"""Calculate inverse relative frequency weighting."""
# Calculate in-degree, i.e. number of incoming edges
inv, cnt = torch.unique(idx, return_counts=True, return_inverse=True)[1:]
return cnt[inv].float().reciprocal()
class InverseInDegreeEdgeWeighting(EdgeWeighting):
"""Normalize messages by inverse in-degree."""
def forward(self, source: torch.LongTensor,
target: torch.LongTensor) -> torch.FloatTensor: # noqa: D102
return _inverse_frequency_weighting(idx=target)
class InverseOutDegreeEdgeWeighting(EdgeWeighting):
"""Normalize messages by inverse out-degree."""
def forward(self, source: torch.LongTensor,
target: torch.LongTensor) -> torch.FloatTensor: # noqa: D102
return _inverse_frequency_weighting(idx=source)
class SymmetricEdgeWeighting(EdgeWeighting):
"""Normalize messages by product of inverse sqrt of in-degree and out-degree."""
def forward(self, source: torch.LongTensor,
target: torch.LongTensor) -> torch.FloatTensor: # noqa: D102
return (_inverse_frequency_weighting(idx=source) *
_inverse_frequency_weighting(idx=target)).sqrt()
edge_weight_resolver = Resolver.from_subclasses(base=EdgeWeighting,
default=SymmetricEdgeWeighting)
self.entity_shape = base.entity_shape
self.relation_shape = base.relation_shape
self.tail_entity_shape = base.tail_entity_shape
# The parameters of the affine transformation: bias
self.bias = nn.Parameter(torch.empty(size=tuple()), requires_grad=trainable_bias)
self.initial_bias = torch.as_tensor(data=[initial_bias], dtype=torch.get_default_dtype())
# scale. We model this as log(scale) to ensure scale > 0, and thus monotonicity
self.log_scale = nn.Parameter(torch.empty(size=tuple()), requires_grad=trainable_scale)
self.initial_log_scale = torch.as_tensor(data=[math.log(initial_scale)], dtype=torch.get_default_dtype())
def reset_parameters(self): # noqa: D102
self.bias.data = self.initial_bias.to(device=self.bias.device)
self.log_scale.data = self.initial_log_scale.to(device=self.bias.device)
def forward(
self,
h: HeadRepresentation,
r: RelationRepresentation,
t: TailRepresentation,
) -> torch.FloatTensor: # noqa: D102
return self.log_scale.exp() * self.base(h=h, r=r, t=t) + self.bias
interaction_resolver = Resolver.from_subclasses(
Interaction, # type: ignore
skip={TranslationalInteraction, FunctionalInteraction, MonotonicAffineTransformationInteraction},
suffix=Interaction.__name__,
)
__all__ = [
"evaluate",
"Evaluator",
"MetricResults",
"RankBasedEvaluator",
"RankBasedMetricResults",
"ClassificationEvaluator",
"ClassificationMetricResults",
"evaluator_resolver",
"metric_resolver",
"get_metric_list",
]
evaluator_resolver = Resolver.from_subclasses(
base=Evaluator, # type: ignore
default=RankBasedEvaluator,
)
_METRICS_SUFFIX = "MetricResults"
_METRICS: Set[Type[MetricResults]] = {
RankBasedMetricResults,
ClassificationMetricResults,
}
metric_resolver = Resolver(
_METRICS,
suffix=_METRICS_SUFFIX,
base=MetricResults,
)
def get_metric_list():
'RESCAL',
'RGCN',
'RotatE',
'SimplE',
'StructuredEmbedding',
'TransD',
'TransE',
'TransH',
'TransR',
'TuckER',
'UnstructuredModel',
# Utils
'model_resolver',
'make_model',
'make_model_cls',
]
model_resolver = Resolver.from_subclasses(
base=Model,
skip={
_NewAbstractModel,
# We might be able to relax this later
ERModel,
LiteralModel,
# Old style models should never be looked up
_OldAbstractModel,
EntityEmbeddingModel,
EntityRelationEmbeddingModel,
},
)
"PomBaseGetter",
"PubChemCompoundGetter",
"RGDGetter",
"ReactomeGetter",
"RheaGetter",
"SCHEMGetter",
"SCOMPGetter",
"SDISGetter",
"SFAMGetter",
"SwissLipidsGetter",
"UMLSGetter",
"UniProtGetter",
"UniProtPtmGetter",
"WikiPathwaysGetter",
"ZFINGetter",
"ontology_resolver",
]
def _assert_sorted():
_sorted = sorted(__all__)
if _sorted != __all__:
raise ValueError(f"unsorted. should be:\n{_sorted}")
_assert_sorted()
del _assert_sorted
ontology_resolver: Resolver[Obo] = Resolver.from_subclasses(base=Obo,
suffix="Getter")
def test_make_many(self):
"""Test the make_many function."""
with self.assertRaises(ValueError):
# no default is given
self.resolver.make_many(None)
with self.assertRaises(ValueError):
# wrong number of kwargs is given
self.resolver.make_many([], [{}, {}])
with self.assertRaises(ValueError):
# wrong number of kwargs is given
self.resolver.make_many(["a", "a", "a"], [{}, {}])
# One class, one kwarg
instances = self.resolver.make_many("a", dict(name="name"))
self.assertEqual([A(name="name")], instances)
instances = self.resolver.make_many("a", [dict(name="name")])
self.assertEqual([A(name="name")], instances)
instances = self.resolver.make_many(["a"], dict(name="name"))
self.assertEqual([A(name="name")], instances)
instances = self.resolver.make_many(["a"], [dict(name="name")])
self.assertEqual([A(name="name")], instances)
# Single class, multiple kwargs
instances = self.resolver.make_many(
"a", [dict(name="name1"), dict(name="name2")])
self.assertEqual([A(name="name1"), A(name="name2")], instances)
instances = self.resolver.make_many(
["a"], [dict(name="name1"), dict(name="name2")])
self.assertEqual([A(name="name1"), A(name="name2")], instances)
# Multiple class, one kwargs
instances = self.resolver.make_many(["a", "b", "c"], dict(name="name"))
self.assertEqual(
[A(name="name"), B(name="name"),
C(name="name")], instances)
instances = self.resolver.make_many(["a", "b", "c"],
[dict(name="name")])
self.assertEqual(
[A(name="name"), B(name="name"),
C(name="name")], instances)
# Multiple class, multiple kwargs
instances = self.resolver.make_many(
["a", "b", "c"],
[dict(name="name1"),
dict(name="name2"),
dict(name="name3")])
self.assertEqual([A(name="name1"),
B(name="name2"),
C(name="name3")], instances)
# One class, No kwargs
instances = self.resolver.make_many("e")
self.assertEqual([E()], instances)
instances = self.resolver.make_many(["e"])
self.assertEqual([E()], instances)
instances = self.resolver.make_many("e", None)
self.assertEqual([E()], instances)
instances = self.resolver.make_many(["e"], None)
self.assertEqual([E()], instances)
instances = self.resolver.make_many(["e"], [None])
self.assertEqual([E()], instances)
# No class
resolver = Resolver.from_subclasses(Base, default=A)
instances = resolver.make_many(None, dict(name="name"))
self.assertEqual([A(name="name")], instances)
# other relations
for r in range(self.num_relations):
source_r, target_r, weights_r = _reduce_relation_specific(
relation=r,
source=source,
target=target,
edge_type=edge_type,
edge_weights=edge_weights,
)
# skip relations without edges
if source_r is None:
continue
# compute message, shape: (num_edges_of_type, num_blocks, block_size)
uniq_source_r, inv_source_r = source_r.unique(return_inverse=True)
w_r = self.blocks[r]
m = torch.einsum('nbi,bij->nbj', x[uniq_source_r], w_r).index_select(dim=0, index=inv_source_r)
# optional message weighting
if weights_r is not None:
m = m * weights_r.unsqueeze(dim=1).unsqueeze(dim=2)
# message aggregation
out.index_add_(dim=0, index=target_r, source=m)
return out.reshape(-1, self.output_dim)
decomposition_resolver = Resolver.from_subclasses(base=Decomposition, default=BasesDecomposition)
.. seealso:: https://github.com/DeepGraphLearning/KnowledgeGraphEmbedding/blob/master/codes/model.py
"""
# -w * log sigma(-(m + n)) - log sigma (m + p)
# p >> -m => m + p >> 0 => sigma(m + p) ~= 1 => log sigma(m + p) ~= 0 => -log sigma(m + p) ~= 0
# p << -m => m + p << 0 => sigma(m + p) ~= 0 => log sigma(m + p) << 0 => -log sigma(m + p) >> 0
neg_loss = functional.logsigmoid(-neg_scores - self.margin)
neg_loss = neg_weights * neg_loss
neg_loss = self._reduction_method(neg_loss)
pos_loss = functional.logsigmoid(self.margin + pos_scores)
pos_loss = self._reduction_method(pos_loss)
loss = -pos_loss - neg_loss
if self._reduction_method is torch.mean:
loss = loss / 2.
return loss
loss_resolver = Resolver.from_subclasses(
Loss,
default=MarginRankingLoss,
skip={
PairwiseLoss,
PointwiseLoss,
SetwiseLoss,
},
)
for _name, _cls in loss_resolver.lookup_dict.items():
for _synonym in _cls.synonyms or []:
loss_resolver.synonyms[_synonym] = _cls