source code in pdearena/data/twod/datapipes/navierstokes2d.py

class NavierStokesDatasetOpener(dp.iter.IterDataPipe):
    """DataPipe to load Navier-Stokes dataset.

    Args:
        dp (dp.iter.IterDataPipe): List of `hdf5` files containing Navier-Stokes data.
        mode (str): Mode to load data from. Can be one of `train`, `val`, `test`.
        limit_trajectories (int, optional): Limit the number of trajectories to load from individual `hdf5` file. Defaults to None.
        usegrid (bool, optional): Whether to output spatial grid or not. Defaults to False.

    Yields:
        (Tuple[torch.Tensor, torch.Tensor, Optional[torch.Tensor], Optional[torch.Tensor]]): Tuple containing particle scalar field, velocity vector field, and optionally buoyancy force parameter value  and spatial grid.
    """

    def __init__(self, dp, mode: str, limit_trajectories: Optional[int] = None, usegrid: bool = False) -> None:
        super().__init__()
        self.dp = dp
        self.mode = mode
        self.limit_trajectories = limit_trajectories
        self.usegrid = usegrid

    def __iter__(self):
        for path in self.dp:
            with h5py.File(path, "r") as f:
                data = f[self.mode]
                if self.limit_trajectories is None or self.limit_trajectories == -1:
                    num = data["u"].shape[0]
                else:
                    num = self.limit_trajectories

                iter_start = 0
                iter_end = num

                for idx in range(iter_start, iter_end):
                    u = torch.tensor(data["u"][idx])
                    vx = torch.tensor(data["vx"][idx])
                    vy = torch.tensor(data["vy"][idx])
                    if "buo_y" in data:
                        cond = torch.tensor(data["buo_y"][idx]).unsqueeze(0).float()
                    else:
                        cond = None

                    v = torch.cat((vx[:, None], vy[:, None]), dim=1)

                    if self.usegrid:
                        gridx = torch.linspace(0, 1, data["x"][idx].shape[0])
                        gridy = torch.linspace(0, 1, data["y"][idx].shape[0])
                        gridx = gridx.reshape(
                            1,
                            gridx.size(0),
                            1,
                        ).repeat(
                            1,
                            1,
                            gridy.size(0),
                        )
                        gridy = gridy.reshape(
                            1,
                            1,
                            gridy.size(0),
                        ).repeat(
                            1,
                            gridx.size(1),
                            1,
                        )
                        grid = torch.cat((gridx[:, None], gridy[:, None]), dim=1)
                    else:
                        grid = None
                    yield u.unsqueeze(1).float(), v.float(), cond, grid

pdearena.data.twod.datapipes.shallowwater2d

ShallowWaterDatasetOpener

Bases: IterDataPipe

DataPipe for loading the shallow water dataset

Parameters:

Name	Type	Description	Default
dp	IterDataPipe	datapipe with paths to load the dataset from.	required
mode	str	"train" or "valid" or "test"	required
limit_trajectories	Optional[int]	number of trajectories to load from the dataset	None
usevort	bool	whether to use vorticity or velocity. If False, velocity is returned.	False
usegrid	bool	whether to use grid or not. If False, no grid is returned.	False
sample_rate	int	sample rate for the data. Default is 1, which means no sub-sampling.	1

Note

We manually manage the data distribution across workers and processes. So make sure not to use torchdata's dp.iter.Sharder with this data pipe.

Source code in pdearena/data/twod/datapipes/shallowwater2d.py

class ShallowWaterDatasetOpener(dp.iter.IterDataPipe):
    """DataPipe for loading the shallow water dataset

    Args:
        dp: datapipe with paths to load the dataset from.
        mode (str): "train" or "valid" or "test"
        limit_trajectories: number of trajectories to load from the dataset
        usevort (bool): whether to use vorticity or velocity. If False, velocity is returned.
        usegrid (bool): whether to use grid or not. If False, no grid is returned.
        sample_rate: sample rate for the data. Default is 1, which means no sub-sampling.

    Note:
        We manually manage the data distribution across workers and processes. So make sure not to use `torchdata`'s [dp.iter.Sharder][torchdata.datapipes.iter.ShardingFilter] with this data pipe.
    """

    def __init__(
        self,
        dp: dp.iter.IterDataPipe,
        mode: str,
        limit_trajectories: Optional[int] = None,
        usevort: bool = False,
        usegrid: bool = False,
        sample_rate: int = 1,
    ) -> None:
        super().__init__()
        self.dp = dp
        self.mode = mode
        self.limit_trajectories = limit_trajectories
        self.usevort = usevort
        self.usegrid = usegrid
        self.sample_rate = sample_rate

    def __iter__(self):
        for path in self.dp:
            if "zarr" in path:
                data = xr.open_zarr(path)
            else:
                # Note that this is much slower
                data = xr.open_mfdataset(
                    os.path.join(path, "seed=*", "run*", "output.nc"),
                    concat_dim="b",
                    combine="nested",
                    parallel=True,
                )

            normstat = torch.load(os.path.join(path, "..", "normstats.pt"))
            if self.limit_trajectories is None or self.limit_trajectories == -1:
                num = data["u"].shape[0]
            else:
                num = self.limit_trajectories

            if dist.is_initialized():
                rank = dist.get_rank()
                world_size = dist.get_world_size()
            else:
                rank = 0
                world_size = 1

            # Different workers should be using different trajectory batches
            worker_info = torch.utils.data.get_worker_info()
            if worker_info is not None:
                num_workers_per_dist = min(worker_info.num_workers, num)
                num_shards = num_workers_per_dist * world_size
                per_worker = int(math.floor(num / float(num_shards)))
                wid = rank * num_workers_per_dist + worker_info.id
                iter_start = wid * per_worker
                iter_end = iter_start + per_worker
            else:
                per_dist = int(math.floor(num / float(world_size)))
                iter_start = rank * per_dist
                iter_end = iter_start + per_dist

            for idx in range(iter_start, iter_end):
                if self.usevort:
                    vort = torch.tensor(data["vor"][idx].to_numpy())
                    vort = (vort - normstat["vor"]["mean"]) / normstat["vor"]["std"]
                else:
                    u = torch.tensor(data["u"][idx].to_numpy())
                    v = torch.tensor(data["v"][idx].to_numpy())
                    vecf = torch.cat((u, v), dim=1)

                pres = torch.tensor(data["pres"][idx].to_numpy())

                pres = (pres - normstat["pres"]["mean"]) / normstat["pres"]["std"]
                pres = pres.unsqueeze(1)

                if self.sample_rate > 1:
                    # TODO: hardocded skip_nt=4
                    pres = pres[4 :: self.sample_rate]
                    if self.usevort:
                        vort = vort[4 :: self.sample_rate]
                    else:
                        vecf = vecf[4 :: self.sample_rate]
                if self.usegrid:
                    raise NotImplementedError("Grid not implemented for weather data")
                else:
                    if self.usevort:
                        yield torch.cat((pres, vort), dim=1).float(), None, None, None
                    else:
                        yield pres.float(), vecf.float(), None, None

---

pdearena.data.threed.datapipes

build_maxwell_datapipes(pde, data_path, limit_trajectories, usegrid, dataset_opener, lister, sharder, filter_fn, mode, time_history=1, time_future=1, time_gap=0, onestep=False)

Build datapipes for training and evaluation.

Parameters:

Name	Type	Description	Default
pde	PDEDataConfig	PDE configuration.	required
data_path	str	Path to the data.	required
limit_trajectories	int	Number of trajectories to use.	required
usegrid	bool	Whether to use spatial grid as input.	required
dataset_opener	Callable[..., IterDataPipe]	Dataset opener.	required
lister	Callable[..., IterDataPipe]	List files.	required
sharder	Callable[..., IterDataPipe]	Shard files.	required
filter_fn	Callable[..., IterDataPipe]	Filter files.	required
mode	str	Mode of the data. ["train", "valid", "test"]	required
time_history	int	Number of time steps in the past. Defaults to 1.	1
time_future	int	Number of time steps in the future. Defaults to 1.	1
time_gap	int	Number of time steps between the past and the future to be skipped. Defaults to 0.	0
onestep	bool	Whether to use one-step prediction. Defaults to False.	False

Returns:

Name	Type	Description
dpipe	IterDataPipe	IterDataPipe for training and evaluation.

Source code in pdearena/data/threed/datapipes.py

def build_maxwell_datapipes(
    pde: PDEDataConfig,
    data_path,
    limit_trajectories,
    usegrid: bool,
    dataset_opener: Callable[..., dp.iter.IterDataPipe],
    lister: Callable[..., dp.iter.IterDataPipe],
    sharder: Callable[..., dp.iter.IterDataPipe],
    filter_fn: Callable[..., dp.iter.IterDataPipe],
    mode: str,
    time_history=1,
    time_future=1,
    time_gap=0,
    onestep=False,
):
    """Build datapipes for training and evaluation.

    Args:
        pde (PDEDataConfig): PDE configuration.
        data_path (str): Path to the data.
        limit_trajectories (int): Number of trajectories to use.
        usegrid (bool): Whether to use spatial grid as input.
        dataset_opener (Callable[..., dp.iter.IterDataPipe]): Dataset opener.
        lister (Callable[..., dp.iter.IterDataPipe]): List files.
        sharder (Callable[..., dp.iter.IterDataPipe]): Shard files.
        filter_fn (Callable[..., dp.iter.IterDataPipe]): Filter files.
        mode (str): Mode of the data. ["train", "valid", "test"]
        time_history (int, optional): Number of time steps in the past. Defaults to 1.
        time_future (int, optional): Number of time steps in the future. Defaults to 1.
        time_gap (int, optional): Number of time steps between the past and the future to be skipped. Defaults to 0.
        onestep (bool, optional): Whether to use one-step prediction. Defaults to False.

    Returns:
        dpipe (IterDataPipe): IterDataPipe for training and evaluation.
    """
    dpipe = lister(
        data_path,
    ).filter(filter_fn=filter_fn)
    if mode == "train":
        dpipe = dpipe.shuffle()

    dpipe = dataset_opener(
        sharder(dpipe),
        mode=mode,
        limit_trajectories=limit_trajectories,
        usegrid=usegrid,
    )
    if mode == "train":
        # Make sure that in expectation we have seen all the data despite randomization
        dpipe = dpipe.cycle(pde.trajlen)

    if mode == "train":
        # Training data is randomized
        dpipe = RandomizedPDETrainData3D(
                dpipe,
                pde,
                time_history,
                time_future,
                time_gap,
            )   
    else:
        # Evaluation data is not randomized.
        if onestep:
            dpipe = PDEEvalTimeStepData3D(
                dpipe,
                pde,
                time_history,
                time_future,
                time_gap,
            )

    return dpipe