Source code for pykeen.models.base

# -*- coding: utf-8 -*-

"""Base module for all KGE models."""

from __future__ import annotations

import inspect
import logging
import os
import pickle
from abc import ABC, abstractmethod
from typing import Any, ClassVar, Iterable, Mapping, Optional, Type, Union

import torch
from class_resolver import HintOrType
from docdata import parse_docdata
from torch import nn

from ..losses import Loss, MarginRankingLoss, loss_resolver
from ..triples import KGInfo, relation_inverter
from ..typing import LABEL_HEAD, LABEL_RELATION, LABEL_TAIL, InductiveMode, MappedTriples, Target
from ..utils import NoRandomSeedNecessary, get_preferred_device, set_random_seed

__all__ = [
    "Model",
]

logger = logging.getLogger(__name__)


[docs]class Model(nn.Module, ABC): """A base module for KGE models. Subclasses of :class:`Model` can decide however they want on how to store entities' and relations' representations, how they want to be looked up, and how they should be scored. The :class:`OModel` provides a commonly used interface for models storing entity and relation representations in the form of :class:`pykeen.nn.Embedding`. """ #: The default strategy for optimizing the model's hyper-parameters hpo_default: ClassVar[Mapping[str, Any]] _random_seed: Optional[int] #: The default loss function class loss_default: ClassVar[Type[Loss]] = MarginRankingLoss #: The default parameters for the default loss function class loss_default_kwargs: ClassVar[Optional[Mapping[str, Any]]] = dict(margin=1.0, reduction="mean") #: The instance of the loss loss: Loss num_entities: int num_relations: int use_inverse_triples: bool can_slice_h: ClassVar[bool] can_slice_r: ClassVar[bool] can_slice_t: ClassVar[bool] def __init__( self, *, triples_factory: KGInfo, loss: HintOrType[Loss] = None, loss_kwargs: Optional[Mapping[str, Any]] = None, predict_with_sigmoid: bool = False, random_seed: Optional[int] = None, ) -> None: """Initialize the module. :param triples_factory: The triples factory facilitates access to the dataset. :param loss: The loss to use. If None is given, use the loss default specific to the model subclass. :param loss_kwargs: keyword-based parameters passed to the loss instance upon instantiation :param predict_with_sigmoid: Whether to apply sigmoid onto the scores when predicting scores. Applying sigmoid at prediction time may lead to exactly equal scores for certain triples with very high, or very low score. When not trained with applying sigmoid (or using BCEWithLogitsLoss), the scores are not calibrated to perform well with sigmoid. :param random_seed: A random seed to use for initialising the model's weights. **Should** be set when aiming at reproducibility. """ super().__init__() # Random seeds have to set before the embeddings are initialized if random_seed is None: logger.warning("No random seed is specified. This may lead to non-reproducible results.") self._random_seed = None elif random_seed is not NoRandomSeedNecessary: set_random_seed(random_seed) self._random_seed = random_seed # Loss if loss is None: self.loss = self.loss_default(**(self.loss_default_kwargs or {})) else: self.loss = loss_resolver.make(loss, pos_kwargs=loss_kwargs) self.use_inverse_triples = triples_factory.create_inverse_triples self.num_entities = triples_factory.num_entities self.num_relations = triples_factory.num_relations """ When predict_with_sigmoid is set to True, the sigmoid function is applied to the logits during evaluation and also for predictions after training, but has no effect on the training. """ self.predict_with_sigmoid = predict_with_sigmoid @property def num_real_relations(self) -> int: """Return the real number of relations (without inverses).""" if self.use_inverse_triples: return self.num_relations // 2 return self.num_relations def __init_subclass__(cls, **kwargs): """Initialize the subclass. This checks for all subclasses if they are tagged with :class:`abc.ABC` with :func:`inspect.isabstract`. All non-abstract deriving models should have citation information. Subclasses can further override ``__init_subclass__``, but need to remember to call ``super().__init_subclass__`` as well so this gets run. :param kwargs: ignored keyword-based parameters """ if not inspect.isabstract(cls): parse_docdata(cls) @property def device(self) -> torch.device: """Return the model's device.""" return get_preferred_device(self, allow_ambiguity=False)
[docs] def reset_parameters_(self): # noqa: D401 """Reset all parameters of the model and enforce model constraints.""" self._reset_parameters_() # TODO: why do we need to empty the cache? torch.cuda.empty_cache() self.post_parameter_update() return self
"""Base methods"""
[docs] def post_forward_pass(self): """Run after calculating the forward loss."""
def _free_graph_and_cache(self): """Run to free the graph and cache.""" """Abstract methods""" @abstractmethod def _reset_parameters_(self): # noqa: D401 """Reset all parameters of the model in-place.""" @abstractmethod def _get_entity_len(self, *, mode: Optional[InductiveMode]) -> Optional[int]: """Get the number of entities depending on the mode parameters."""
[docs] def post_parameter_update(self) -> None: """Has to be called after each parameter update."""
"""Abstract methods - Scoring"""
[docs] @abstractmethod def score_hrt(self, hrt_batch: torch.LongTensor, *, mode: Optional[InductiveMode] = None) -> torch.FloatTensor: """Forward pass. This method takes head, relation and tail of each triple and calculates the corresponding score. :param hrt_batch: shape: (batch_size, 3), dtype: long The indices of (head, relation, tail) triples. :param mode: The pass mode, which is None in the transductive setting and one of "training", "validation", or "testing" in the inductive setting. :return: shape: (batch_size, 1), dtype: float The score for each triple. """
[docs] @abstractmethod def score_t( self, hr_batch: torch.LongTensor, *, slice_size: Optional[int] = None, mode: Optional[InductiveMode] = None, tails: Optional[torch.LongTensor] = None, ) -> torch.FloatTensor: """Forward pass using right side (tail) prediction. This method calculates the score for all possible tails for each (head, relation) pair. :param hr_batch: shape: (batch_size, 2), dtype: long The indices of (head, relation) pairs. :param slice_size: >0 The divisor for the scoring function when using slicing. :param mode: The pass mode, which is None in the transductive setting and one of "training", "validation", or "testing" in the inductive setting. :param tails: shape: (num_tails,) | (batch_size, num_tails) tail entity indices to score against. If `None`, scores against all entities (from the given mode). :return: shape: (batch_size, num_tails), dtype: float For each h-r pair, the scores for all possible tails. """
[docs] @abstractmethod def score_r( self, ht_batch: torch.LongTensor, *, slice_size: Optional[int] = None, mode: Optional[InductiveMode] = None, relations: Optional[torch.LongTensor] = None, ) -> torch.FloatTensor: """Forward pass using middle (relation) prediction. This method calculates the score for all possible relations for each (head, tail) pair. :param ht_batch: shape: (batch_size, 2), dtype: long The indices of (head, tail) pairs. :param slice_size: >0 The divisor for the scoring function when using slicing. :param mode: The pass mode, which is None in the transductive setting and one of "training", "validation", or "testing" in the inductive setting. :param relations: shape: (num_relations,) | (batch_size, num_relations) relation indices to score against. If None, scores against all relations (from the given mode). :return: shape: (batch_size, num_real_relations), dtype: float For each h-t pair, the scores for all possible relations. """
# TODO: this currently compute (batch_size, num_relations) instead, # i.e., scores for normal and inverse relations
[docs] @abstractmethod def score_h( self, rt_batch: torch.LongTensor, *, slice_size: Optional[int] = None, mode: Optional[InductiveMode] = None, heads: Optional[torch.LongTensor] = None, ) -> torch.FloatTensor: """Forward pass using left side (head) prediction. This method calculates the score for all possible heads for each (relation, tail) pair. :param rt_batch: shape: (batch_size, 2), dtype: long The indices of (relation, tail) pairs. :param slice_size: >0 The divisor for the scoring function when using slicing. :param mode: The pass mode, which is None in the transductive setting and one of "training", "validation", or "testing" in the inductive setting. :param heads: shape: (num_heads,) | (batch_size, num_heads) head entity indices to score against. If None, scores against all entities (from the given mode). :return: shape: (batch_size, num_heads), dtype: float For each r-t pair, the scores for all possible heads. """
[docs] @abstractmethod def collect_regularization_term(self) -> torch.FloatTensor: """Get the regularization term for the loss function."""
"""Concrete methods"""
[docs] def get_grad_params(self) -> Iterable[nn.Parameter]: """Get the parameters that require gradients.""" # TODO: Why do we need that? The optimizer takes care of filtering the parameters. return filter(lambda p: p.requires_grad, self.parameters())
@property def num_parameter_bytes(self) -> int: """Calculate the number of bytes used for all parameters of the model.""" return sum(param.numel() * param.element_size() for param in self.parameters(recurse=True)) @property def num_parameters(self) -> int: """Calculate the number of parameters of the model.""" return sum(param.numel() for param in self.parameters(recurse=True))
[docs] def save_state(self, path: Union[str, os.PathLike]) -> None: """Save the state of the model. :param path: Path of the file where to store the state in. """ torch.save(self.state_dict(), path, pickle_protocol=pickle.HIGHEST_PROTOCOL)
[docs] def load_state(self, path: Union[str, os.PathLike]) -> None: """Load the state of the model. :param path: Path of the file where to load the state from. """ self.load_state_dict(torch.load(path, map_location=self.device))
"""Prediction methods""" def _prepare_batch(self, batch: torch.LongTensor, index_relation: int) -> torch.LongTensor: # send to device batch = batch.to(self.device) # special handling of inverse relations if not self.use_inverse_triples: return batch # when trained on inverse relations, the internal relation ID is twice the original relation ID return relation_inverter.map(batch=batch, index=index_relation, invert=False)
[docs] def predict_hrt(self, hrt_batch: torch.LongTensor, *, mode: Optional[InductiveMode] = None) -> torch.FloatTensor: """Calculate the scores for triples. This method takes head, relation and tail of each triple and calculates the corresponding score. Additionally, the model is set to evaluation mode. :param hrt_batch: shape: (number of triples, 3), dtype: long The indices of (head, relation, tail) triples. :param mode: The pass mode. Is None for transductive and "training" / "validation" / "testing" in inductive. :return: shape: (number of triples, 1), dtype: float The score for each triple. """ self.eval() # Enforce evaluation mode scores = self.score_hrt(self._prepare_batch(batch=hrt_batch, index_relation=1), mode=mode) if self.predict_with_sigmoid: scores = torch.sigmoid(scores) return scores
[docs] def predict_h( self, rt_batch: torch.LongTensor, **kwargs, ) -> torch.FloatTensor: """Forward pass using left side (head) prediction for obtaining scores of all possible heads. This method calculates the score for all possible heads for each (relation, tail) pair. .. note:: If the model has been trained with inverse relations, the task of predicting the head entities becomes the task of predicting the tail entities of the inverse triples, i.e., $f(*,r,t)$ is predicted by means of $f(t,r_{inv},*)$. Additionally, the model is set to evaluation mode. :param rt_batch: shape: (batch_size, 2), dtype: long The indices of (relation, tail) pairs. :param kwargs: additional keyword-based parameters passed to :meth:`Model.score_h` :return: shape: (batch_size, num_heads), dtype: float For each r-t pair, the scores for all possible heads. """ self.eval() # Enforce evaluation mode rt_batch = self._prepare_batch(batch=rt_batch, index_relation=0) if self.use_inverse_triples: scores = self.score_h_inverse(rt_batch=rt_batch, **kwargs) else: scores = self.score_h(rt_batch, **kwargs) if self.predict_with_sigmoid: scores = torch.sigmoid(scores) return scores
[docs] def predict_t( self, hr_batch: torch.LongTensor, **kwargs, ) -> torch.FloatTensor: """Forward pass using right side (tail) prediction for obtaining scores of all possible tails. This method calculates the score for all possible tails for each (head, relation) pair. Additionally, the model is set to evaluation mode. :param hr_batch: shape: (batch_size, 2), dtype: long The indices of (head, relation) pairs. :param kwargs: additional keyword-based parameters passed to :meth:`Model.score_t` :return: shape: (batch_size, num_tails), dtype: float For each h-r pair, the scores for all possible tails. .. note:: We only expect the right side-predictions, i.e., $(h,r,*)$ to change its default behavior when the model has been trained with inverse relations (mainly because of the behavior of the LCWA training approach). This is why the :func:`predict_h` has different behavior depending on if inverse triples were used in training, and why this function has the same behavior regardless of the use of inverse triples. """ self.eval() # Enforce evaluation mode hr_batch = self._prepare_batch(batch=hr_batch, index_relation=1) scores = self.score_t(hr_batch, **kwargs) if self.predict_with_sigmoid: scores = torch.sigmoid(scores) return scores
[docs] def predict_r( self, ht_batch: torch.LongTensor, **kwargs, ) -> torch.FloatTensor: """Forward pass using middle (relation) prediction for obtaining scores of all possible relations. This method calculates the score for all possible relations for each (head, tail) pair. Additionally, the model is set to evaluation mode. :param ht_batch: shape: (batch_size, 2), dtype: long The indices of (head, tail) pairs. :param kwargs: additional keyword-based parameters passed to :meth:`Model.score_r` :return: shape: (batch_size, num_relations), dtype: float For each h-t pair, the scores for all possible relations. """ self.eval() # Enforce evaluation mode ht_batch = ht_batch.to(self.device) scores = self.score_r(ht_batch, **kwargs) if self.predict_with_sigmoid: scores = torch.sigmoid(scores) return scores
[docs] def predict( self, hrt_batch: MappedTriples, target: Target, full_batch: bool = True, ids: Optional[torch.LongTensor] = None, **kwargs, ) -> torch.FloatTensor: """ Predict scores for the given target. :param hrt_batch: shape: (batch_size, 3) or (batch_size, 2) the full batch, or the relevant part of it :param target: the target to predict :param full_batch: whether `hrt_batch` is the full batch, or only the "input" part of the target prediction method :param ids: restrict prediction to only those ids :param kwargs: additional keyword-based parameters passed to the specific target prediction method. :raises ValueError: if the target is invalid :return: shape: (batch_size, num) the scores """ if target == LABEL_TAIL: if full_batch: hrt_batch = hrt_batch[:, 0:2] return self.predict_t(hrt_batch, **kwargs, tails=ids) if target == LABEL_RELATION: if full_batch: hrt_batch = hrt_batch[:, 0::2] return self.predict_r(hrt_batch, **kwargs, relations=ids) if target == LABEL_HEAD: if full_batch: hrt_batch = hrt_batch[:, 1:3] return self.predict_h(hrt_batch, **kwargs, heads=ids) raise ValueError(f"Unknown target={target}")
"""Inverse scoring""" def _prepare_inverse_batch(self, batch: torch.LongTensor, index_relation: int) -> torch.LongTensor: if not self.use_inverse_triples: raise ValueError( "Your model is not configured to predict with inverse relations." " Set ``create_inverse_triples=True`` when creating the dataset/triples factory" " or using the pipeline().", ) return relation_inverter.invert_(batch=batch, index=index_relation).flip(1)
[docs] def score_hrt_inverse( self, hrt_batch: torch.LongTensor, *, mode: Optional[InductiveMode] = None, ) -> torch.FloatTensor: r""" Score triples based on inverse triples, i.e., compute $f(h,r,t)$ based on $f(t,r_{inv},h)$. When training with inverse relations, the model produces two (different) scores for a triple $(h,r,t) \in K$. The forward score is calculated from $f(h,r,t)$ and the inverse score is calculated from $f(t,r_{inv},h)$. This function enables users to inspect the scores obtained by using the corresponding inverse triples. :param hrt_batch: shape: (b, 3) the batch of triples :param mode: the inductive mode, or None for transductive :return: the triple scores obtained by inverse relations """ t_r_inv_h = self._prepare_inverse_batch(batch=hrt_batch, index_relation=1) return self.score_hrt(hrt_batch=t_r_inv_h, mode=mode)
[docs] def score_t_inverse(self, hr_batch: torch.LongTensor, *, tails: Optional[torch.LongTensor] = None, **kwargs): """Score all tails for a batch of (h,r)-pairs using the head predictions for the inverses $(*,r_{inv},h)$.""" r_inv_h = self._prepare_inverse_batch(batch=hr_batch, index_relation=1) return self.score_h(rt_batch=r_inv_h, heads=tails, **kwargs)
[docs] def score_h_inverse(self, rt_batch: torch.LongTensor, *, heads: Optional[torch.LongTensor] = None, **kwargs): """Score all heads for a batch of (r,t)-pairs using the tail predictions for the inverses $(t,r_{inv},*)$.""" t_r_inv = self._prepare_inverse_batch(batch=rt_batch, index_relation=0) return self.score_t(hr_batch=t_r_inv, tails=heads, **kwargs)