credit.ensemble.bred_vector

credit.ensemble.bred_vector#

Attributes#

logger

Classes#

BredVector

Functions#

generate_bred_vectors(x_batch, model[, num_cycles, ...])

Generate bred vectors and initialize initial conditions for the given batch.

generate_bred_vectors_cycle(initial_condition, ...[, ...])

Generate bred vectors and initialize initial conditions for the given batch.

clone_dataset(dataset)

Clones a PyTorch Dataset by creating a deep copy.

adjust_start_times(time_ranges[, hours])

Adjusts the start times by subtracting 24 hours.

Module Contents#

class credit.ensemble.bred_vector.BredVector(model: Callable[[torch.Tensor], torch.Tensor], noise_amplitude: float = 0.15, num_cycles: int = 5, integration_steps: int = 1, perturbation_method: Callable[[torch.Tensor, collections.OrderedDict[str, numpy.ndarray]], torch.Tensor] | None = None, hemispheric_rescale: Callable[[torch.Tensor, torch.Tensor], torch.Tensor] | None = None, terrain_file: str = None, perturb_channel_idx: int = None, ensemble_perturb: bool = False, clamp: bool = False, clamp_min: float = None, clamp_max: float = None, input_static_dim: int = 3, varnum_diag: int = 0, post_conf: dict = {})#
model#
noise_amplitude = 0.15#
num_cycles = 5#
integration_steps = 1#
perturbation_method = None#
hemispheric_rescale#
ensemble_perturb = False#
perturb_channel_idx = None#
clamp = False#
clamp_min = None#
clamp_max = None#
input_static_dim = 3#
varnum_diag = 0#
post_conf#
flag_mass_conserve = False#
flag_water_conserve = False#
flag_energy_conserve = False#
use_post_block = False#
perturb(x_input: torch.Tensor, forecast_step: int = 1) → torch.Tensor#
__call__(initial_condition: torch.Tensor, dataset, return_delta_x=False) → list[torch.Tensor]#
credit.ensemble.bred_vector.generate_bred_vectors(x_batch, model, num_cycles=5, perturbation_std=0.15, epsilon=1.0, flag_clamp=False, clamp_min=None, clamp_max=None)#

Generate bred vectors and initialize initial conditions for the given batch.

Parameters:

  • x_batch (torch.Tensor) – The input batch.

  • batch (dict) – A dictionary containing additional batch data.

  • model (nn.Module) – The model used for predictions.

  • num_cycles (int) – Number of perturbation cycles.

  • perturbation_std (float) – Magnitude of initial perturbations.

  • epsilon (float) – Scaling factor for bred vectors.

  • flag_clamp (bool, optional) – Whether to clamp inputs. Defaults to False.

  • clamp_min (float, optional) – Minimum clamp value. Required if flag_clamp is True.

  • clamp_max (float, optional) – Maximum clamp value. Required if flag_clamp is True.

Returns:

List of initial conditions generated using bred vectors.

Return type:

list[torch.Tensor]

credit.ensemble.bred_vector.generate_bred_vectors_cycle(initial_condition, dataset, model, num_cycles=5, perturbation_std=0.15, epsilon=1.0, flag_clamp=False, clamp_min=None, clamp_max=None, device='cuda', history_len=1, varnum_diag=None, static_dim_size=None, post_conf={})#

Generate bred vectors and initialize initial conditions for the given batch.

Parameters:

  • x_batch (torch.Tensor) – The input batch.

  • batch (dict) – A dictionary containing additional batch data.

  • model (nn.Module) – The model used for predictions.

  • num_cycles (int) – Number of perturbation cycles.

  • perturbation_std (float) – Magnitude of initial perturbations.

  • epsilon (float) – Scaling factor for bred vectors.

  • flag_clamp (bool, optional) – Whether to clamp inputs. Defaults to False.

  • clamp_min (float, optional) – Minimum clamp value. Required if flag_clamp is True.

  • clamp_max (float, optional) – Maximum clamp value. Required if flag_clamp is True.

Returns:

List of initial conditions generated using bred vectors.

Return type:

list[torch.Tensor]

credit.ensemble.bred_vector.clone_dataset(dataset)#

Clones a PyTorch Dataset by creating a deep copy.

Parameters:

dataset (torch.utils.data.Dataset) – The original dataset.

Returns:

A cloned dataset.

Return type:

torch.utils.data.Dataset

credit.ensemble.bred_vector.adjust_start_times(time_ranges, hours=24)#

Adjusts the start times by subtracting 24 hours.

Parameters:

time_ranges (list of lists) – Each sublist contains [start_time, end_time] as strings.

Returns:

Adjusted time ranges [[start_time - 24hrs, start_time], …]

Return type:

list of lists

credit.ensemble.bred_vector.logger#