# -*- coding: utf-8 -*-
"""Implementation of MuRE."""
from typing import Any, ClassVar, Mapping, Optional
from torch.nn.init import normal_, uniform_, zeros_
from ..nbase import ERModel
from ...constants import DEFAULT_EMBEDDING_HPO_EMBEDDING_DIM_RANGE
from ...nn.modules import MuREInteraction
from ...typing import Hint, Initializer
__all__ = [
"MuRE",
]
[docs]class MuRE(ERModel):
r"""An implementation of MuRE from [balazevic2019b]_.
---
citation:
author: Balažević
year: 2019
link: https://arxiv.org/abs/1905.09791
"""
#: The default strategy for optimizing the model's hyper-parameters
hpo_default: ClassVar[Mapping[str, Any]] = dict(
embedding_dim=DEFAULT_EMBEDDING_HPO_EMBEDDING_DIM_RANGE,
p=dict(type=int, low=1, high=2),
)
def __init__(
self,
*,
embedding_dim: int = 200,
p: int = 2,
power_norm: bool = True,
entity_initializer: Hint[Initializer] = normal_,
entity_initializer_kwargs: Optional[Mapping[str, Any]] = None,
entity_bias_initializer: Hint[Initializer] = zeros_,
relation_initializer: Hint[Initializer] = normal_,
relation_initializer_kwargs: Optional[Mapping[str, Any]] = None,
relation_matrix_initializer: Hint[Initializer] = uniform_,
relation_matrix_initializer_kwargs: Optional[Mapping[str, Any]] = None,
**kwargs,
) -> None:
r"""Initialize MuRE via the :class:`pykeen.nn.modules.MuREInteraction` interaction.
:param embedding_dim: The entity embedding dimension $d$. Defaults to 200. Is usually $d \in [50, 300]$.
:param p: The $l_p$ norm. Defaults to 2.
:param power_norm: Should the power norm be used? Defaults to true.
:param entity_initializer: Entity initializer function. Defaults to :func:`torch.nn.init.normal_`
:param entity_initializer_kwargs: Keyword arguments to be used when calling the entity initializer
:param entity_bias_initializer: Entity bias initializer function. Defaults to :func:`torch.nn.init.zeros_`
:param relation_initializer: Relation initializer function. Defaults to :func:`torch.nn.init.normal_`
:param relation_initializer_kwargs: Keyword arguments to be used when calling the relation initializer
:param relation_matrix_initializer: Relation matrix initializer function.
Defaults to :func:`torch.nn.init.uniform_`
:param relation_matrix_initializer_kwargs: Keyword arguments to be used when calling the
relation matrix initializer
:param kwargs: Remaining keyword arguments passed through to :class:`pykeen.models.ERModel`.
"""
# comment:
# https://github.com/ibalazevic/multirelational-poincare/blob/34523a61ca7867591fd645bfb0c0807246c08660/model.py#L52
# uses float64
super().__init__(
interaction=MuREInteraction,
interaction_kwargs=dict(p=p, power_norm=power_norm),
entity_representations_kwargs=[
dict(
shape=embedding_dim,
initializer=entity_initializer,
initializer_kwargs=entity_initializer_kwargs
or dict(
std=1.0e-03,
),
),
# entity bias for head
dict(
shape=tuple(), # scalar
initializer=entity_bias_initializer,
),
# entity bias for tail
dict(
shape=tuple(), # scalar
initializer=entity_bias_initializer,
),
],
relation_representations_kwargs=[
# relation offset
dict(
shape=embedding_dim,
initializer=relation_initializer,
initializer_kwargs=relation_initializer_kwargs
or dict(
std=1.0e-03,
),
),
# diagonal relation transformation matrix
dict(
shape=embedding_dim,
initializer=relation_matrix_initializer,
initializer_kwargs=relation_matrix_initializer_kwargs
or dict(
a=-1,
b=1,
),
),
],
**kwargs,
)