TripleREInteraction

class TripleREInteraction(u=1.0, p=1, power_norm=False)[source]

Bases: NormBasedInteraction[FloatTensor, Tuple[FloatTensor, FloatTensor, FloatTensor], FloatTensor]

A stateful module for the TripleRE interaction function from [yu2021].

\[score(h, (r_h, r, r_t), t) = h * (r_h + u) - t * (r_t + u) + r\]

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.

See also

pykeen.nn.functional.triple_re_interaction()

See also

https://github.com/LongYu-360/TripleRE-Add-NodePiece

Note

this interaction is equivalent to LineaREInteraction except the u term

Initialize the module.

Parameters:

  • u (Optional[float]) – the relation factor offset. can be set to None to disable it.

  • p (int) – The norm used with torch.linalg.vector_norm(). Defaults to 1 for TripleRE.

  • power_norm (bool) – 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. Defaults to False for TripleRE.

Attributes Summary

relation_shape

The symbolic shapes for relation representations

Methods Summary

func(r_head, r_mid, r_tail, t[, u, p, ...])

Evaluate the TripleRE interaction function.

Attributes Documentation

relation_shape: Sequence[str] = ('d', 'd', 'd')

The symbolic shapes for relation representations

Methods Documentation

func(r_head, r_mid, r_tail, t, u=None, p=2, power_norm=False)

Evaluate the TripleRE interaction function.

See also

pykeen.nn.modules.TripleREInteraction for the stateful interaction module

Parameters:

  • h (FloatTensor) – shape: (*batch_dims, rank, dim) The head representations.

  • r_head (FloatTensor) – shape: (*batch_dims, rank, dim) The relation-specific head multiplicator representations.

  • r_mid (FloatTensor) – shape: (*batch_dims, rank, dim) The relation representations.

  • r_tail (FloatTensor) – shape: (*batch_dims, rank, dim) The relation-specific tail multiplicator representations.

  • t (FloatTensor) – shape: (*batch_dims, rank, dim) The tail representations.

  • u (Optional[float]) – the relation factor offset. If u is not None or 0, this corresponds to TripleREv2.

  • p (int) – The p for the norm. cf. torch.linalg.vector_norm().

  • power_norm (bool) – Whether to return the powered norm.

Return type:

FloatTensor

Returns:

shape: batch_dims The scores.