phlower.nn.SimilarityEquivariantMLP

class phlower.nn.SimilarityEquivariantMLP(nodes, scale_names, activations=None, dropouts=None, bias=True, create_linear_weight=False, norm_function_name='identity', disable_en_equivariance=False, invariant=False, centering=False, coeff_amplify=1.0, cross_interaction=False, normalize=True, unbatch=True)

Bases: IPhlowerCoreModule, Module

Similarity-equivariant Multi Layer Perceptron as in https://proceedings.mlr.press/v235/horie24a.html.

Parameters:

• nodes (list[int]) – List of feature dimension sizes (The last value of tensor shape).

• scale_names (dict[str, str]) – Name of scales to be used to convert input features to dimensionless. Keys should be PhysicalDimensionSymbolType (T, L, M, …) and values should be names of fields.

• activations (list[str] | None (optional)) – List of activation functions to apply to the output. Defaults to None.

• dropouts (list[float] | None (optional)) – List of dropout rates to apply to the output. Defaults to None.

• bias (bool) – Whether to use bias.

• create_linear_weight (bool) – Whether to create a linear weight.

• norm_function_name (str) – Name of the normalization function to apply to the output. Defaults to “identity”.

• disable_en_equivariance (bool) – Whether to disable en-equivariance.

• invariant (bool) – Whether to make the output invariant to the input.

• centering (bool) – Whether to center the output.

• coeff_amplify (float) – Coefficient to amplify the output.

• cross_interaction (bool) – If True, use interaction layer proposed by https://arxiv.org/abs/2203.06442.

• normalize (bool) – If True, use eq (12) in https://arxiv.org/abs/2203.06442 when cross_interaction is True.

• unbatch (bool) – If True, perform unbatching. Set False when you call this module from another module which perform unbaching. Note that this option cannot be set from yaml. Default: True

Examples

>>> similarity_equivariant_mlp = SimilarityEquivariantMLP(
...     nodes=[10, 20, 30],
...     scale_names={"T": "time", "L": "length"},
...     activations=["relu", "relu", "relu"],
...     dropouts=[0.1, 0.1, 0.1],
...     bias=True,
...     create_linear_weight=True,
...     norm_function_name="identity",
...     disable_en_equivariance=False,
...     invariant=False,
...     centering=False,
...     coeff_amplify=1.0,
...     cross_interaction=False,
...     normalize=True,
... )
>>> similarity_equivariant_mlp(data, field_data=field_data)

Methods

forward(data, *[, field_data])	forward function which overloads torch.nn.Module
from_setting(setting)	Generate model from setting object
get_nn_name()	Return neural network name

Attributes

T_destination
call_super_init
dump_patches
training

forward(data, *, field_data=None, **kwards)

forward function which overloads 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)

Generate model from setting object

Parameters:

setting (SimilarityEquivariantMLPSetting) – setting object

Returns:

SimilarityEquivariantMLP object

Return type:

Self

classmethod get_nn_name()

Return neural network name

Returns:

name

Return type:

str