"""Stateful interaction functions."""
from __future__ import annotations
import dataclasses
import itertools as itt
import logging
import math
from abc import ABC, abstractmethod
from collections import Counter
from collections.abc import Collection, Iterable, Mapping, MutableMapping, Sequence
from operator import itemgetter
from typing import Any, Callable, ClassVar, Generic, Optional, Union, cast, overload
import more_itertools
import numpy
import torch
from class_resolver import (
ClassResolver,
Hint,
LookupOrType,
OptionalKwargs,
ResolverKey,
update_docstring_with_resolver_keys,
)
from class_resolver.contrib.torch import activation_resolver
from docdata import parse_docdata
from torch import nn
from torch.nn.init import xavier_normal_
from typing_extensions import Self
from . import init, quaternion
from .sim import KG2ESimilarity, kg2e_similarity_resolver
from .utils import apply_optional_bn
from ..metrics.utils import ValueRange
from ..typing import (
FloatTensor,
GaussianDistribution,
HeadRepresentation,
HintOrType,
Initializer,
RelationRepresentation,
Representation,
Sign,
TailRepresentation,
)
from ..utils import (
add_cudnn_error_hint,
at_least_eps,
batched_dot,
circular_correlation,
clamp_norm,
einsum,
ensure_complex,
ensure_tuple,
estimate_cost_of_sequence,
make_ones_like,
negative_norm,
negative_norm_of_sum,
project_entity,
tensor_product,
tensor_sum,
unpack_singletons,
)
# TODO: split file into multiple smaller ones?
__all__ = [
"interaction_resolver",
# Base Classes
"Interaction",
"FunctionalInteraction",
"NormBasedInteraction",
# Adapter classes
"MonotonicAffineTransformationInteraction",
"ClampedInteraction",
"DirectionAverageInteraction",
# Concrete Classes
"AutoSFInteraction",
"BoxEInteraction",
"ComplExInteraction",
"ConvEInteraction",
"ConvKBInteraction",
"CPInteraction",
"CrossEInteraction",
"DistMAInteraction",
"DistMultInteraction",
"ERMLPEInteraction",
"ERMLPInteraction",
"HolEInteraction",
"KG2EInteraction",
"LineaREInteraction",
"MultiLinearTuckerInteraction",
"MuREInteraction",
"NTNInteraction",
"PairREInteraction",
"ProjEInteraction",
"QuatEInteraction",
"RESCALInteraction",
"RotatEInteraction",
"SEInteraction",
"SimplEInteraction",
"TorusEInteraction",
"TransDInteraction",
"TransEInteraction",
"TransFInteraction",
"TransformerInteraction",
"TransHInteraction",
"TransRInteraction",
"TripleREInteraction",
"TuckERInteraction",
"UMInteraction",
]
logger = logging.getLogger(__name__)
def parallel_slice_batches(
*representations: Representation,
split_size: int,
dim: int,
) -> Iterable[Sequence[Representation]]:
"""
Slice representations along the given dimension.
:param representations:
the representations to slice
:param split_size:
the slice size
:param dim:
the dimension along which to slice
:yields: batches of sliced representations
"""
# normalize input
rs: Sequence[Sequence[FloatTensor]] = ensure_tuple(*representations)
# get number of head/relation/tail representations
length = list(map(len, rs))
splits = numpy.cumsum([0] + length)
# flatten list
rsl: Sequence[FloatTensor] = sum(map(list, rs), [])
# split tensors
parts = [r.split(split_size, dim=dim) for r in rsl]
# broadcasting
n_parts = max(map(len, parts))
parts = [r_parts if len(r_parts) == n_parts else r_parts * n_parts for r_parts in parts]
# yield batches
for batch in zip(*parts):
# complex typing
yield unpack_singletons(*(batch[start:stop] for start, stop in zip(splits, splits[1:]))) # type: ignore
# docstr-coverage:excused `overload`
@overload
def parallel_unsqueeze(x: Sequence[FloatTensor], dim: int) -> Sequence[FloatTensor]: ...
# docstr-coverage:excused `overload`
@overload
def parallel_unsqueeze(x: FloatTensor, dim: int) -> FloatTensor: ...
def parallel_unsqueeze(x: FloatTensor | Sequence[FloatTensor], dim: int) -> FloatTensor | Sequence[FloatTensor]:
"""Unsqueeze all representations along the given dimension."""
if not isinstance(x, Sequence):
return x.unsqueeze(dim=dim)
return cast(Sequence[FloatTensor], [xx.unsqueeze(dim=dim) for xx in x])
[docs]
class Interaction(nn.Module, Generic[HeadRepresentation, RelationRepresentation, TailRepresentation], ABC):
"""Base class for interaction functions."""
#: The symbolic shapes for entity representations
entity_shape: Sequence[str] = ("d",)
#: The symbolic shapes for relation representations
relation_shape: Sequence[str] = ("d",)
# if the interaction function's head parameter should only receive a subset of entity representations
_head_indices: Sequence[int] | None = None
# if the interaction function's tail parameter should only receive a subset of entity representations
_tail_indices: Sequence[int] | None = None
# TODO: does not seem to be used
#: the interaction's value range (for unrestricted input)
value_range: ClassVar[ValueRange] = ValueRange()
# TODO: annotate modelling capabilities? cf., e.g., https://arxiv.org/abs/1902.10197, Table 2
# TODO: annotate properties, e.g., symmetry, and use them for testing?
# TODO: annotate complexity?
#: whether the interaction is defined on complex input
is_complex: ClassVar[bool] = False
@property
def head_shape(self) -> Sequence[str]:
"""Return the symbolic shape for head entity representations."""
if self._head_indices is None:
return self.entity_shape
return [self.entity_shape[i] for i in self._head_indices]
@property
def tail_shape(self) -> Sequence[str]:
"""Return the symbolic shape for tail entity representations."""
if self._tail_indices is None:
return self.entity_shape
return [self.entity_shape[i] for i in self._tail_indices]
@property
def head_indices(self) -> Sequence[int]:
"""Return the entity representation indices used for the head representations."""
if self._head_indices is None:
return range(len(self.entity_shape))
return self._head_indices
@property
def tail_indices(self) -> Sequence[int]:
"""Return the entity representation indices used for the tail representations."""
if self._tail_indices is None:
return range(len(self.tail_shape))
return self._tail_indices
@property
def dimensions(self) -> set[str]:
"""Get all the relevant dimension keys.
This draws from :data:`Interaction.entity_shape`, and :data:`Interaction.relation_shape`.
:returns: a set of strings representing the dimension keys.
"""
return set(itt.chain(self.entity_shape, self.relation_shape))
[docs]
@abstractmethod
def forward(
self,
h: HeadRepresentation,
r: RelationRepresentation,
t: TailRepresentation,
) -> FloatTensor:
"""Compute broadcasted triple scores given broadcasted representations for head, relation and tails.
In general, each interaction function (class) expects a certain format for each of head, relation and
tail representations. This format is composed of the *number* and the shape of the representations.
Many simple interaction functions such as :class:`~pykeen.nn.modules.TransEInteraction`
operate on a single representation, however there are also interactions such as
:class:`~pykeen.nn.modules.TransDInteraction`, which requires two representations for each slot, or
:class:`~pykeen.nn.modules.PairREInteraction`, which requires two relation representations, but only a single
representation for head and tail entity respectively.
Each individual representation has a *shape*. This can be a simple $d$-dimensional vector, but also comprise
matrices, or even high-order tensors.
This method supports the general batched calculation, i.e., each of the representations can have a
preceding batch dimensions. Those batch dimensions do not necessarily need to be exactly the same, but they
need to be broadcastable. A good explanation of broadcasting rules can be found in
`NumPy's documentation <https://numpy.org/doc/stable/user/basics.broadcasting.html>`_.
.. seealso::
- :ref:`representations` for an overview about different ways how to obtain individual representations.
:param h: shape: ``(*batch_dims, *dims)``
The head representations.
:param r: shape: ``(*batch_dims, *dims)``
The relation representations.
:param t: shape: ``(*batch_dims, *dims)``
The tail representations.
:return: shape: batch_dims
The scores.
"""
[docs]
def score(
self,
h: HeadRepresentation,
r: RelationRepresentation,
t: TailRepresentation,
slice_size: int | None = None,
slice_dim: int = 1,
) -> FloatTensor:
"""Compute broadcasted triple scores with optional slicing.
.. note ::
At most one of the slice sizes may be not None.
.. todo::
we could change that to slicing along multiple dimensions, if necessary
:param h: shape: (`*batch_dims`, `*dims`)
The head representations.
:param r: shape: (`*batch_dims`, `*dims`)
The relation representations.
:param t: shape: (`*batch_dims`, `*dims`)
The tail representations.
:param slice_size:
The slice size.
:param slice_dim:
The dimension along which to slice. From {0, ..., len(batch_dims)}
:return: shape: batch_dims
The scores.
"""
if slice_size is None:
return self(h=h, r=r, t=t)
return torch.cat(
[
self(h=h_batch, r=r_batch, t=t_batch)
for h_batch, r_batch, t_batch in parallel_slice_batches(h, r, t, split_size=slice_size, dim=slice_dim)
],
dim=slice_dim,
)
[docs]
def score_hrt(
self,
h: HeadRepresentation,
r: RelationRepresentation,
t: TailRepresentation,
) -> FloatTensor:
"""Score a batch of triples.
:param h: shape: (batch_size, d_e)
The head representations.
:param r: shape: (batch_size, d_r)
The relation representations.
:param t: shape: (batch_size, d_e)
The tail representations.
:return: shape: (batch_size, 1)
The scores.
"""
return self.score(h=h, r=r, t=t).unsqueeze(dim=-1)
[docs]
def score_h(
self,
all_entities: HeadRepresentation,
r: RelationRepresentation,
t: TailRepresentation,
slice_size: int | None = None,
) -> FloatTensor:
"""Score all head entities.
:param all_entities: shape: (num_entities, d_e)
The head representations.
:param r: shape: (batch_size, d_r)
The relation representations.
:param t: shape: (batch_size, d_e)
The tail representations.
:param slice_size:
The slice size.
:return: shape: (batch_size, num_entities)
The scores.
"""
return self.score(
h=parallel_unsqueeze(all_entities, dim=0),
r=parallel_unsqueeze(r, dim=1),
t=parallel_unsqueeze(t, dim=1),
slice_size=slice_size,
)
[docs]
def score_r(
self,
h: HeadRepresentation,
all_relations: RelationRepresentation,
t: TailRepresentation,
slice_size: int | None = None,
) -> FloatTensor:
"""Score all relations.
:param h: shape: (batch_size, d_e)
The head representations.
:param all_relations: shape: (num_relations, d_r)
The relation representations.
:param t: shape: (batch_size, d_e)
The tail representations.
:param slice_size:
The slice size.
:return: shape: (batch_size, num_entities)
The scores.
"""
return self.score(
h=parallel_unsqueeze(h, dim=1),
r=parallel_unsqueeze(all_relations, dim=0),
t=parallel_unsqueeze(t, dim=1),
slice_size=slice_size,
)
[docs]
def score_t(
self,
h: HeadRepresentation,
r: RelationRepresentation,
all_entities: TailRepresentation,
slice_size: int | None = None,
) -> FloatTensor:
"""Score all tail entities.
:param h: shape: (batch_size, d_e)
The head representations.
:param r: shape: (batch_size, d_r)
The relation representations.
:param all_entities: shape: (num_entities, d_e)
The tail representations.
:param slice_size:
The slice size.
:return: shape: (batch_size, num_entities)
The scores.
"""
return self.score(
h=parallel_unsqueeze(h, dim=1),
r=parallel_unsqueeze(r, dim=1),
t=parallel_unsqueeze(all_entities, dim=0),
slice_size=slice_size,
)
[docs]
def reset_parameters(self):
"""Reset parameters the interaction function may have."""
for mod in self.modules():
if mod is self:
continue
if hasattr(mod, "reset_parameters"):
mod.reset_parameters()
[docs]
class FunctionalInteraction(Interaction, Generic[HeadRepresentation, RelationRepresentation, TailRepresentation]):
"""Base class for interaction functions."""
#: The functional interaction form
func: Callable[..., FloatTensor]
[docs]
def forward(
self,
h: HeadRepresentation,
r: RelationRepresentation,
t: TailRepresentation,
) -> FloatTensor:
"""Compute broadcasted triple scores given broadcasted representations for head, relation and tails.
:param h: shape: (`*batch_dims`, `*dims`)
The head representations.
:param r: shape: (`*batch_dims`, `*dims`)
The relation representations.
:param t: shape: (`*batch_dims`, `*dims`)
The tail representations.
:return: shape: batch_dims
The scores.
"""
return self.__class__.func(**self._prepare_for_functional(h=h, r=r, t=t))
def _prepare_for_functional(
self,
h: HeadRepresentation,
r: RelationRepresentation,
t: TailRepresentation,
) -> Mapping[str, FloatTensor]:
"""Conversion utility to prepare the arguments for the functional form."""
kwargs = self._prepare_hrt_for_functional(h=h, r=r, t=t)
kwargs.update(self._prepare_state_for_functional())
return kwargs
# docstr-coverage: inherited
@classmethod
def _prepare_hrt_for_functional(
cls,
h: HeadRepresentation,
r: RelationRepresentation,
t: TailRepresentation,
) -> MutableMapping[str, FloatTensor]: # noqa: D102
"""Conversion utility to prepare the h/r/t representations for the functional form."""
# TODO: we only allow single-tensor representations here, but could easily generalize
assert all(torch.is_tensor(x) for x in (h, r, t))
if cls.is_complex:
h, r, t = ensure_complex(h, r, t)
return dict(h=h, r=r, t=t)
def _prepare_state_for_functional(self) -> MutableMapping[str, Any]:
"""Conversion utility to prepare the state to be passed to the functional form."""
return dict()
[docs]
class NormBasedInteraction(Interaction, Generic[HeadRepresentation, RelationRepresentation, TailRepresentation], ABC):
"""Norm-based interactions use a (powered) $p$-norm in their scoring function."""
def __init__(self, p: int, power_norm: bool = False):
"""Initialize the norm-based interaction function.
:param p:
The norm used with :func:`torch.linalg.vector_norm`. Typically is 1 or 2.
:param power_norm:
Whether to use the p-th power of the $L_p$ norm. It has the advantage of being differentiable around 0,
and numerically more stable.
"""
super().__init__()
self.p = p
self.power_norm = power_norm
# docstr-coverage: inherited
def _prepare_state_for_functional(self) -> MutableMapping[str, Any]: # noqa: D102
raise AssertionError("This is a relic.")
[docs]
@parse_docdata
class TransEInteraction(NormBasedInteraction[FloatTensor, FloatTensor, FloatTensor]):
r"""The state-less norm-based TransE interaction function.
TransE models relations as a translation from head to tail entities in :math:`\textbf{e}`:
.. math::
\textbf{e}_h + \textbf{e}_r \approx \textbf{e}_t
This equation is rearranged and the :math:`l_p` norm is applied to create the TransE interaction function.
.. math::
f(h, r, t) = - \|\textbf{e}_h + \textbf{e}_r - \textbf{e}_t\|_{p}
While this formulation is computationally efficient, it inherently cannot model one-to-many, many-to-one, and
many-to-many relationships. For triples :math:`(h,r,t_1), (h,r,t_2) \in \mathcal{K}` where :math:`t_1 \neq t_2`,
the model adapts the embeddings in order to ensure :math:`\textbf{e}_h + \textbf{e}_r \approx \textbf{e}_{t_1}`
and :math:`\textbf{e}_h + \textbf{e}_r \approx \textbf{e}_{t_2}` which results in
:math:`\textbf{e}_{t_1} \approx \textbf{e}_{t_2}`.
---
citation:
author: Bordes
year: 2013
link: http://papers.nips.cc/paper/5071-translating-embeddings-for-modeling-multi-relational-data.pdf
"""
[docs]
def forward(self, h: FloatTensor, r: FloatTensor, t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
return negative_norm_of_sum(h, r, -t, p=self.p, power_norm=self.power_norm)
[docs]
@parse_docdata
class TransFInteraction(Interaction[FloatTensor, FloatTensor, FloatTensor]):
r"""The state-less norm-based TransF interaction function.
It is given by
.. math ::
f(\mathbf{h}, \mathbf{r}, \mathbf{t}) =
(\mathbf{h} + \mathbf{r})^T \mathbf{t} + \mathbf{h}^T (\mathbf{r} - \mathbf{t})
for head entity, relation, and tail entity representations $\mathbf{h}, \mathbf{r}, \mathbf{t} \in \mathbb{R}^d$.
The interaction function can be simplified as
.. math ::
f(\mathbf{h}, \mathbf{r}, \mathbf{t}) &=&
(\mathbf{h} + \mathbf{r})^T \mathbf{t} + \mathbf{h}^T (\mathbf{t} - \mathbf{r}) \\
&=&
\langle \mathbf{h}, \mathbf{t}\rangle
+ \langle \mathbf{r}, \mathbf{t}\rangle
+ \langle \mathbf{h}, \mathbf{t}\rangle
- \langle \mathbf{h}, \mathbf{r}\rangle \\
&=&
2 \cdot \langle \mathbf{h}, \mathbf{t}\rangle
+ \langle \mathbf{r}, \mathbf{t}\rangle
- \langle \mathbf{h}, \mathbf{r}\rangle
.. note ::
This is the *balanced* variant from the paper.
.. todo ::
Implement the unbalanced version, too:
$f(\mathbf{h}, \mathbf{r}, \mathbf{t}) = (\mathbf{h} + \mathbf{r})^T \mathbf{t}$
---
citation:
author: Feng
year: 2016
link: https://www.aaai.org/ocs/index.php/KR/KR16/paper/view/12887
arxiv: 1505.05253
"""
# TODO: implement the unbalanced variant from the paper: f(h, r, t) = (h + r)^T t
[docs]
def forward(self, h: FloatTensor, r: FloatTensor, t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
return tensor_sum(2 * batched_dot(h, t), batched_dot(r, t), -batched_dot(h, r))
[docs]
@parse_docdata
class ComplExInteraction(FunctionalInteraction[FloatTensor, FloatTensor, FloatTensor]):
r"""The ComplEx interaction proposed by [trouillon2016]_.
ComplEx operates on complex-valued entity and relation representations, i.e.,
$\textbf{e}_i, \textbf{r}_i \in \mathbb{C}^d$ and calculates the plausibility score via the Hadamard product:
.. math::
f(h,r,t) = Re(\mathbf{e}_h\odot\mathbf{r}_r\odot\bar{\mathbf{e}}_t)
Which expands to:
.. math::
f(h,r,t) = \left\langle Re(\mathbf{e}_h),Re(\mathbf{r}_r),Re(\mathbf{e}_t)\right\rangle
+ \left\langle Im(\mathbf{e}_h),Re(\mathbf{r}_r),Im(\mathbf{e}_t)\right\rangle
+ \left\langle Re(\mathbf{e}_h),Im(\mathbf{r}_r),Im(\mathbf{e}_t)\right\rangle
- \left\langle Im(\mathbf{e}_h),Im(\mathbf{r}_r),Re(\mathbf{e}_t)\right\rangle
where $Re(\textbf{x})$ and $Im(\textbf{x})$ denote the real and imaginary parts of the complex valued vector
$\textbf{x}$. Because the Hadamard product is not commutative in the complex space, ComplEx can model
anti-symmetric relations in contrast to DistMult.
.. seealso ::
Official implementation: https://github.com/ttrouill/complex/
.. note::
this method generally expects all tensors to be of complex datatype, i.e., `torch.is_complex(x)` to evaluate to
`True`. However, for backwards compatibility and convenience in use, you can also pass real tensors whose shape
is compliant with :func:`torch.view_as_complex`, cf. :func:`pykeen.utils.ensure_complex`.
---
citation:
arxiv: 1606.06357
author: Trouillon
github: ttrouill/complex
link: https://arxiv.org/abs/1606.06357
year: 2016
"""
is_complex: ClassVar[bool] = True
# TODO: update class docstring
# TODO: give this a better name?
[docs]
@staticmethod
def func(h: FloatTensor, r: FloatTensor, t: FloatTensor) -> FloatTensor:
r"""Evaluate the interaction function.
:param h: shape: (`*batch_dims`, dim)
The complex head representations.
:param r: shape: (`*batch_dims`, dim)
The complex relation representations.
:param t: shape: (`*batch_dims`, dim)
The complex tail representations.
:return: shape: batch_dims
The scores.
"""
return torch.real(einsum("...d, ...d, ...d -> ...", h, r, torch.conj(t)))
@dataclasses.dataclass
class ConvEResolvedImageShape:
"""The resolved shape of the ConvE 'image'."""
dim: int
width: int
height: int
channels: int
@property
def is_valid(self) -> bool:
"""Determine whether the given shape is a valid factorization of the embedding dimension."""
return self.channels * self.width * self.height == self.dim
@classmethod
def make(cls, channels: int | None, dim: int | None, height: int | None, width: int | None) -> Self:
"""
Automatically calculates missing dimensions for ConvE.
The dimensions need to fulfil $channels * height * width = dim$.
:param channels:
the number of input channels
:param dim:
the embedding dimension
:param height:
the "image" height
:param width:
the "image" width
:return:
a resolve shape information.
:raises ValueError:
when the constraints cannot be satisfied.
"""
if dim is None:
if channels is None or width is None or height is None:
raise ValueError(
f"When {dim=} none of the other dimensions may be None, "
f"but {channels=}, {width=}, and {height=}"
)
dim = channels * width * height
# All are None -> try and make closest to square
if channels is None and width is None and height is None:
result_sqrt = math.floor(math.sqrt(dim))
height = max(factor for factor in range(1, result_sqrt + 1) if dim % factor == 0)
width = dim // height
return cls(dim=dim, width=width, height=height, channels=1)
# Only input channels is None
if channels is None and width is not None and height is not None:
return cls(dim=dim, width=width, height=height, channels=dim // (width * height))
# Only width is None
if channels is not None and width is None and height is not None:
return cls(dim=dim, width=dim // (height * channels), height=height, channels=channels)
# Only height is none
if height is None and width is not None and channels is not None:
return cls(dim=dim, width=width, height=dim // (width * channels), channels=channels)
# Height and input_channels are None -> set input_channels to 1 and calculage height
if channels is None and height is None and width is not None:
return cls(dim=dim, width=width, height=dim // width, channels=1)
# Width and input channels are None -> set input channels to 1 and calculate width
if channels is None and height is not None and width is None:
return cls(dim=dim, width=dim // height, height=height, channels=1)
raise ValueError(f"Could not resolve {channels=}, {height=}, {width=} = {dim=}.")
@dataclasses.dataclass
class ConvEShapeInformation:
"""Resolved ConvE shape information."""
#: the embedding dimension
embedding_dim: int
#: the number of input channels of the convolution
input_channels: int
#: the embedding "image" height
image_height: int
#: the embedding "image" width
image_width: int
#: the number of output channels of the convolution
output_channels: int
#: the convolution kernel height
kernel_height: int
#: the convolution kernel width
kernel_width: int
@property
def num_in_features(self) -> int:
"""The number of input features to the linear layer."""
return (
self.output_channels
* (2 * self.image_height - self.kernel_height + 1)
* (self.image_width - self.kernel_width + 1)
)
@classmethod
def make(
cls,
embedding_dim: int | None,
image_width: int | None = None,
image_height: int | None = None,
input_channels: int | None = None,
output_channels: int = 32,
kernel_width: int = 3,
kernel_height: int | None = None,
) -> Self:
"""Automatically calculates missing dimensions for ConvE.
:param embedding_dim:
The embedding dimension.
:param image_width:
The width of the embedding "image".
:param image_height:
The height of the embedding "image".
:param input_channels:
The number of input channels for the convolution.
:param output_channels:
The number of output channels for the convolution.
:param kernel_width:
The width of the convolution kernel.
:param kernel_height:
The height of the convolution kernel.
:return: Fully resolve shapes.
:raises ValueError:
If no factorization could be found.
"""
# resolve image shape
logger.info(f"Resolving {input_channels} * {image_width} * {image_height} = {embedding_dim}.")
# Store initial input for error message
original = (input_channels, image_width, image_height)
# infer open dimensions from the remainder
image_shape = ConvEResolvedImageShape.make(
dim=embedding_dim,
height=image_height,
width=image_width,
channels=input_channels,
)
if not image_shape.is_valid:
raise ValueError(f"Could not resolve {original} to a valid factorization of {embedding_dim}.")
# resolve kernel size defaults
kernel_height = kernel_height or kernel_width
return cls(
embedding_dim=image_shape.dim,
input_channels=image_shape.channels,
image_width=image_shape.width,
image_height=image_shape.height,
kernel_height=kernel_height,
kernel_width=kernel_width,
output_channels=output_channels,
)
[docs]
@parse_docdata
class ConvEInteraction(Interaction[FloatTensor, FloatTensor, tuple[FloatTensor, FloatTensor]]):
r"""The stateful ConvE interaction function.
ConvE is a CNN-based approach. For input representations $\mathbf{h}, \mathbf{r}, \mathbf{t} \in \mathbb{R}^d$,
it first combines $\mathbf{h}$ and $\mathbf{r}$ into a matrix matrix $\mathbf{A} \in \mathbb{R}^{2 \times d}$,
where the first row of $\mathbf{A}$ represents $\mathbf{h}$ and the second row represents $\mathbf{r}$.
$\mathbf{A}$ is reshaped to a matrix $\mathbf{B} \in \mathbb{R}^{m \times n}$
where the first $m/2$ half rows represent $\mathbf{h}$ and the remaining $m/2$ half rows represent $\mathbf{r}$.
In the convolution layer, a set of *2-dimensional* convolutional filters
$\Omega = \{\omega_i \mid \omega_i \in \mathbb{R}^{r \times c}\}$ are applied on $\mathbf{B}$
that capture interactions between $\mathbf{h}$ and $\mathbf{r}$.
The resulting feature maps are reshaped and concatenated in order to create a feature vector
$\mathbf{v} \in \mathbb{R}^{|\Omega|rc}$.
In the next step, $\mathbf{v}$ is mapped into the entity space using a linear transformation
$\mathbf{W} \in \mathbb{R}^{|\Omega|rc \times d}$, that is $\mathbf{e}_{h,r} = \mathbf{v}^{T} \mathbf{W}$.
The score is then obtained by:
.. math::
f(\mathbf{h}, \mathbf{r}, \mathbf{t}) = \mathbf{e}_{h,r} \mathbf{t}
Since the interaction model can be decomposed into
$f(\mathbf{h}, \mathbf{r}, \mathbf{t}) = \left\langle f'(\mathbf{h}, \mathbf{r}), \mathbf{t} \right\rangle$
the model is particularly designed to 1-N scoring, i.e. efficient computation of scores for
$(h,r,t)$ for fixed $h,r$ and many different $t$.
The default setting uses batch normalization. Batch normalization normalizes the output of the activation functions,
in order to ensure that the weights of the NN don't become imbalanced and to speed up training.
However, batch normalization is not the only way to achieve more robust and effective training [santurkar2018]_.
Therefore, we added the flag ``apply_batch_normalization`` to turn batch normalization on/off (it's turned on as
default).
---
citation:
author: Dettmers
year: 2018
link: https://www.aaai.org/ocs/index.php/AAAI/AAAI18/paper/view/17366
github: TimDettmers/ConvE
arxiv: 1707.01476
"""
# vector & scalar offset
entity_shape = ("d", "")
# the offset is only used for tails
_head_indices = (0,)
#: The head-relation encoder operating on 2D "images"
hr2d: nn.Module
#: The head-relation encoder operating on the 1D flattened version
hr1d: nn.Module
def __init__(
self,
input_channels: int | None = None,
output_channels: int = 32,
embedding_height: int | None = None,
embedding_width: int | None = None,
kernel_width: int = 3,
kernel_height: int | None = None,
input_dropout: float = 0.2,
feature_map_dropout: float = 0.2,
output_dropout: float = 0.3,
embedding_dim: int = 200,
apply_batch_normalization: bool = True,
):
"""
Initialize the interaction module.
:param input_channels:
the number of input channels for the convolution operation. Can be inferred from other parameters,
cf. :func:`_calculate_missing_shape_information`.
:param output_channels:
the number of input channels for the convolution operation
:param embedding_height:
the height of the "image" after reshaping the concatenated head and relation embedding. Can be inferred
from other parameters, cf. :func:`_calculate_missing_shape_information`.
:param embedding_width:
the width of the "image" after reshaping the concatenated head and relation embedding. Can be inferred
from other parameters, cf. :func:`_calculate_missing_shape_information`.
:param kernel_width:
the width of the convolution kernel
:param kernel_height:
the height of the convolution kernel. Defaults to `kernel_width`
:param input_dropout:
the dropout applied *before* the convolution
:param feature_map_dropout:
the dropout applied *after* the convolution
:param output_dropout:
the dropout applied after the linear projection
:param embedding_dim:
the embedding dimension of entities and relations
:param apply_batch_normalization:
whether to apply batch normalization
"""
super().__init__()
# Parameter need to fulfil:
# input_channels * embedding_height * embedding_width = embedding_dim
self.shape_info = ConvEShapeInformation.make(
embedding_dim=embedding_dim,
input_channels=input_channels,
image_width=embedding_width,
image_height=embedding_height,
kernel_width=kernel_width,
kernel_height=kernel_height,
output_channels=output_channels,
)
# encoders
# 1: 2D encoder: BN?, DO, Conv, BN?, Act, DO
hr2d_layers = [
nn.BatchNorm2d(self.shape_info.input_channels) if apply_batch_normalization else None,
nn.Dropout(input_dropout),
nn.Conv2d(
in_channels=self.shape_info.input_channels,
out_channels=self.shape_info.output_channels,
kernel_size=(self.shape_info.kernel_height, self.shape_info.kernel_width),
stride=1,
padding=0,
bias=True,
),
nn.BatchNorm2d(self.shape_info.output_channels) if apply_batch_normalization else None,
nn.ReLU(),
nn.Dropout2d(feature_map_dropout),
]
self.hr2d = nn.Sequential(*(layer for layer in hr2d_layers if layer is not None))
# 2: 1D encoder: FC, DO, BN?, Act
hr1d_layers = [
nn.Linear(self.shape_info.num_in_features, self.shape_info.embedding_dim),
nn.Dropout(output_dropout),
nn.BatchNorm1d(self.shape_info.embedding_dim) if apply_batch_normalization else None,
nn.ReLU(),
]
self.hr1d = nn.Sequential(*(layer for layer in hr1d_layers if layer is not None))
[docs]
@add_cudnn_error_hint
def forward(
self,
h: FloatTensor,
r: FloatTensor,
t: tuple[FloatTensor, FloatTensor],
) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``
The relation representations.
:param t: two vectors of shape: ``(*batch_dims, d)`` and ``batch_dims``
The tail representations, comprising the tail entity embedding and bias.
:return: shape: ``batch_dims``
The scores.
"""
t_emb, t_bias = t
# repeat if necessary, and concat head and relation
# shape: -1, num_input_channels, 2*height, width
x = torch.cat(
torch.broadcast_tensors(
h.view(
*h.shape[:-1],
self.shape_info.input_channels,
self.shape_info.image_height,
self.shape_info.image_width,
),
r.view(
*r.shape[:-1],
self.shape_info.input_channels,
self.shape_info.image_height,
self.shape_info.image_width,
),
),
dim=-2,
)
prefix_shape = x.shape[:-3]
x = x.view(-1, self.shape_info.input_channels, 2 * self.shape_info.image_height, self.shape_info.image_width)
# shape: -1, num_input_channels, 2*height, width
x = self.hr2d(x)
# -1, num_output_channels * (2 * height - kernel_height + 1) * (width - kernel_width + 1)
x = x.view(-1, self.shape_info.num_in_features)
x = self.hr1d(x)
# reshape: (-1, dim) -> (*batch_dims, dim)
x = x.view(*prefix_shape, h.shape[-1])
# For efficient calculation, each of the convolved [h, r] rows has only to be multiplied with one t row
# output_shape: batch_dims
x = batched_dot(x, t_emb)
# add bias term
return x + t_bias
[docs]
@parse_docdata
class ConvKBInteraction(Interaction[FloatTensor, FloatTensor, FloatTensor]):
r"""The stateful ConvKB interaction function.
ConvKB uses a convolutional neural network (CNN) whose feature maps capture global interactions of the input.
For given input representations for head entity, relation and tail entity, denoted by
$\mathbf{h}, \mathbf{r}, \mathbf{t} \in \mathbb{R}^d$, it first combines them to a matrix
$\mathbf{A} = [\mathbf{h}; \mathbf{r}; \mathbf{t}] \in \mathbb{R}^{d \times 3}$.
In the convolution layer, a set of convolutional filters
$\omega_i \in \mathbb{R}^{1 \times 3}$, $i=1, \dots, \tau,$ are applied on the input in order to compute for
each dimension global interactions of the embedded triple. Each $\omega_i$ is applied on every row of
$\mathbf{A}$ creating a feature map $\mathbf{v}_i = [v_{i,1},...,v_{i,d}] \in \mathbb{R}^d$:
.. math::
\mathbf{v}_i = g(\omega_j \mathbf{A} + \mathbf{b})
where $\mathbf{b} \in \mathbb{R}$ denotes a bias term and $g$ an activation function which is employed element-wise.
Based on the resulting feature maps $\mathbf{v}_1, \dots, \mathbf{v}_{\tau}$, the plausibility score of a triple
is given by:
.. math::
f(h,r,t) = [\mathbf{v}_i; \ldots ;\mathbf{v}_\tau] \cdot \mathbf{w}
where $[\mathbf{v}_i; \ldots ;\mathbf{v}_\tau] \in \mathbb{R}^{\tau d \times 1}$ and
$\mathbf{w} \in \mathbb{R}^{\tau d \times 1}$ is a shared weight vector.
ConvKB may be seen as a restriction of :class:`~pykeen.nn.modules.ERMLPInteraction` with a certain weight sharing
pattern in the first layer.
---
citation:
author: Nguyen
year: 2018
link: https://www.aclweb.org/anthology/N18-2053
github: daiquocnguyen/ConvKB
arxiv: 1712.02121
"""
def __init__(
self,
hidden_dropout_rate: float = 0.0,
embedding_dim: int = 200,
num_filters: int = 400,
):
"""
Initialize the interaction module.
:param hidden_dropout_rate:
the dropout rate applied on the hidden layer
:param embedding_dim:
the entity and relation embedding dimension
:param num_filters:
the number of filters (=output channels) of the convolution
"""
super().__init__()
self.embedding_dim = embedding_dim
self.num_filters = num_filters
# The interaction model
self.conv = nn.Conv2d(in_channels=1, out_channels=num_filters, kernel_size=(1, 3), bias=True)
self.activation = nn.ReLU()
self.hidden_dropout = nn.Dropout(p=hidden_dropout_rate)
self.linear = nn.Linear(embedding_dim * num_filters, 1, bias=True)
# docstr-coverage: inherited
[docs]
def reset_parameters(self): # noqa: D102
# Use Xavier initialization for weight; bias to zero
nn.init.xavier_uniform_(self.linear.weight, gain=nn.init.calculate_gain("relu"))
nn.init.zeros_(self.linear.bias)
# Initialize all filters to [0.1, 0.1, -0.1],
# c.f. https://github.com/daiquocnguyen/ConvKB/blob/master/model.py#L34-L36
nn.init.constant_(self.conv.weight[..., :2], 0.1)
nn.init.constant_(self.conv.weight[..., 2], -0.1)
nn.init.zeros_(self.conv.bias)
[docs]
def forward(self, h: FloatTensor, r: FloatTensor, t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
# cat into shape (..., 1, d, 3)
x = torch.stack(torch.broadcast_tensors(h, r, t), dim=-1).unsqueeze(dim=-3)
s = x.shape
x = x.view(-1, *s[-3:])
x = self.conv(x)
x = x.view(*s[:-3], -1)
x = self.activation(x)
# Apply dropout, cf. https://github.com/daiquocnguyen/ConvKB/blob/master/model.py#L54-L56
x = self.hidden_dropout(x)
# Linear layer for final scores; use flattened representations, shape: (*batch_dims, d * f)
x = self.linear(x)
return x.squeeze(dim=-1)
[docs]
@parse_docdata
class DistMultInteraction(FunctionalInteraction[FloatTensor, FloatTensor, FloatTensor]):
r"""The stateless DistMult interaction function.
This interaction is given by
.. math::
f(\mathbf{h}, \mathbf{r}, \mathbf{t}) = \sum \limits_{i} \mathbf{h}_i \cdot \mathbf{r}_{i} \cdot \mathbf{t}_i
where $\mathbf{h}, \mathbf{r}, \mathbf{t} \in \mathbb{R}^{d}$ are the representations for the head entity,
the relation, and the tail entity.
For a single triple of $d$-dimensional vectors, the computational complexity is given as $\mathcal{O}(d)$.
The interaction function is symmetric in the entities, i.e.,
.. math::
f(h, r, t) = f(t, r, h)
---
citation:
author: Yang
year: 2014
link: https://arxiv.org/abs/1412.6575
arxiv: 1412.6575
"""
[docs]
@staticmethod
def func(h: FloatTensor, r: FloatTensor, t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
return tensor_product(h, r, t).sum(dim=-1)
[docs]
@parse_docdata
class DistMAInteraction(FunctionalInteraction[FloatTensor, FloatTensor, FloatTensor]):
r"""The stateless DistMA interaction function from [shi2019]_.
For head entity, relation, and tail representations $\mathbf{h}, \mathbf{r}, \mathbf{t} \in \mathbb{R}^d$,
the interaction functions is given by
.. math ::
\langle \mathbf{h}, \mathbf{r}\rangle
+ \langle \mathbf{r}, \mathbf{t}\rangle
+ \langle \mathbf{h}, \mathbf{t}\rangle
.. note ::
This interaction function is the symmetric part $E_1$ from the respective paper, and not the combination
with :class:`~pykeen.nn.modules.ComplExInteraction`.
---
citation:
author: Shi
year: 2019
link: https://www.aclweb.org/anthology/D19-1075.pdf
"""
[docs]
@staticmethod
def func(h: FloatTensor, r: FloatTensor, t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
return batched_dot(h, r) + batched_dot(r, t) + batched_dot(h, t)
[docs]
@parse_docdata
class ERMLPInteraction(Interaction[FloatTensor, FloatTensor, FloatTensor]):
r"""The ER-MLP stateful interaction function.
ER-MLP uses a multi-layer perceptron based approach with a single hidden layer.
The $d$-dimensional representations of head entity, relation, and tail entity are concatenated
and passed to the hidden layer. The output-layer consists of a single neuron that computes the plausibility score:
.. math::
f(\mathbf{h}, \mathbf{r}, \mathbf{t}) = \mathbf{w}^{T} g(\mathbf{W} [\mathbf{h}; \mathbf{r}; \mathbf{t}]),
where $\textbf{W} \in \mathbb{R}^{k \times 3d}$ represents the weight matrix of the hidden layer,
$\textbf{w} \in \mathbb{R}^{k}$, the weights of the output layer, and $g$ denotes an activation function such
as the hyperbolic tangent.
---
name: ER-MLP
citation:
author: Dong
year: 2014
link: https://storage.googleapis.com/pub-tools-public-publication-data/pdf/45634.pdf
"""
@update_docstring_with_resolver_keys(
ResolverKey(name="activation", resolver="class_resolver.contrib.torch.activation_resolver")
)
def __init__(
self,
embedding_dim: int,
hidden_dim: int | None = None,
activation: HintOrType[nn.Module] = nn.ReLU,
activation_kwargs: OptionalKwargs = None,
):
"""Initialize the interaction module.
:param embedding_dim:
The embedding vector dimension for entities and relations.
:param hidden_dim:
The hidden dimension of the MLP. Defaults to `embedding_dim`.
:param activation:
The activation function or a hint thereof.
:param activation_kwargs:
Additional keyword-based parameters passed to the activation's constructor, if the activation is not
pre-instantiated.
"""
super().__init__()
# normalize hidden_dim
hidden_dim = hidden_dim or embedding_dim
self.hidden = nn.Linear(in_features=3 * embedding_dim, out_features=hidden_dim, bias=True)
self.activation = activation_resolver.make(activation, activation_kwargs)
self.hidden_to_score = nn.Linear(in_features=hidden_dim, out_features=1, bias=True)
[docs]
def forward(self, h: FloatTensor, r: FloatTensor, t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
# shortcut for same shape
if h.shape == r.shape and h.shape == t.shape:
x = self.hidden(torch.cat([h, r, t], dim=-1))
else:
# split weight into head-/relation-/tail-specific sub-matrices
*prefix, dim = h.shape
x = tensor_sum(
self.hidden.bias.view(*make_ones_like(prefix), -1),
*(
einsum("...i, ji -> ...j", xx, weight)
for xx, weight in zip([h, r, t], self.hidden.weight.split(split_size=dim, dim=-1))
),
)
return self.hidden_to_score(self.activation(x)).squeeze(dim=-1)
# docstr-coverage: inherited
[docs]
def reset_parameters(self): # noqa: D102
# Initialize biases with zero
nn.init.zeros_(self.hidden.bias)
nn.init.zeros_(self.hidden_to_score.bias)
# In the original formulation,
nn.init.xavier_uniform_(self.hidden.weight)
nn.init.xavier_uniform_(
self.hidden_to_score.weight,
gain=nn.init.calculate_gain(self.activation.__class__.__name__.lower()),
)
[docs]
@parse_docdata
class ERMLPEInteraction(Interaction[FloatTensor, FloatTensor, FloatTensor]):
r"""The stateful ER-MLP (E) interaction function.
This interaction uses a neural network-based approach similar to ER-MLP and with slight modifications.
In :class:`~pykeen.nn.modules.ERMLPInteraction`, the interaction is:
.. math::
f(h, r, t) = \textbf{w}^{T} g(\textbf{W} [\textbf{h}; \textbf{r}; \textbf{t}])
whereas here it is:
.. math::
f(h, r, t) = \textbf{t}^{T} f(\textbf{W} (g(\textbf{W} [\textbf{h}; \textbf{r}]))
including dropouts and batch-norms between each two hidden layers. Thus,
:class:`~pykeen.nn.modules.ConvEInteraction` can be seen as a special case of ERMLP (E).
---
name: ER-MLP (E)
citation:
author: Sharifzadeh
year: 2019
link: https://github.com/pykeen/pykeen
github: pykeen/pykeen
"""
def __init__(
self,
embedding_dim: int = 256,
input_dropout: float = 0.2,
hidden_dim: int | None = None,
hidden_dropout: float | None = None,
):
"""
Initialize the interaction module.
:param embedding_dim:
the embedding dimension of entities and relations
:param hidden_dim:
the hidden dimension of the MLP. Defaults to `embedding_dim`.
:param input_dropout:
the dropout applied *before* the first layer
:param hidden_dropout:
the dropout applied *after* the first layer
"""
super().__init__()
hidden_dim = hidden_dim or embedding_dim
hidden_dropout = input_dropout if hidden_dropout is None else hidden_dropout
self.mlp = nn.Sequential(
nn.Dropout(input_dropout),
nn.Linear(2 * embedding_dim, hidden_dim),
nn.Dropout(hidden_dropout),
nn.BatchNorm1d(hidden_dim),
nn.ReLU(),
nn.Linear(hidden_dim, embedding_dim),
nn.Dropout(hidden_dropout),
nn.BatchNorm1d(embedding_dim),
nn.ReLU(),
)
[docs]
def forward(self, h: FloatTensor, r: FloatTensor, t: FloatTensor) -> FloatTensor:
"""Compute broadcasted triple scores given broadcasted representations for head, relation and tails.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
# repeat if necessary, and concat head and relation, (*batch_dims, 2 * embedding_dim)
x = torch.cat(torch.broadcast_tensors(h, r), dim=-1)
# Predict t embedding, shape: (*batch_dims, d)
*batch_dims, dim = x.shape
x = self.mlp(x.view(-1, dim)).view(*batch_dims, -1)
# dot product
return batched_dot(x, t)
[docs]
@parse_docdata
class TransRInteraction(NormBasedInteraction[FloatTensor, tuple[FloatTensor, FloatTensor], FloatTensor]):
r"""The state-less norm-based TransR interaction function.
It is given by
.. math ::
-\|c(\mathbf{M}_{r}\mathbf{h}) + \mathbf{r} - c(\mathbf{M}_{r}\mathbf{t})\|_{2}^2
for head and tail entity representations $\mathbf{h}, \mathbf{t} \in \mathbb{R}^d$,
relation representation $\mathbf{r} \in \mathbb{R}^k$,
and a relation-specific projection matrix $\mathbf{M}_r \in \mathbb{R}^{k \times d}$.
$c$ enforces the constraint $\|\cdot\| \leq 1$, cf. :func:`pykeen.utils.clamp_norm`.
.. note ::
:class:`pykeen.models.TransR` additionally also enforces $\|\cdot\| \leq 1$ on all embeddings.
---
citation:
author: Lin
year: 2015
link: https://aaai.org/papers/9491-learning-entity-and-relation-embeddings-for-knowledge-graph-completion/
"""
relation_shape = ("e", "de")
def __init__(self, p: int, power_norm: bool = True, max_projection_norm: float = 1.0):
"""
Initialize the interaction module.
.. seealso::
The parameter ``p`` and ``power_norm`` are directly passed to
:class:`~pykeen.nn.modules.NormBasedInteraction`.
:param p:
The norm used with :func:`torch.linalg.vector_norm`. Typically is 1 or 2.
:param power_norm:
Whether to use the p-th power of the $L_p$ norm. It has the advantage of being differentiable around 0,
and numerically more stable.
:param max_projection_norm:
The maximum norm to be clamped after projection.
"""
super().__init__(p=p, power_norm=power_norm)
self.max_projection_norm = max_projection_norm
[docs]
def forward(self, h: FloatTensor, r: tuple[FloatTensor, FloatTensor], t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, k)`` and ``(*batch_dims, d, k)``
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
r, m_r = r
# project to relation specific subspace
h_bot = einsum("...e, ...er -> ...r", h, m_r)
t_bot = einsum("...e, ...er -> ...r", t, m_r)
# ensure constraints
h_bot = clamp_norm(h_bot, p=self.p, dim=-1, maxnorm=self.max_projection_norm)
t_bot = clamp_norm(t_bot, p=self.p, dim=-1, maxnorm=self.max_projection_norm)
return negative_norm_of_sum(h_bot, r, -t_bot, p=self.p, power_norm=self.power_norm)
[docs]
@parse_docdata
class RotatEInteraction(FunctionalInteraction[FloatTensor, FloatTensor, FloatTensor]):
r"""The RotatE interaction function proposed by [sun2019]_.
RotatE operates on complex-valued entity and relation representations, i.e.,
$\textbf{e}_i, \textbf{r}_i \in \mathbb{C}^d$.
.. note::
this method generally expects all tensors to be of complex datatype, i.e., `torch.is_complex(x)` to evaluate to
`True`. However, for backwards compatibility and convenience in use, you can also pass real tensors whose shape
is compliant with :func:`torch.view_as_complex`, cf. :func:`pykeen.utils.ensure_complex`.
---
citation:
arxiv: 1902.10197
author: Sun
github: DeepGraphLearning/KnowledgeGraphEmbedding
link: https://arxiv.org/abs/1902.10197
year: 2019
"""
# TODO: update docstring
is_complex: ClassVar[bool] = True
# TODO: give this a better name?
[docs]
@staticmethod
def func(h: FloatTensor, r: FloatTensor, t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. note::
this method expects all tensors to be of complex datatype, i.e., `torch.is_complex(x)` to evaluate to
`True`.
:param h: shape: (`*batch_dims`, dim)
The head representations.
:param r: shape: (`*batch_dims`, dim)
The relation representations.
:param t: shape: (`*batch_dims`, dim)
The tail representations.
:return: shape: batch_dims
The scores.
"""
if estimate_cost_of_sequence(h.shape, r.shape) < estimate_cost_of_sequence(r.shape, t.shape):
# r expresses a rotation in complex plane.
# rotate head by relation (=Hadamard product in complex space)
h = h * r
else:
# rotate tail by inverse of relation
# The inverse rotation is expressed by the complex conjugate of r.
# The score is computed as the distance of the relation-rotated head to the tail.
# Equivalently, we can rotate the tail by the inverse relation, and measure the distance to the head, i.e.
# |h * r - t| = |h - conj(r) * t|
t = t * torch.conj(r)
return negative_norm(h - t, p=2, power_norm=False)
[docs]
@parse_docdata
class HolEInteraction(FunctionalInteraction[FloatTensor, FloatTensor, FloatTensor]):
r"""The stateless HolE interaction function.
Holographic embeddings (HolE) make use of the circular correlation operator to compute interactions between
latent features of entities and relations:
.. math::
f(h,r,t) = \textbf{r}^{T}(\textbf{h} \star \textbf{t})
where the circular correlation $\star: \mathbb{R}^d \times \mathbb{R}^d \rightarrow \mathbb{R}^d$ is defined as:
.. math::
[\textbf{a} \star \textbf{b}]_i = \sum_{k=0}^{d-1} \textbf{a}_{k} * \textbf{b}_{(i+k)\ mod \ d}
By using the correlation operator each component $[\textbf{h} \star \textbf{t}]_i$ represents a sum over a
fixed partition over pairwise interactions. This enables the model to put semantic similar interactions into the
same partition and share weights through $\textbf{r}$. Similarly irrelevant interactions of features could also
be placed into the same partition which could be assigned a small weight in $\textbf{r}$.
---
citation:
author: Nickel
year: 2016
link: https://www.aaai.org/ocs/index.php/AAAI/AAAI16/paper/viewFile/12484/11828
github: mnick/holographic-embeddings
arxiv: 1510.04935
"""
[docs]
@staticmethod
def func(
h: torch.FloatTensor,
r: torch.FloatTensor,
t: torch.FloatTensor,
) -> torch.FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
# composite: (*batch_dims, d)
composite = circular_correlation(h, t)
# inner product with relation embedding
return batched_dot(r, composite)
[docs]
@parse_docdata
class ProjEInteraction(Interaction[FloatTensor, FloatTensor, FloatTensor]):
r"""The state-ful ProjE interaction function.
The activation function is given as
.. math ::
g(
f(
\mathbf{d}_h \odot \mathbf{h}
+ \mathbf{d}_r \odot \mathbf{r}
+ \mathbf{b}
)^T \mathbf{t} + b_p
)
where $\mathbf{h}, \mathbf{r}, \mathbf{t} \in \mathbb{R}^d$ are the head entity, relation, and tail entity
representations, $\mathbf{d}_h, \mathbf{d}_r, \mathbf{b} \in \mathbb{R}^d$ and $b_p \in \mathbb{R}$ are global
parameters, and $f, g$ activation functions.
It can be interpreted as a two-layer neural network with a very special sparsity pattern on the weight matrices.
The paper also describes the first operation
.. math ::
\mathbf{y} = f(\mathbf{d}_h \odot \mathbf{h}
+ \mathbf{d}_r \odot \mathbf{r}
+ \mathbf{b}
)
as the *combination* operation and the second part
.. math ::
g(\mathbf{y}^T \mathbf{t} + b_p)
as the *projection* layer.
.. note ::
While the original paper describes using either sigmoid or softmax as the final activation,
this implementation defaults to no activation on the final layer. This allows the use of numerically stable
implementations of merged activation and loss, such as :class:`torch.nn.BCEWithLogitsLoss` (for sigmoid),
or :class:`torch.nn.CrossEntropyLoss` (for softmax).
---
citation:
author: Shi
year: 2017
link: https://aaai.org/papers/10677-aaai-31-2017/
github: nddsg/ProjE
arxiv: 1611.05425
"""
@update_docstring_with_resolver_keys(
ResolverKey(name="inner_activation", resolver="class_resolver.contrib.torch.activation_resolver"),
ResolverKey(name="outer_activation", resolver="class_resolver.contrib.torch.activation_resolver"),
)
def __init__(
self,
embedding_dim: int = 50,
inner_activation: HintOrType[nn.Module] = None,
inner_activation_kwargs: OptionalKwargs = None,
outer_activation: HintOrType[nn.Module] = None,
outer_activation_kwargs: OptionalKwargs = None,
bias_initializer: Hint[Initializer] = init.xavier_uniform_,
bias_initializer_kwargs: OptionalKwargs = None,
projection_initializer: Hint[Initializer] = init.xavier_uniform_,
projection_initializer_kwargs: OptionalKwargs = None,
):
"""
Initialize the interaction module.
:param embedding_dim:
the embedding dimension of entities and relations
:param inner_activation:
the inner non-linearity, or a hint thereof. Defaults to :class:`nn.Tanh`.
Disable by passing :class:`nn.Idenity`
:param inner_activation_kwargs:
additional keyword-based parameters used to instantiate the inner activation function.
:param outer_activation:
the outer non-linearity, or a hint thereof. Defaults to :class:`nn.Identity`, i.e., no activation.
:param outer_activation_kwargs:
additional keyword-based parameters used to instantiate the outer activation function.
:param bias_initializer:
the initializer to use for the biases; defaults to :func:`pykeen.nn.init.xavier_uniform_`.
:param bias_initializer_kwargs:
additional keyword-based parameters passed to the bias initializer.
:param projection_initializer:
the initializer to use for the projection; defaults to :func:`pykeen.nn.init.xavier_uniform_`.
:param projection_initializer_kwargs:
additional keyword-based parameters passed to the projection initializer.
"""
super().__init__()
# Global entity projection
self.d_e = nn.Parameter(torch.empty(embedding_dim), requires_grad=True)
# Global relation projection
self.d_r = nn.Parameter(torch.empty(embedding_dim), requires_grad=True)
self.bias_initializer = init.initializer_resolver.make(bias_initializer, bias_initializer_kwargs)
# Global combination bias
self.b_c = nn.Parameter(torch.empty(embedding_dim), requires_grad=True)
# Global combination bias
self.b_p = nn.Parameter(torch.empty(tuple()), requires_grad=True)
self.projection_initializer = init.initializer_resolver.make(
projection_initializer, projection_initializer_kwargs
)
if inner_activation is None:
inner_activation = nn.Tanh
if outer_activation is None:
outer_activation = nn.Identity
self.inner_activation = activation_resolver.make(inner_activation, inner_activation_kwargs)
self.outer_activation = activation_resolver.make(outer_activation, outer_activation_kwargs)
[docs]
def forward(self, h: FloatTensor, r: FloatTensor, t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
# global projections
h = einsum("...d, d -> ...d", h, self.d_e)
r = einsum("...d, d -> ...d", r, self.d_r)
# combination, shape: (*batch_dims, d)
x = self.inner_activation(tensor_sum(h, r, self.b_c))
# dot product with t
return self.outer_activation(batched_dot(x, t) + self.b_p)
# docstr-coverage: inherited
[docs]
def reset_parameters(self): # noqa: D102
self.projection_initializer(self.d_e)
self.projection_initializer(self.d_r)
self.bias_initializer(self.b_c)
self.bias_initializer(self.b_p)
[docs]
@parse_docdata
class RESCALInteraction(Interaction[FloatTensor, FloatTensor, FloatTensor]):
r"""The state-less RESCAL interaction function.
For head and tail entity representations $\mathbf{h}, \mathbf{t} \in \mathbb{R}^d$
and relation representation $\mathbf{R} \in \mathbb{R}^{d \times d}$, the interaction function is given as
.. math::
\mathbf{h}^T \textbf{R} \textbf{t}
= \sum_{i=1}^{d} \sum_{j=1}^{d} \mathbf{h}_i \mathbf{R}_{i, j} \mathbf{t}_{i}
Thus, the relation matrices $\textbf{R}$ contain weights $\textbf{R}_{i, j}$ that capture the amount of interaction
between the $i$-th latent factor of the head representation and the $j$-th latent factor.
The computational complexity is given by $\mathcal{O}(d^2)$.
---
citation:
author: Nickel
year: 2011
link: https://icml.cc/2011/papers/438_icmlpaper.pdf
"""
relation_shape = ("dd",)
[docs]
def forward(self, h: FloatTensor, r: FloatTensor, t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
return einsum("...d,...de,...e->...", h, r, t)
[docs]
@parse_docdata
class SEInteraction(NormBasedInteraction[FloatTensor, tuple[FloatTensor, FloatTensor], FloatTensor]):
r"""The Structured Embedding (SE) interaction function.
SE applies role- and relation-specific projection matrices
$\textbf{M}_{r}^{h}, \textbf{M}_{r}^{t} \in \mathbb{R}^{d \times d}$ to the head and tail
entities' representations $\mathbf{h}, \mathbf{t} \in \mathbb{R}^d$ before computing their distance.
.. math::
f(\textbf{h}, (\textbf{M}_{r}^{h}, \textbf{M}_{r}^{t}), \textbf{t})
= -\|\textbf{M}_{r}^{h} \textbf{h} - \textbf{M}_{r}^{t} \textbf{t}\|_p
---
name: Structured Embedding
citation:
author: Bordes
year: 2011
link: https://www.aaai.org/ocs/index.php/AAAI/AAAI11/paper/download/3659/3898
"""
relation_shape = ("dd", "dd")
[docs]
def forward(self, h: FloatTensor, r: FloatTensor, t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d, d)`` and ``(*batch_dims, d, d)``.
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
r_h, r_t = r
# projections
p_h = einsum("...rd,...d->...r", r_h, h)
p_t = einsum("...rd,...d->...r", r_t, t)
return negative_norm(p_h - p_t, p=self.p, power_norm=self.power_norm)
[docs]
@parse_docdata
class TuckERInteraction(Interaction[FloatTensor, FloatTensor, FloatTensor]):
r"""The stateful TuckER interaction function.
The interaction function is inspired by the `Tucker tensor decomposition <https://en.wikipedia.org/wiki/Tucker_decomposition>`_.
The base form is given as
.. math ::
\mathbf{Z} \times_1 \mathbf{h} \times_2 \mathbf{r} \times_3 \mathbf{t}
= \sum_{1 \leq i, k \leq d_e, 1 \leq j \leq d_r}
\mathbf{Z}_{i, j, k} \cdot \mathbf{h}_{i} \cdot \mathbf{r}_{j} \cdot \mathbf{t}_{k}
where $\mathbf{h}, \mathbf{t} \in \mathbb{R}^{d_e}$ are the head and tail entity representation,
$\mathbf{r} \in \mathbb{R}^{d_r}$ is the relation representation, and
$\mathbf{Z} \in \mathbb{R}^{d_e \times d_r \times d_e}$ is a *global* parameter, and $\times_k$ denotes the tensor
product along the $k$-th dimension.
The implementation further adds :class:`~torch.nn.BatchNorm1d` and :class:`~torch.nn.Dropout`
layers at the following locations:
.. math ::
\textit{DO}_{hr}(\textit{BN}_{hr}(
\textit{DO}_h(\textit{BN}_h(\mathbf{h}))
\times_1
\textit{DO}_r(\mathbf{Z} \times_2 \mathbf{r})
) \times_3 \mathbf{t}
The implementation a has complexity of $\mathcal{O}(d_e^2 d_r)$, and requires $\mathcal{O}(d_e^2 d_r)$
global trainable parameters.
---
citation:
author: Balažević
year: 2019
arxiv: 1901.09590
link: https://arxiv.org/abs/1901.09590
github: ibalazevic/TuckER
"""
# default core tensor initialization
# cf. https://github.com/ibalazevic/TuckER/blob/master/model.py#L12
default_core_initializer: ClassVar[Initializer] = staticmethod(nn.init.uniform_) # type: ignore
default_core_initializer_kwargs: Mapping[str, Any] = {"a": -1.0, "b": 1.0}
@update_docstring_with_resolver_keys(
ResolverKey(name="core_initializer", resolver="pykeen.nn.init.initializer_resolver")
)
def __init__(
self,
embedding_dim: int = 200,
relation_dim: int | None = None,
head_dropout: float = 0.3,
relation_dropout: float = 0.4,
head_relation_dropout: float = 0.5,
apply_batch_normalization: bool = True,
core_initializer: Hint[Initializer] = None,
core_initializer_kwargs: OptionalKwargs = None,
):
"""Initialize the Tucker interaction function.
:param embedding_dim:
The entity embedding dimension.
:param relation_dim:
The relation embedding dimension. Defaults to ``embedding_dim``.
:param head_dropout:
The dropout rate applied to the head representations.
:param relation_dropout:
The dropout rate applied to the relation representations.
:param head_relation_dropout:
The dropout rate applied to the combined head and relation representations.
:param apply_batch_normalization:
Whether to use batch normalization on head representations and the combination of head and relation.
:param core_initializer:
The core tensor's initializer, or a hint thereof.
Defaults to :attr:`~pykeen.nn.modules.TuckerInteraction.default_core_initializer`.
:param core_initializer_kwargs:
Additional keyword-based parameters for the initializer.
Defaults to :attr:`~pykeen.nn.modules.TuckerInteraction.default_core_initializer_kwargs`.
"""
super().__init__()
# normalize initializer
if core_initializer is None:
core_initializer = self.default_core_initializer
self.core_initializer = core_initializer
if core_initializer is self.default_core_initializer and core_initializer_kwargs is None:
core_initializer_kwargs = self.default_core_initializer_kwargs
self.core_initializer_kwargs = core_initializer_kwargs
# normalize relation dimension
if relation_dim is None:
relation_dim = embedding_dim
# Core tensor
# Note: we use a different dimension permutation as in the official implementation to match the paper.
self.core_tensor = nn.Parameter(
torch.empty(embedding_dim, relation_dim, embedding_dim),
requires_grad=True,
)
# Dropout
self.head_dropout = nn.Dropout(head_dropout)
self.relation_dropout = nn.Dropout(relation_dropout)
self.head_relation_dropout = nn.Dropout(head_relation_dropout)
if apply_batch_normalization:
self.head_batch_norm = nn.BatchNorm1d(embedding_dim)
self.head_relation_batch_norm = nn.BatchNorm1d(embedding_dim)
else:
self.head_batch_norm = self.head_relation_batch_norm = None
self.reset_parameters()
# docstr-coverage: inherited
[docs]
def reset_parameters(self): # noqa: D102
# instantiate here to make module easily serializable
core_initializer = init.initializer_resolver.make(
self.core_initializer, pos_kwargs=self.core_initializer_kwargs
)
core_initializer(self.core_tensor)
# batch norm gets reset automatically, since it defines reset_parameters
[docs]
def forward(self, h: FloatTensor, r: FloatTensor, t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
h = self.head_dropout(apply_optional_bn(x=h, batch_norm=self.head_batch_norm))
# x_2 contraction
r = einsum("ijk,...j->...ik", self.core_tensor, r)
r = self.relation_dropout(r)
# x_1 contraction
return batched_dot(
self.head_relation_dropout(
apply_optional_bn(
x=einsum("...ik,...i->...k", r, h),
batch_norm=self.head_relation_batch_norm,
)
),
t,
)
[docs]
@parse_docdata
class UMInteraction(NormBasedInteraction[FloatTensor, tuple[()], FloatTensor]):
r"""The Unstructured Model (UM) interaction function.
UM calculates the score as the negative distance between head and tail entities:
.. math::
-\|\textbf{h} - \textbf{t}\|_p^2
It is appropriate for networks with a single relationship type that is undirected.
.. warning::
In UM, neither the relations nor the directionality are considered, so it can't distinguish between them.
However, it may serve as a baseline for comparison against relation-aware models.
---
name: Unstructured Model
citation:
author: Bordes
year: 2014
link: https://link.springer.com/content/pdf/10.1007%2Fs10994-013-5363-6.pdf
"""
# shapes
relation_shape: Sequence[str] = tuple()
def __init__(self, p: int, power_norm: bool = True):
"""Initialize the norm-based interaction function.
.. seealso::
The parameter ``p`` and ``power_norm`` are directly passed to
:class:`~pykeen.nn.modules.NormBasedInteraction`.
:param p:
The norm used with :func:`torch.linalg.vector_norm`. Typically is 1 or 2.
:param power_norm:
Whether to use the p-th power of the $L_p$ norm. It has the advantage of being differentiable around 0,
and numerically more stable.
"""
super().__init__(p=p, power_norm=power_norm)
[docs]
def forward(self, h: FloatTensor, r: tuple[()], t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r:
No relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
return negative_norm(h - t, p=self.p, power_norm=self.power_norm)
[docs]
@parse_docdata
class TorusEInteraction(NormBasedInteraction[FloatTensor, FloatTensor, FloatTensor]):
"""The TorusE interaction function from [ebisu2018].
.. note ::
This only implements the two L_p norm based variants.
---
citation:
author: Ebisu
year: 2018
link: https://www.aaai.org/ocs/index.php/AAAI/AAAI18/paper/view/16227
arxiv: 1711.05435
github: TakumaE/TorusE
"""
def __init__(self, p: int = 2, power_norm: bool = False):
"""
Initialize the interaction module.
.. seealso::
The parameter ``p`` and ``power_norm`` are directly passed to
:class:`~pykeen.nn.modules.NormBasedInteraction`.
:param p:
The norm used with :func:`torch.linalg.vector_norm`. Typically is 1 or 2.
:param power_norm:
Whether to use the p-th power of the $L_p$ norm. It has the advantage of being differentiable around 0,
and numerically more stable.
"""
super().__init__(p=p, power_norm=power_norm)
[docs]
def forward(self, h: FloatTensor, r: FloatTensor, t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
d = tensor_sum(h, r, -t)
d = d - torch.floor(d)
d = torch.minimum(d, 1.0 - d)
return negative_norm(d, p=self.p, power_norm=self.power_norm)
[docs]
@parse_docdata
class TransDInteraction(
NormBasedInteraction[
tuple[FloatTensor, FloatTensor], tuple[FloatTensor, FloatTensor], tuple[FloatTensor, FloatTensor]
]
):
r"""The TransD interaction function.
TransD is an extension of :class:`~pykeen.nn.modules.TransRInteraction` that, like TransR,
considers entities and relations as objects living in different vector spaces.
However, instead of performing the same relation-specific projection for all entity embeddings,
entity-relation-specific projection matrices
$\mathbf{M}_{r, h}, \mathbf{M}_{t, h} \in \mathbb{R}^{k \times d}$
are constructed.
To do so, all head entities, tail entities, and relations are represented by two vectors,
$\mathbf{h}_v, \mathbf{h}_p, \mathbf{t}_v, \mathbf{t}_p \in \mathbb{R}^d$
and $\mathbf{r}_v, \mathbf{r}_v \in \mathbb{R}^k$, respectively.
The first set of representations is used for calculating the entity-relation-specific projection matrices:
.. math::
\mathbf{M}_{r, h} &=& \mathbf{r}_p \mathbf{h}_p^{T} + \tilde{\mathbf{I}}
\mathbf{M}_{r, t} &=& \mathbf{r}_p \mathbf{t}_p^{T} + \tilde{\mathbf{I}}
where $\tilde{\textbf{I}} \in \mathbb{R}^{k \times d}$ is a $k \times d$ matrix with ones on the diagonal and
zeros elsewhere. Next, $\mathbf{h}_v$ and $\mathbf{t}_v$ are projected into the relation space by means of the
constructed projection matrices, before calculating a distance similar to
:class:`~pykeen.nn.modules.TransEInteraction`:
.. math::
-\|c(\mathbf{M}_{r, h} \mathbf{h}_v) + \mathbf{r}_v - c(\mathbf{M}_{r, t} \mathbf{t}_v)\|_{2}^2
where $c$ enforces the constraint $\|\cdot\| \leq 1$.
.. note ::
:class:`~pykeen.models.TransD` additionally enforces $\|\mathbf{h}\|, \|\mathbf{r}\|, \|\mathbf{t}\| \leq 1$.
---
citation:
author: Ji
year: 2015
link: http://www.aclweb.org/anthology/P15-1067
"""
entity_shape = ("d", "d")
relation_shape = ("e", "e")
def __init__(self, p: int = 2, power_norm: bool = True):
"""
Initialize the interaction module.
.. seealso::
The parameters ``p`` and ``power_norm`` are directly passed to
:class:`~pykeen.nn.modules.NormBasedInteraction`
:param p:
The norm used with :func:`torch.linalg.vector_norm`. Typically is 1 or 2.
:param power_norm:
Whether to use the p-th power of the $L_p$ norm. It has the advantage of being differentiable around 0,
and numerically more stable.
"""
super().__init__(p=p, power_norm=power_norm)
[docs]
def forward(
self, h: tuple[FloatTensor, FloatTensor], r: tuple[FloatTensor, FloatTensor], t: tuple[FloatTensor, FloatTensor]
) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)`` and ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, e)`` and ``(*batch_dims, e)``
The relation representations.
:param t: shape: ``(*batch_dims, d)`` and ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
h_emb, h_p = h
r_emb, r_p = r
t_emb, t_p = t
# Project entities
h_bot = project_entity(e=h_emb, e_p=h_p, r_p=r_p)
t_bot = project_entity(e=t_emb, e_p=t_p, r_p=r_p)
return negative_norm_of_sum(h_bot, r_emb, -t_bot, p=self.p, power_norm=self.power_norm)
[docs]
@parse_docdata
class NTNInteraction(
Interaction[FloatTensor, tuple[FloatTensor, FloatTensor, FloatTensor, FloatTensor, FloatTensor], FloatTensor],
):
r"""The state-less Neural Tensor Network (NTN) interaction function.
It is given by
.. math::
\mathbf{r}_{u}^{T} \cdot \sigma(
\mathbf{h} \mathbf{R}_{3} \mathbf{t}
+ \mathbf{R}_{2} [\mathbf{h};\mathbf{t}]
+ \mathbf{r}_1
)
with $\mathbf{W}_3 \in \mathbb{R}^{d \times d \times k}$, $\textbf{R}_2 \in \mathbb{R}^{k \times 2d}$,
the bias vector $\textbf{r}_1$, the final projection $\textbf{r}_u \in \mathbb{R}^k$, and a non-linear activation
function $\sigma$ (which defaults to :class:`~torch.nn.Tanh`).
It can be seen as an extension of a two-layer MLP with relation-specific weights
and an additional bi-linear tensor in the input layer.
A separately parameterized neural network for each relationship makes the model very expressive,
but also computationally expensive ($\mathcal{O}(kd^2)$).
.. note::
We split the original $k \times 2d$-dimensional $\mathbf{R}_2$ matrix into two parts of shape $k \times d$ to
support more efficient 1:n scoring, e.g., in the :meth:`~pykeen.models.Model.score_h` or
:meth:`~pykeen.models.Model.score_t` setting.
---
citation:
author: Socher
year: 2013
link: https://proceedings.neurips.cc/paper/2013/file/b337e84de8752b27eda3a12363109e80-Paper.pdf
github: khurram18/NeuralTensorNetworks
"""
relation_shape = ("kdd", "kd", "kd", "k", "k")
@update_docstring_with_resolver_keys(
ResolverKey(name="activation", resolver="class_resolver.contrib.torch.activation_resolver")
)
def __init__(
self,
activation: HintOrType[nn.Module] = None,
activation_kwargs: Mapping[str, Any] | None = None,
):
"""Initialize NTN with the given non-linear activation function.
:param activation: A non-linear activation function. Defaults to the hyperbolic
tangent :class:`torch.nn.Tanh` if ``None``.
:param activation_kwargs: If the ``activation`` is passed as a class, these keyword arguments
are used during its instantiation.
"""
super().__init__()
if activation is None:
activation = nn.Tanh()
self.activation = activation_resolver.make(activation, activation_kwargs)
[docs]
def forward(
self, h: FloatTensor, r: tuple[FloatTensor, FloatTensor, FloatTensor, FloatTensor, FloatTensor], t: FloatTensor
) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, k, d, d)``, ``(*batch_dims, k, d)``, ``(*batch_dims, k, d)``,
``(*batch_dims, k)``, and ``(*batch_dims, k)``
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
w, vh, vt, b, u = r
return batched_dot(
u,
self.activation(
tensor_sum(
einsum("...d,...kde,...e->...k", h, w, t),
einsum("...d, ...kd->...k", h, vh),
einsum("...d, ...kd->...k", t, vt),
b,
)
),
)
[docs]
@parse_docdata
class KG2EInteraction(
Interaction[tuple[FloatTensor, FloatTensor], tuple[FloatTensor, FloatTensor], tuple[FloatTensor, FloatTensor]]
):
r"""The stateless KG2E interaction function.
Inspired by :class:`~pykeen.nn.modules.TransEInteraction`, relations are modeled as transformations
from head to tail entities $\mathcal{H} - \mathcal{T} \approx \mathcal{R}$, where
.. math ::
\mathcal{H} \sim \mathcal{N}(\mu_h, \Sigma_h)\\
\mathcal{T} \sim \mathcal{N}(\mu_t, \Sigma_t)\\
\mathcal{R} \sim \mathcal{N}(\mu_r, \Sigma_r)
and thus, since head and tail entities are considered independent with respect to the relations,
.. math ::
\mathcal{P}_e = \mathcal{H} - \mathcal{T} \sim \mathcal{N}(\mu_h - \mu_t, \Sigma_h + \Sigma_t)
To obtain scores, the interaction measures the similarity between $\mathcal{P}_e$ and
$\mathcal{P}_r = \mathcal{N}(\mu_r, \Sigma_r)$, either by means of the (asymmetric)
:class:`~pykeen.nn.sim.NegativeKullbackLeiblerDivergence`, or a symmetric variant with
:class:`~pykeen.nn.sim.ExpectedLikelihood`.
.. note ::
This interaction module does *not* sub-class from :class:`~pykeen.nn.modules.FunctionalInteraction`
just for the technical reason that the choice of the similarity represents some "state". However, it
does not contain any trainable parameters.
---
citation:
author: He
year: 2015
link: https://dl.acm.org/doi/10.1145/2806416.2806502
"""
entity_shape = ("d", "d")
relation_shape = ("d", "d")
similarity: KG2ESimilarity
@update_docstring_with_resolver_keys(
ResolverKey(name="similarity", resolver="pykeen.nn.sim.kg2e_similarity_resolver")
)
def __init__(self, similarity: HintOrType[KG2ESimilarity] | None = None, similarity_kwargs: OptionalKwargs = None):
"""
Initialize the interaction module.
:param similarity:
The similarity measures for gaussian distributions. Defaults to
:class:`~pykeen.nn.sim.NegativeKullbackLeiblerDivergence`.
:param similarity_kwargs:
Additional keyword-based parameters used to instantiate the similarity.
"""
super().__init__()
self.similarity = kg2e_similarity_resolver.make(similarity, similarity_kwargs)
[docs]
def forward(
self, h: tuple[FloatTensor, FloatTensor], r: tuple[FloatTensor, FloatTensor], t: tuple[FloatTensor, FloatTensor]
) -> FloatTensor:
"""Evaluate the interaction function.
:param h: both shape: (`*batch_dims`, `d`)
The head representations, mean and (diagonal) variance.
:param r: shape: (`*batch_dims`, `d`)
The relation representations, mean and (diagonal) variance.
:param t: shape: (`*batch_dims`, `d`)
The tail representations, mean and (diagonal) variance.
:return: shape: batch_dims
The scores.
"""
h_mean, h_var = h
r_mean, r_var = r
t_mean, t_var = t
return self.similarity(
h=GaussianDistribution(mean=h_mean, diagonal_covariance=h_var),
r=GaussianDistribution(mean=r_mean, diagonal_covariance=r_var),
t=GaussianDistribution(mean=t_mean, diagonal_covariance=t_var),
)
[docs]
@parse_docdata
class TransHInteraction(NormBasedInteraction[FloatTensor, tuple[FloatTensor, FloatTensor], FloatTensor]):
r"""The norm-based TransH interaction function.
This model extends :class:`~pykeen.models.TransEInteraction` by applying the translation from head to tail entity
in a relation-specific hyperplane in order to address its inability to model one-to-many, many-to-one, and
many-to-many relations.
In TransH, each relation is represented by a hyperplane, or more specifically a normal vector of this hyperplane
$\mathbf{r}_{w} \in \mathbb{R}^d$ and a vector $\mathbf{r}_{d} \in \mathbb{R}^d$ that lies in the hyperplane.
To obtain a plausibility score, the head representation $\mathbf{h} \in \mathbb{R}^d$,
and the tail embedding $\mathbf{t} \in \mathbb{R}^d$ are first projected onto the relation-specific hyperplane:
.. math::
\mathbf{h}_{r} = \mathbf{h} - \mathbf{r}_{w}^T \mathbf{h} \mathbf{r}_w
\mathbf{t}_{r} = \mathbf{t} - \mathbf{r}_{w}^T \mathbf{t} \mathbf{r}_w
Then, the projected representations are used to compute the score as in
:class:`~pykeen.nn.modules.TransEInteraction`:
.. math::
-\|\textbf{h}_{r} + \textbf{r}_d - \textbf{t}_{r}\|_{p}^2
---
citation:
author: Wang
year: 2014
link: https://aaai.org/papers/8870-knowledge-graph-embedding-by-translating-on-hyperplanes/
"""
relation_shape = ("d", "d")
[docs]
def forward(self, h: FloatTensor, r: tuple[FloatTensor, FloatTensor], t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
w_r, d_r = r
return negative_norm_of_sum(
# h projection to hyperplane
h,
-einsum("...i, ...i, ...j -> ...j", h, w_r, w_r),
# r
d_r,
# -t projection to hyperplane
-t,
einsum("...i, ...i, ...j -> ...j", t, w_r, w_r),
p=self.p,
power_norm=self.power_norm,
)
[docs]
@parse_docdata
class MuREInteraction(
NormBasedInteraction[
tuple[FloatTensor, FloatTensor], tuple[FloatTensor, FloatTensor], tuple[FloatTensor, FloatTensor]
],
):
r"""The norm-based MuRE interaction function from [balazevic2019b]_.
For $\mathbf{h}, \mathbf{r}, \mathbf{R}, \mathbf{t} \in \mathbb{R}^d$, and $b_h, b_t \in \mathbb{R}$, it is given
by
.. math ::
-\|\mathbf{R} \odot \mathbf{h} + \mathbf{r} - \mathbf{t}\| + b_h + b_t
where $\mathbf{h}, \mathbf{r}, \mathbf{t}$ are head entity, relation, and tail entity embedding vectors,
$\mathbf{R}$ is a diagonal relation matrix, and $b_h, b_t$ are head and tail entity biases.
.. note::
This module implements a slightly more generic function, where the norm $\| \cdot \|_p$ can be chosen,
as well as a variant which uses $\| \cdot \|_p^p$, cf. :class:`~pykeen.nn.modules.NormBasedInteraction`.
---
citation:
author: Balažević
year: 2019
link: https://arxiv.org/abs/1905.09791
arxiv: 1905.09791
"""
# there are separate biases for entities in head and tail position
entity_shape = ("d", "", "")
_head_indices = (0, 1)
_tail_indices = (0, 2)
relation_shape = ("d", "d")
[docs]
def forward(
self, h: tuple[FloatTensor, FloatTensor], r: tuple[FloatTensor, FloatTensor], t: tuple[FloatTensor, FloatTensor]
) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)`` and ``(*batch_dims)``
The head representations.
:param r: shape: ``(*batch_dims, d)`` and ``(*batch_dims, d)``
The relation representations.
:param t: shape: ``(*batch_dims, d)`` and ``(*batch_dims)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
h_emb, h_bias = h
t_emb, t_bias = t
r_vec, r_mat = r
return (
negative_norm_of_sum(h_emb * r_mat, r_vec, -t_emb, p=self.p, power_norm=self.power_norm) + h_bias + t_bias
)
Clamp = Union[tuple[Optional[float], float], tuple[float, Optional[float]]]
[docs]
class ClampedInteraction(Interaction[HeadRepresentation, RelationRepresentation, TailRepresentation]):
"""An adapter to clamp scores to a minimum or maximum value.
.. warning::
The used :func:`torch.clamp` function has zero gradient for scores below the minimum of above the maximum value.
Thus, it aggravates gradient-based optimization.
"""
clamp_score: Clamp | None
base: Interaction[HeadRepresentation, RelationRepresentation, TailRepresentation]
@update_docstring_with_resolver_keys(ResolverKey(name="base", resolver="interaction_resolver"))
def __init__(
self,
base: LookupOrType[Interaction[HeadRepresentation, RelationRepresentation, TailRepresentation]],
base_kwargs: OptionalKwargs = None,
clamp_score: Clamp | float | None = None,
):
"""
Initialize the interaction module.
:param base:
the base interaction.
:param base_kwargs:
keyword-based parameters used to instantiate the base interaction
:param clamp_score:
whether to clamp scores into a fixed interval
"""
super().__init__()
if isinstance(clamp_score, float):
clamp_score = (-clamp_score, clamp_score)
self.clamp_score = clamp_score
self.base = interaction_resolver.make(base, base_kwargs)
@property
def entity_shape(self) -> Sequence[str]: # type:ignore[override]
"""Expose the base interaction's entity shape."""
return self.base.entity_shape
@property
def relation_shape(self) -> Sequence[str]: # type:ignore[override]
"""Expose the base interaction's relation shape."""
return self.base.relation_shape
# docstr-coverage: inherited
[docs]
def forward(self, h: HeadRepresentation, r: RelationRepresentation, t: TailRepresentation) -> FloatTensor:
scores = self.base(h, r, t)
if self.clamp_score is None:
return scores
low, high = self.clamp_score
return torch.clamp(scores, min=low, max=high)
[docs]
class DirectionAverageInteraction(
Interaction[
tuple[FloatTensor, FloatTensor],
tuple[FloatTensor, FloatTensor],
tuple[FloatTensor, FloatTensor],
],
):
r"""The directional average interaction module.
This can be considered as a generalization of the SimplE interaction module that can be parametrized
with any other interaction module, rather than just :class:`pykeen.nn.modules.DistMultInteraction`.
A separate representation is learned for each entity $e \in \mathcal{E}$ for when it appears as the
subject of a triple $\mathbf{e}_h \in \mathbb{R}^d$ and as the object of a triple $\mathbf{e}_t \in \mathbb{R}^d$.
Similarly, two representations are learned for each relationship for a forward $\textbf{r}_{\rightarrow}$
and backward triple $\textbf{r}_{\leftarrow}$.
The score is then obtained by averaging the *forward* and the *backward* interaction function value:
.. math::
\frac{
f(\textbf{h}_{h}, \textbf{r}_{\rightarrow}, \textbf{t}_{t})
+ f(\textbf{t}_{h}, \textbf{r}_{\leftarrow}, \textbf{h}_{t})
}{2}
Where ``f`` is the interaction model used. If :class:`pykeen.nn.modules.DistMultInteraction` is used,
then this becomes :class:`pykeen.nn.modules.SimplEInteraction`.
.. todo:: can we generalize the type annotations for this from FloatTensor to HeadRepresentation, etc.?
"""
@update_docstring_with_resolver_keys(ResolverKey(name="base", resolver="interaction_resolver"))
def __init__(
self,
base: LookupOrType[Interaction[FloatTensor, FloatTensor, FloatTensor]],
base_kwargs: OptionalKwargs = None,
):
"""
Initialize the interaction module.
:param base:
the base interaction.
:param base_kwargs:
keyword-based parameters used to instantiate the base interaction
"""
super().__init__()
self.base = interaction_resolver.make(base, base_kwargs)
[docs]
def forward(
self,
h: tuple[FloatTensor, FloatTensor],
r: tuple[FloatTensor, FloatTensor],
t: tuple[FloatTensor, FloatTensor],
) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)`` and ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)`` and ``(*batch_dims, d)``
The relation representations.
:param t: shape: ``(*batch_dims, d)`` and ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
h_fwd, h_bwd = h
r_fwd, r_bwd = r
t_fwd, t_bwd = t
return 0.5 * (self.base(h_fwd, r_fwd, t_fwd) + self.base(t_bwd, r_bwd, h_bwd))
[docs]
@parse_docdata
class SimplEInteraction(DirectionAverageInteraction):
r"""The SimplE interaction function.
SimplE can be regarded as extension of (a special case of) :class:`pykeen.nn.modules.CPInteraction`,
an early tensor factorization approach in which each entity
$e \in \mathcal{E}$ is represented by two vectors $\mathbf{e}_h, \mathbf{e}_t \in \mathbb{R}^d$ and each
relation by a single vector $\mathbf{r} \in \mathbb{R}^d$. Depending whether an entity participates in a
triple as the head or tail entity, either $\mathbf{e}_h$ or $\mathbf{e}_t$ is used. Both entity
representations are learned independently, i.e. observing a triple $(h,r,t)$, the method only updates
$\mathbf{h}_h$ and $\mathbf{t}_t$.
In contrast to :class:`~pykeen.nn.modules.CPInteraction`, SimplE introduces separate weights for each relation:
$\textbf{r}_{\rightarrow}$ and $\textbf{r}_{\leftarrow}$ for the inverse relation.
The interaction model is based on both:
.. math::
\frac{1}{2}\left(
\left\langle\textbf{h}_{h}, \textbf{r}_{\rightarrow}, \textbf{t}_{t}\right\rangle
+ \left\langle\textbf{t}_{h}, \textbf{r}_{\leftarrow}, \textbf{h}_{t}\right\rangle
\right)
---
citation:
author: Kazemi
year: 2018
link: https://papers.nips.cc/paper/7682-simple-embedding-for-link-prediction-in-knowledge-graphs
github: Mehran-k/SimplE
"""
entity_shape = ("d", "d")
relation_shape = ("d", "d")
def __init__(self):
"""Initialize the interaction module."""
super().__init__(DistMultInteraction)
[docs]
@parse_docdata
class PairREInteraction(NormBasedInteraction[FloatTensor, tuple[FloatTensor, FloatTensor], FloatTensor]):
r"""The state-less norm-based PairRE interaction function.
It is given by
.. math ::
-\|\mathbf{h} \odot \mathbf{r}_h - \mathbf{t} \odot \mathbf{r}_t \|
where $\mathbf{h}, \mathbf{r}_h, \mathbf{r}_t, \mathbf{t} \in \mathbb{R}$ are representations for head entity,
relation-specific head projection, relation-specific tail projection, and tail entity, respectively.
.. note ::
:class:`pykeen.models.PairRE` additionally enforces $\|\mathbf{h}\| = \|\mathbf{t}\| = 1$.
---
citation:
author: Chao
year: 2020
link: http://arxiv.org/abs/2011.03798
arxiv: 2011.03798
github: alipay/KnowledgeGraphEmbeddingsViaPairedRelationVectors_PairRE
"""
relation_shape = ("d", "d")
[docs]
def forward(self, h: FloatTensor, r: tuple[FloatTensor, FloatTensor], t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)`` and ``(*batch_dims, d)``
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
r_h, r_t = r
return negative_norm_of_sum(
h * r_h,
-t * r_t,
p=self.p,
power_norm=self.power_norm,
)
[docs]
@parse_docdata
class QuatEInteraction(Interaction[FloatTensor, FloatTensor, FloatTensor]):
r"""The state-less QuatE interaction function.
It is given as
.. math ::
\langle \mathbf{h} \otimes \mathbf{r}, \mathbf{t} \rangle
where $\mathbf{h}, \mathbf{r}, \mathbf{t} \in \mathbb{H}^d$ are quanternion representations,
$\otimes$ denotes the Hamilton product, and $\langle \cdot, \cdot \rangle$ the inner product.
.. warning ::
In order to representation a rotation, $\mathbf{r}$ must be normalized to unit length,
cf. :func:`pykeen.nn.quaternion.normalize`.
.. seealso::
- https://en.wikipedia.org/wiki/Quaternion
---
citation:
author: Zhang
year: 2019
arxiv: 1904.10281
link: https://arxiv.org/abs/1904.10281
github: cheungdaven/quate
"""
# with k=4
entity_shape: Sequence[str] = ("dk",)
relation_shape: Sequence[str] = ("dk",)
def __init__(self) -> None:
"""Initialize the interaction module."""
super().__init__()
self.register_buffer(name="table", tensor=quaternion.multiplication_table())
[docs]
def forward(self, h: FloatTensor, r: tuple[FloatTensor, FloatTensor], t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function of QuatE for given embeddings.
The embeddings have to be in a broadcastable shape.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: (`*batch_dims`, dim, 4)
The head representations.
:param r: shape: (`*batch_dims`, dim, 4)
The head representations.
:param t: shape: (`*batch_dims`, dim, 4)
The tail representations.
:return: shape: (...)
The scores.
"""
# TODO: this sign is in the official code, too, but why do we need it?
# note: this is a fused kernel for computing the Hamilton product and the inner product at once
return -einsum("...di, ...dj, ...dk, ijk -> ...", h, r, t, self.table)
[docs]
@parse_docdata
class CrossEInteraction(Interaction[FloatTensor, tuple[FloatTensor, FloatTensor], FloatTensor]):
r"""The stateful interaction function of CrossE.
The interaction function is given by
.. math ::
\textit{drop}(
\textit{act}(
\mathbf{c}_r \odot \mathbf{h} + \mathbf{c}_r \odot \mathbf{h} \odot \mathbf{r} + \mathbf{b})
)
)^T
\mathbf{t}
where $\mathbf{h}, \mathbf{c}_r, \mathbf{r}, \mathbf{t} \in \mathbb{R}^d$ is the head embedding, the relation
interaction vector, the relation embedding, and the tail embedding, respectively.
$\mathbf{b} \in \mathbb{R}^d$ is a global bias vector (which makes this interaction function stateful).
$\textit{drop}$ denotes dropout, and $\textit{act}$ is the activation function.
.. note ::
The CrossE paper describes an additional sigmoid activation as part of the interaction function. Since using a
log-likelihood loss can cause numerical problems (due to explicitly calling sigmoid before log), we do not use
it in our implementation, but opt for the numerically stable variant. However, the model itself has an option
``predict_with_sigmoid``, which can be used to force the use of sigmoid during inference. This can also affect
rank-based scoring, since limited numerical precision can lead to exactly equal scores for multiple choices.
The definition of a rank is not clear in this case, and there are several competing ways to break ties.
See :ref:`understanding-evaluation` for more information.
---
citation:
author: Zhang
year: 2019
link: https://arxiv.org/abs/1903.04750
arxiv: 1903.04750
github: https://github.com/wencolani/CrossE
"""
relation_shape = ("d", "d")
@update_docstring_with_resolver_keys(
ResolverKey("combination_activation", "class_resolver.contrib.torch.activation_resolver")
)
def __init__(
self,
embedding_dim: int = 50,
combination_activation: HintOrType[nn.Module] = nn.Tanh,
combination_activation_kwargs: Mapping[str, Any] | None = None,
combination_dropout: float | None = 0.5,
):
"""
Instantiate the interaction module.
:param embedding_dim:
The embedding dimension.
:param combination_activation:
The combination activation function.
:param combination_activation_kwargs:
Additional keyword-based arguments passed to the constructor of the combination activation function (if
not already instantiated).
:param combination_dropout:
An optional dropout applied after the combination and before the dot product similarity.
"""
super().__init__()
self.activation = activation_resolver.make(
combination_activation,
pos_kwargs=combination_activation_kwargs,
)
# TODO: expose initialization?
self.bias = nn.Parameter(data=torch.zeros(embedding_dim))
self.dropout = nn.Dropout(combination_dropout) if combination_dropout else None
[docs]
def forward(
self,
h: FloatTensor,
r: tuple[FloatTensor, FloatTensor],
t: FloatTensor,
) -> FloatTensor:
r"""
Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: (`*batch_dims`, dim)
The head representations.
:param r: shape: (`*batch_dims`, dim)
The relation representations and relation-specific interaction vector.
:param t: shape: (`*batch_dims`, dim)
The tail representations.
:return: shape: batch_dims
The scores.
"""
r_emb, c_r = r
# head interaction
h = c_r * h
# relation interaction (notice that h has been updated)
r_emb = h * r_emb
# combination
x = self.activation(self.bias.view(*make_ones_like(h.shape[:-1]), -1) + h + r_emb)
if self.dropout is not None:
x = self.dropout(x)
# similarity
return batched_dot(x, t)
[docs]
@parse_docdata
class BoxEInteraction(
NormBasedInteraction[
tuple[FloatTensor, FloatTensor],
tuple[FloatTensor, FloatTensor, FloatTensor, FloatTensor, FloatTensor, FloatTensor],
tuple[FloatTensor, FloatTensor],
]
):
r"""
The BoxE interaction from [abboud2020]_.
Entities are represented by two $d$-dimensional vectors describing the *base position* as well
as the *translational bump*, which translates all the entities co-occuring in a fact with this entity
from their base positions to their final embeddings, called "bumping".
Relations are represented as a fixed number of hyper-rectangles corresponding to the relation's arity.
Since we are only considering single-hop link predition here, the arity is always two, i.e., one box
for the head position and another one for the tail position. There are different possibilities to
parametrize a hyper-rectangle, where the most common may be its description as the coordinate of to
opposing vertices. BoxE suggests a different parametrization for each box by
- a base position given by its center, a $d$-dimensional vector $\mathbf{c} \in \mathbb{R}^d$
- an extent in each dimension. This size is further factored in
- a scalar global scalar scaling factor, $s \in \mathbb{R}$
- a normalized extent in each dimension, i.e., the extents sum to one, given as $\mathbf{e} \in \mathbb{R}^d$,
with $\|\mathbf{e}\| = 1$ and $0 \leq \mathbf{e}_i$ for all $i$.
---
citation:
author: Abboud
year: 2020
link: https://arxiv.org/abs/2007.06267
github: ralphabb/BoxE
"""
relation_shape = ("d", "d", "s", "d", "d", "s") # Boxes are 2xd (size) each, x 2 sets of boxes: head and tail
entity_shape = ("d", "d") # Base position and bump
def __init__(self, tanh_map: bool = True, p: int = 2, power_norm: bool = False):
r"""
Instantiate the interaction module.
.. seealso::
The parameter ``p`` and ``power_norm`` are directly passed to
:class:`~pykeen.nn.modules.NormBasedInteraction`.
:param tanh_map:
Whether to use tanh mapping after BoxE computation (defaults to true). The hyperbolic tangent mapping
restricts the embedding space to the range [-1, 1], and thus this map implicitly
regularizes the space to prevent loss reduction by growing boxes arbitrarily large.
:param p:
The norm used with :func:`torch.linalg.vector_norm`. Typically is 1 or 2.
:param power_norm:
Whether to use the p-th power of the $L_p$ norm. It has the advantage of being differentiable around 0,
and numerically more stable.
"""
super().__init__(p=p, power_norm=power_norm)
self.tanh_map = tanh_map
[docs]
def forward(
self,
h: tuple[FloatTensor, FloatTensor],
r: tuple[FloatTensor, FloatTensor, FloatTensor, FloatTensor, FloatTensor, FloatTensor],
t: tuple[FloatTensor, FloatTensor],
) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)`` and ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``, ``(*batch_dims, d)``, ``(*batch_dims, s)``,
``(*batch_dims, d)``, ``(*batch_dims, d)``, and ``(*batch_dims, s)``
The relation representations.
:param t: shape: ``(*batch_dims, d)`` and ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
# relation box head; relation box tail
rh_base, rh_delta, rh_size, rt_base, rt_delta, rt_size = r
# head position and bump
h_pos, h_bump = h
# tail position and bump
t_pos, t_bump = t
# head score
return BoxEInteraction.boxe_kg_arity_position_score(
# head box score
entity_pos=h_pos,
other_entity_bump=t_bump,
relation_box=BoxEInteraction.compute_box(base=rh_base, delta=rh_delta, size=rh_size),
tanh_map=self.tanh_map,
p=self.p,
power_norm=self.power_norm,
) + BoxEInteraction.boxe_kg_arity_position_score(
# tail box score
entity_pos=t_pos,
other_entity_bump=h_bump,
relation_box=BoxEInteraction.compute_box(base=rt_base, delta=rt_delta, size=rt_size),
tanh_map=self.tanh_map,
p=self.p,
power_norm=self.power_norm,
)
[docs]
@staticmethod
def product_normalize(x: FloatTensor, dim: int = -1) -> FloatTensor:
r"""Normalize a tensor along a given dimension so that the geometric mean is 1.0.
:param x: shape: s
An input tensor
:param dim:
the dimension along which to normalize the tensor
:return: shape: s
An output tensor where the given dimension is normalized to have a geometric mean of 1.0.
"""
return x / at_least_eps(at_least_eps(x.abs()).log().mean(dim=dim, keepdim=True).exp())
[docs]
@staticmethod
def point_to_box_distance(
points: FloatTensor,
box_lows: FloatTensor,
box_highs: FloatTensor,
) -> FloatTensor:
r"""Compute the point to box distance function proposed by [abboud2020]_ in an element-wise fashion.
:param points: shape: ``(*, d)``
the positions of the points being scored against boxes
:param box_lows: shape: ``(*, d)``
the lower corners of the boxes
:param box_highs: shape: ``(*, d)``
the upper corners of the boxes
:returns:
Element-wise distance function scores as per the definition above
Given points $p$, box_lows $l$, and box_highs $h$, the following quantities are
defined:
- Width $w$ is the difference between the upper and lower box bound: $w = h - l$
- Box centers $c$ are the mean of the box bounds: $c = (h + l) / 2$
Finally, the point to box distance $dist(p,l,h)$ is defined as
the following piecewise function:
.. math::
dist(p,l,h) = \begin{cases}
|p-c|/(w+1) & l <= p <+ h \\
|p-c|*(w+1) - 0.5*w*((w+1)-1/(w+1)) & otherwise \\
\end{cases}
"""
widths = box_highs - box_lows
# compute width plus 1
widths_p1 = widths + 1
# compute box midpoints
# TODO: we already had this before, as `base`
centres = 0.5 * (box_lows + box_highs)
return torch.where(
# inside box?
torch.logical_and(points >= box_lows, points <= box_highs),
# yes: |p - c| / (w + 1)
torch.abs(points - centres) / widths_p1,
# no: (w + 1) * |p - c| - 0.5 * w * (w - 1/(w + 1))
widths_p1 * torch.abs(points - centres) - (0.5 * widths) * (widths_p1 - 1 / widths_p1),
)
[docs]
@classmethod
def boxe_kg_arity_position_score(
cls,
entity_pos: FloatTensor,
other_entity_bump: FloatTensor,
relation_box: tuple[FloatTensor, FloatTensor],
tanh_map: bool,
p: int,
power_norm: bool,
) -> FloatTensor:
r"""Perform the BoxE computation at a single arity position.
.. note::
this computation is parallelizable across all positions
.. note ::
`entity_pos`, `other_entity_bump`, `relation_box_low` and `relation_box_high` have to be in broadcastable
shape.
:param entity_pos: shape: ``(*s_p, d)``
This is the base entity position of the entity appearing in the target position. For example,
for a fact $r(h, t)$ and the head arity position, `entity_pos` is the base position of $h$.
:param other_entity_bump: shape: ``(*s_b, d)``
This is the bump of the entity at the other position in the fact. For example, given a
fact $r(h, t)$ and the head arity position, `other_entity_bump` is the bump of $t$.
:param relation_box: shape: ``(*s_r, d)``
The lower/upper corner of the relation box at the target arity position.
:param tanh_map:
whether to apply the tanh map regularizer
:param p:
The norm used with :func:`torch.linalg.vector_norm`. Typically is 1 or 2.
:param power_norm:
Whether to use the p-th power of the $L_p$ norm. It has the advantage of being differentiable around 0,
and numerically more stable.
:return: shape: ``*s``
Arity-position score for the entity relative to the target relation box. Larger is better. The shape is the
broadcasted shape from position, bump and box, where the last dimension has been removed.
"""
# Step 1: Apply the other entity bump
bumped_representation = entity_pos + other_entity_bump
relation_box_low, relation_box_high = relation_box
# Step 2: Apply tanh if tanh_map is set to True.
if tanh_map:
relation_box_low = torch.tanh(relation_box_low)
relation_box_high = torch.tanh(relation_box_high)
bumped_representation = torch.tanh(bumped_representation)
# Compute the distance function output element-wise
element_wise_distance = cls.point_to_box_distance(
points=bumped_representation,
box_lows=relation_box_low,
box_highs=relation_box_high,
)
# Finally, compute the norm
return negative_norm(element_wise_distance, p=p, power_norm=power_norm)
[docs]
@classmethod
def compute_box(
cls,
base: FloatTensor,
delta: FloatTensor,
size: FloatTensor,
) -> tuple[FloatTensor, FloatTensor]:
r"""Compute the lower and upper corners of a resulting box.
:param base: shape: ``(*, d)``
the base position (box center) of the input relation embeddings
:param delta: shape: ``(*, d)``
the base shape of the input relation embeddings
:param size: shape: ``(*, d)``
the size scalar vectors of the input relation embeddings
:return: shape: ``(*, d)`` each
lower and upper bounds of the box whose embeddings are provided as input.
"""
# Enforce that sizes are strictly positive by passing through ELU
size_pos = torch.nn.functional.elu(size) + 1
# Shape vector is normalized using the above helper function
delta_norm = cls.product_normalize(delta)
# Size is learned separately and applied to normalized shape
delta_final = size_pos * delta_norm
# Compute potential boundaries by applying the shape in substraction
first_bound = base - 0.5 * delta_final
# and in addition
second_bound = base + 0.5 * delta_final
# Compute box upper bounds using min and max respectively
box_low = torch.minimum(first_bound, second_bound)
box_high = torch.maximum(first_bound, second_bound)
return box_low, box_high
[docs]
@parse_docdata
class CPInteraction(Interaction[FloatTensor, FloatTensor, FloatTensor]):
r"""
The Canonical Tensor Decomposition interaction as described [lacroix2018]_ (originally from [hitchcock1927]_).
The interaction function is given as
.. math::
\sum_{1 \leq i \leq k, 1 \leq j \leq d} \mathbf{h}_{i, j} \cdot \mathbf{r}_{i, j} \cdot \mathbf{t}_{i, j}
.. note ::
For $k=1$, this interaction is the same as :class:`~pykeen.nn.modules.DistMultInteraction`.
However, in contrast to :class:`~pykeen.models.DistMult`, entities should have different representations for the
head and the tail role.
---
name: Canonical Tensor Decomposition
citation:
author: Lacroix
year: 2018
arxiv: 1806.07297
link: https://arxiv.org/abs/1806.07297
github: facebookresearch/kbc
"""
entity_shape = ("kd", "kd")
relation_shape = ("kd",)
_head_indices = (0,)
_tail_indices = (1,)
[docs]
def forward(self, h: FloatTensor, r: FloatTensor, t: FloatTensor) -> FloatTensor:
r"""
Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, k, d)``
The head representations.
:param r: shape: ``(*batch_dims, k, d)``
The relation representations.
:param t: shape: ``(*batch_dims, k, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
return einsum("...kd, ...kd, ...kd -> ...", h, r, t)
[docs]
@parse_docdata
class MultiLinearTuckerInteraction(Interaction[FloatTensor, FloatTensor, FloatTensor]):
"""
An implementation of the original (multi-linear) TuckER interaction as described [tucker1966]_.
.. note ::
For small tensors, there are more efficient algorithms to compute the decomposition, e.g.,
http://tensorly.org/stable/modules/generated/tensorly.decomposition.Tucker.html
---
name: MultiLinearTucker
citation:
author: Tucker
year: 1966
link: https://dx.doi.org/10.1007/BF02289464
"""
entity_shape = ("d", "f")
_head_indices = (0,)
_tail_indices = (1,)
relation_shape = ("e",)
def __init__(
self,
head_dim: int = 64,
relation_dim: int | None = None,
tail_dim: int | None = None,
):
"""
Initialize the Tucker interaction function.
:param head_dim:
The head entity embedding dimension.
:param relation_dim:
The relation embedding dimension. Defaults to `head_dim`.
:param tail_dim:
The tail entity embedding dimension. Defaults to `head_dim`.
"""
super().__init__()
# input normalization
relation_dim = relation_dim or head_dim
tail_dim = tail_dim or head_dim
# Core tensor
self.core_tensor = nn.Parameter(
torch.empty(head_dim, relation_dim, tail_dim),
requires_grad=True,
)
# docstr-coverage: inherited
[docs]
def reset_parameters(self): # noqa: D102
# initialize core tensor
nn.init.normal_(
self.core_tensor,
mean=0,
std=numpy.sqrt(numpy.prod(numpy.reciprocal(numpy.asarray(self.core_tensor.shape)))),
)
[docs]
def forward(self, h: FloatTensor, r: FloatTensor, t: FloatTensor) -> FloatTensor:
r"""
Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, head_dim)``
The head representations.
:param r: shape: ``(*batch_dims, relation_dim)``
The relation representations.
:param t: shape: ``(*batch_dims, tail_dim)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
return einsum("ijk, ...i, ...j, ...k -> ...", self.core_tensor, h, r, t)
[docs]
@parse_docdata
class TripleREInteraction(NormBasedInteraction[FloatTensor, tuple[FloatTensor, FloatTensor, FloatTensor], FloatTensor]):
r"""The TripleRE interaction function from [yu2021]_.
It is given by
.. math ::
\mathbf{h} \odot (\mathbf{r}_h + u) - \mathbf{t} \odot (\mathbf{r}_t + u) + \mathbf{r}
with head entity, relation and tail entity representations $\mathbf{h}, \mathbf{r}, \mathbf{t} \in \mathbb{R}^d$,
relation specific head and tail multipliers $\mathbf{r}_h, \mathbf{r}_t \in \mathbb{R}^d$,
and a scalar relation factor offset $u \in \mathbb{R}$.
.. note ::
This interaction is equivalent to :class:`~pykeen.nn.modules.LineaREInteraction` except the $u$ term.
The $u$ is only non-zero for the version 2 from the paper.
.. note ::
For equivalence to the paper version, `h` and `t` should be normalized to unit
Euclidean length, and `p` and `power_norm` be kept at their default values.
.. seealso:: https://github.com/LongYu-360/TripleRE-Add-NodePiece
---
name: TripleRE
citation:
author: Yu
year: 2021
link: https://vixra.org/abs/2112.0095
"""
# r_head, r_mid, r_tail
relation_shape = ("d", "d", "d")
def __init__(self, u: float | None = 1.0, p: int = 1, power_norm: bool = False):
"""
Initialize the module.
.. seealso::
The parameter ``p`` and ``power_norm`` are directly passed to
:class:`~pykeen.nn.modules.NormBasedInteraction`.
:param u:
The relation factor offset. Can be set to `None` (or 0) to disable it.
:param p:
The norm used with :func:`torch.linalg.vector_norm`. Typically is 1 or 2.
:param power_norm:
Whether to use the p-th power of the $L_p$ norm. It has the advantage of being differentiable around 0,
and numerically more stable.
"""
super().__init__(p=p, power_norm=power_norm)
self.u = u
[docs]
def forward(self, h: FloatTensor, r: tuple[FloatTensor, FloatTensor, FloatTensor], t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``, 3 times
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
u = self.u
r_head, r_mid, r_tail = r
# note: normalization should be done from the representations
# cf. https://github.com/LongYu-360/TripleRE-Add-NodePiece/blob/994216dcb1d718318384368dd0135477f852c6a4/TripleRE%2BNodepiece/ogb_wikikg2/model.py#L317-L328 # noqa: E501
# version 2
if u is not None:
# r_head = r_head + u * torch.ones_like(r_head)
# r_tail = r_tail + u * torch.ones_like(r_tail)
r_head = r_head + u
r_tail = r_tail + u
return negative_norm_of_sum(h * r_head, -t * r_tail, r_mid, p=self.p, power_norm=self.power_norm)
# type alias for AutoSF block description
# head_index, relation_index, tail_index, sign
AutoSFBlock = tuple[int, int, int, Sign]
[docs]
@parse_docdata
class AutoSFInteraction(FunctionalInteraction[HeadRepresentation, RelationRepresentation, TailRepresentation]):
r"""
The AutoSF interaction as described by [zhang2020]_.
This interaction function is a parametrized way to express bi-linear models
with block structure. It divides the entity and relation representations into blocks,
and expresses the interaction as a sequence of 4-tuples $(i_h, i_r, i_t, s)$,
where $i_h, i_r, i_t$ index a _block_ of the head, relation, or tail representation,
and $s \in {-1, 1}$ is the sign.
The interaction function is then given as
.. math::
\sum_{(i_h, i_r, i_t, s) \in \mathcal{C}} s \cdot \langle h[i_h], r[i_r], t[i_t] \rangle
where $\langle \cdot, \cdot, \cdot \rangle$ denotes the tri-linear dot product.
This parametrization allows to express several well-known interaction functions, e.g.
- :class:`pykeen.nn.DistMultInteraction`:
one block, $\mathcal{C} = \{(0, 0, 0, 1)\}$
- :class:`pykeen.nn.ComplExInteraction`:
two blocks, $\mathcal{C} = \{(0, 0, 0, 1), (0, 1, 1, 1), (1, 0, 1, -1), (1, 0, 1, 1)\}$
- :class:`pykeen.nn.SimplEInteraction`:
two blocks: $\mathcal{C} = \{(0, 0, 1, 1), (1, 1, 0, 1)\}$
While in theory, we can have up to `num_blocks**3` unique triples, usually, a smaller number is preferable to have
some sparsity.
---
citation:
author: Zhang
year: 2020
arxiv: 1904.11682
link: https://arxiv.org/abs/1904.11682
github: AutoML-Research/AutoSF
"""
#: a description of the block structure
coefficients: tuple[AutoSFBlock, ...]
@staticmethod
def _check_coefficients(
coefficients: Collection[AutoSFBlock], num_entity_representations: int, num_relation_representations: int
):
"""Check coefficients.
:param coefficients:
the block description
:param num_entity_representations:
the number of entity representations / blocks
:param num_relation_representations:
the number of relation representations / blocks
:raises ValueError:
if there are duplicate coefficients
"""
counter = Counter(coef[:3] for coef in coefficients)
duplicates = {k for k, v in counter.items() if v > 1}
if duplicates:
raise ValueError(f"Cannot have duplicates in coefficients! Duplicate entries for {duplicates}")
for entities, num_blocks in ((True, num_entity_representations), (False, num_relation_representations)):
missing_ids = set(range(num_blocks)).difference(
AutoSFInteraction._iter_ids(coefficients, entities=entities)
)
if missing_ids:
label = "entity" if entities else "relation"
logger.warning(f"Unused {label} blocks: {missing_ids}. This may indicate an error.")
@staticmethod
def _iter_ids(coefficients: Collection[AutoSFBlock], entities: bool) -> Iterable[int]:
"""Iterate over selected parts of the blocks.
:param coefficients:
the block coefficients
:param entities:
whether to select entity or relation ids, i.e., components `(0, 2)` for entities, or `(1,)` for relations.
:yields: the used indices
"""
indices = (0, 2) if entities else (1,)
yield from itt.chain.from_iterable(map(itemgetter(i), coefficients) for i in indices)
@staticmethod
def _infer_number(coefficients: Collection[AutoSFBlock], entities: bool) -> int:
"""Infer the number of blocks from the given coefficients.
:param coefficients:
the block coefficients
:param entities:
whether to select entity or relation ids, i.e., components `(0, 2)` for entities, or `(1,)` for relations.
:return:
the inferred number of blocks
"""
return 1 + max(AutoSFInteraction._iter_ids(coefficients, entities=entities))
def __init__(
self,
coefficients: Iterable[AutoSFBlock],
*,
num_blocks: int | None = None,
num_entity_representations: int | None = None,
num_relation_representations: int | None = None,
) -> None:
"""
Initialize the interaction function.
:param coefficients:
the coefficients for the individual blocks, cf. :class:`pykeen.nn.AutoSFInteraction`
:param num_blocks:
the number of blocks. If given, will be used for both, entity and relation representations.
:param num_entity_representations:
an explicit number of entity representations / blocks. Only used if `num_blocks` is `None`.
If `num_entity_representations` is `None`, too, this number if inferred from `coefficients`.
:param num_relation_representations:
an explicit number of relation representations / blocks. Only used if `num_blocks` is `None`.
If `num_relation_representations` is `None`, too, this number if inferred from `coefficients`.
"""
super().__init__()
# convert to tuple
coefficients = tuple(coefficients)
# infer the number of entity and relation representations
num_entity_representations = (
num_blocks or num_entity_representations or self._infer_number(coefficients, entities=True)
)
num_relation_representations = (
num_blocks or num_relation_representations or self._infer_number(coefficients, entities=False)
)
# verify coefficients
self._check_coefficients(
coefficients=coefficients,
num_entity_representations=num_entity_representations,
num_relation_representations=num_relation_representations,
)
self.coefficients = coefficients
# dynamic entity / relation shapes
self.entity_shape = tuple(["d"] * num_entity_representations)
self.relation_shape = tuple(["d"] * num_relation_representations)
[docs]
@classmethod
def from_searched_sf(cls, coefficients: Sequence[int], **kwargs) -> AutoSFInteraction:
"""
Instantiate AutoSF interaction from the "official" serialization format.
> The first 4 values (a,b,c,d) represent h_1 * r_1 * t_a + h_2 * r_2 * t_b + h_3 * r_3 * t_c + h_4 * r_4 * t_d.
> For the others, every 4 values represent one adding block: index of r, index of h, index of t, the sign s.
:param coefficients:
the coefficients in the "official" serialization format.
:param kwargs:
additional keyword-based parameters passed to :meth:`pykeen.nn.AutoSFInteraction.__init__`
:return:
An AutoSF interaction module
.. seealso::
https://github.com/AutoML-Research/AutoSF/blob/07b7243ccf15e579176943c47d6e65392cd57af3/searched_SFs.txt
"""
return cls(
coefficients=[(i, ri, i, 1) for i, ri in enumerate(coefficients[:4])]
+ [(hi, ri, ti, s) for ri, hi, ti, s in more_itertools.chunked(coefficients[4:], 4)],
**kwargs,
)
[docs]
@staticmethod
def func(
h: HeadRepresentation,
r: RelationRepresentation,
t: TailRepresentation,
coefficients: Collection[AutoSFBlock],
) -> FloatTensor:
r"""Evaluate an AutoSF-style interaction function as described by [zhang2020]_.
:param h: each shape: (`*batch_dims`, dim)
The list of head representations.
:param r: each shape: (`*batch_dims`, dim)
The list of relation representations.
:param t: each shape: (`*batch_dims`, dim)
The list of tail representations.
:param coefficients:
the coefficients, cf. :class:`pykeen.nn.AutoSFInteraction`
:return: shape: `batch_dims`
The scores
"""
return sum(sign * tensor_product(h[hi], r[ri], t[ti]).sum(dim=-1) for hi, ri, ti, sign in coefficients)
def _prepare_state_for_functional(self) -> MutableMapping[str, Any]:
return dict(coefficients=self.coefficients)
# docstr-coverage: inherited
@staticmethod
def _prepare_hrt_for_functional(
h: HeadRepresentation,
r: RelationRepresentation,
t: TailRepresentation,
) -> MutableMapping[str, FloatTensor]: # noqa: D102
return dict(zip("hrt", ensure_tuple(h, r, t)))
[docs]
def extend(self, *new_coefficients: tuple[int, int, int, Sign]) -> AutoSFInteraction:
"""Extend AutoSF function, as described in the greedy search algorithm in the paper."""
return AutoSFInteraction(coefficients=self.coefficients + tuple(new_coefficients))
[docs]
def latex_visualize(self) -> str:
"""Create the LaTeX + tikz visualization as shown in the paper."""
n = len(self.entity_shape)
return "\n".join(
[
r"\begin{tikzpicture}[yscale=-1]",
rf"\draw (0, 0) grid ({n}, {n});",
]
+ [
rf"\draw ({ti}.5, {hi}.5) node {{${'-' if s < 0 else ''}D^r_{{{ri + 1}}}$}};"
for hi, ri, ti, s in self.coefficients
]
+ [
r"\end{tikzpicture}",
],
)
[docs]
@parse_docdata
class LineaREInteraction(NormBasedInteraction[FloatTensor, tuple[FloatTensor, FloatTensor, FloatTensor], FloatTensor]):
r"""
The LineaRE interaction described by [peng2020]_.
It is given by
.. math ::
\mathbf{h} \odot \mathbf{r}_h - \mathbf{t} \odot \mathbf{r}_t + \mathbf{r}
where $\mathbf{r}_{h}, \mathbf{r}, \mathbf{r}_{t} \in \mathbb{R}^d$ are relation-specific terms,
and $\mathbf{h}, \mathbf{t} \in \mathbb{R}^n$ the head and tail entity representation.
.. note ::
the original paper only describes the interaction for $L_1$ norm, but we extend it to the general $L_p$
norm as well as its powered variant.
.. note ::
This interaction is equivalent to :class:`TripleREInteraction` without the $u$ term.
---
name: LineaRE
citation:
author: Peng
year: 2020
arxiv: 2004.10037
github: pengyanhui/LineaRE
link: https://arxiv.org/abs/2004.10037
"""
# r_head, r_bias, r_tail
relation_shape = ("d", "d", "d")
[docs]
def forward(self, h: FloatTensor, r: tuple[FloatTensor, FloatTensor, FloatTensor], t: FloatTensor) -> FloatTensor:
"""Evaluate the interaction function.
.. seealso::
:meth:`Interaction.forward <pykeen.nn.modules.Interaction.forward>` for a detailed description about
the generic batched form of the interaction function.
:param h: shape: ``(*batch_dims, d)``
The head representations.
:param r: shape: ``(*batch_dims, d)``, 3 times
The relation representations.
:param t: shape: ``(*batch_dims, d)``
The tail representations.
:return: shape: ``batch_dims``
The scores.
"""
r_head, r_mid, r_tail = r
return negative_norm_of_sum(h * r_head, -t * r_tail, r_mid, p=self.p, power_norm=self.power_norm)
#: A resolver for stateful interaction functions
interaction_resolver: ClassResolver[Interaction] = ClassResolver.from_subclasses(
Interaction,
skip={
NormBasedInteraction,
FunctionalInteraction,
MonotonicAffineTransformationInteraction,
ClampedInteraction,
DirectionAverageInteraction,
},
default=TransEInteraction,
)