phlower.nn.InterpolatorPresetGroupModule

class phlower.nn.InterpolatorPresetGroupModule(source_position_name, target_position_name, source_position_from_field=True, target_position_from_field=True, variable_names=None, input_to_output_map=None, source_field_name=None, target_field_name=None, n_head=1, weight_name=None, length_scale=0.1, normalize_row=True, conservative=False, bidirectional=False, k_nns=9, d_nns=None, learnable_length_scale=False, learnable_weight_transpose=False, learnable_weight_distance=False, recompute_distance_k_nns_if_grad_enabled=False, recompute_distance_d_nns_if_grad_enabled=False, unbatch=True, epsilon=1e-05)[source]

Bases: Module, IGenericPhlowerCoreModule[IPhlowerPresetGroupParameters, IPhlowerTensorCollections]

Semi-Lagrangian Attention is a neural network module for convection-dominated problems based on the Semi-Lagrangian scheme.

Parameters:

  • source_position_name (str) – Name to be used for source position data.

  • target_position_name (str) – Name to be used for target position data.

  • source_position_from_field (bool = True) – If False, take source position from tensors rather than field.

  • target_position_from_field (bool = True) – If False, take target position from tensors rather than field.

  • variable_names (list[str] | None = None) – Variable names to be interpolated. If not fed, all data are interpolated.

  • source_field_name (str | None = None) – Name of the source to use for unbatching source fields.

  • target_field_name (str | None = None) – Name of the target to use for unbatching source fields.

  • n_head (int = 1) – The number of head for the multihead attention mechanism.

  • weight_name (str | None = None) – Name to be used for weight data.

  • length_scale (float = 1.0e-1) – Length scale to normalize distance-based attention.

  • normalize_row (bool = True) – If True, normalize the attention matrix in the row direction.

  • conservative (bool = False) – If True, normalize the attention matrix further to achieve conservation.

  • bidirectional (bool = False) – If True, compute bidirectional interpolation assuming both source and target represent the same shape.

  • k_nns (int = 9) – The number to be used for k nearest neighbor search. It should be larger than 1.

  • d_nns (float | None = None) – If fed, perform nearest neighbor search based on the fed distance in addition to the k-NN search.

  • learnable_length_scale (bool = False) – If True, learn length scale. In that case, the length_scale parameter above is used to initialize trainable weights.

  • learnable_weight_transpose (bool = False) – If True, learn weight for the transposed attention matrix.

  • learnable_weight_distance (bool = False) – If True, learn weight for the distance based attention matrix.

  • recompute_distance_k_nns_if_grad_enabled (bool = False) – If True and torch.is_grad_enabled, re-compute distance for k-NN attention matrix to retain gradients.

  • recompute_distance_d_nns_if_grad_enabled (bool = False) – If True and torch.is_grad_enabled, re-compute distance for distance attention matrix to retain gradient.

  • unbatch (bool = True) – If True, unbatch the input data and field.

  • epsilon (float = 1.0e-5) – Epsilon value to decimate small numbers.

  • input_to_output_map (dict[str, str] | None)

Examples

>>> interpolator = InterpolatorPresetGroupModule(
...     source_position_name="source_position",
...     target_position_name="target_position",
... )
>>> output = interpolator(data, field_data=field_data)

Methods

forward(data, *, field_data, **kwards)

forward function which overload torch.nn.Module

from_setting(setting)

Create InterpolatorPresetGroupModule from InterpolatorPresetGroupSetting instance.

get_nn_name()

Return neural network name

Attributes

T_destination

call_super_init

dump_patches

source_field_name_for_unbatch

target_field_name_for_unbatch

training

forward(data, *, field_data, **kwards)[source]

forward function which overload torch.nn.Module

Parameters:

  • data (IPhlowerTensorCollections) – IPhlowerTensorCollections data which receives from predecessors

  • field_data (ISimulationField | None) – ISimulationField | None Constant information through training or prediction

Returns:

Tensor object

Return type:

PhlowerTensor

classmethod from_setting(setting)[source]

Create InterpolatorPresetGroupModule from InterpolatorPresetGroupSetting instance.

Parameters:

  • InterpolatorPresetGroupSetting) (setting) – setting object

  • setting (InterpolatorPresetGroupSetting)

Returns:

InterpolatorPresetGroupModule object

Return type:

InterpolatorPresetGroupModule

classmethod get_nn_name()[source]

Return neural network name

Returns:

name

Return type:

str