Negative Sampling

Uniform negative sampler

class torchkge.sampling.UniformNegativeSampler(kg, kg_val=None, kg_test=None, n_neg=1)[source]

Uniform negative sampler as presented in 2013 paper by Bordes et al.. Either the head or the tail of a triplet is replaced by another entity at random. The choice of head/tail is uniform. This class inherits from the torchkge.sampling.NegativeSampler interface. It then has its attributes as well.

References

Parameters
corrupt_batch(heads, tails, relations=None, n_neg=None)[source]

For each true triplet, produce a corrupted one not different from any other true triplet. If heads and tails are cuda objects , then the returned tensors are on the GPU.

Parameters

  • heads (torch.Tensor, dtype: torch.long, shape: (batch_size)) – Tensor containing the integer key of heads of the relations in the current batch.

  • tails (torch.Tensor, dtype: torch.long, shape: (batch_size)) – Tensor containing the integer key of tails of the relations in the current batch.

  • relations (torch.Tensor, dtype: torch.long, shape: (batch_size)) – Tensor containing the integer key of relations in the current batch. This is optional here and mainly present because of the interface with other NegativeSampler objects.

  • n_neg (int (opt)) – Number of negative sample to create from each fact. It overwrites the value set at the construction of the sampler.

Returns

  • neg_heads (torch.Tensor, dtype: torch.long, shape: (batch_size)) – Tensor containing the integer key of negatively sampled heads of the relations in the current batch.

  • neg_tails (torch.Tensor, dtype: torch.long, shape: (batch_size)) – Tensor containing the integer key of negatively sampled tails of the relations in the current batch.

Bernoulli negative sampler

class torchkge.sampling.BernoulliNegativeSampler(kg, kg_val=None, kg_test=None, n_neg=1)[source]

Bernoulli negative sampler as presented in 2014 paper by Wang et al.. Either the head or the tail of a triplet is replaced by another entity at random. The choice of head/tail is done using probabilities taking into account profiles of the relations. See the paper for more details. This class inherits from the torchkge.sampling.NegativeSampler interface. It then has its attributes as well.

References

Parameters
bern_probs

Bernoulli sampling probabilities. See paper for more details.

Type

torch.Tensor, dtype: torch.float, shape: (kg.n_rel)

corrupt_batch(heads, tails, relations, n_neg=None)[source]

For each true triplet, produce a corrupted one assumed to be different from any other true triplet. If heads and tails are cuda objects, then the returned tensors are on the GPU.

Parameters

  • heads (torch.Tensor, dtype: torch.long, shape: (batch_size)) – Tensor containing the integer key of heads of the relations in the current batch.

  • tails (torch.Tensor, dtype: torch.long, shape: (batch_size)) – Tensor containing the integer key of tails of the relations in the current batch.

  • relations (torch.Tensor, dtype: torch.long, shape: (batch_size)) – Tensor containing the integer key of relations in the current batch.

  • n_neg (int (opt)) – Number of negative sample to create from each fact. It overwrites the value set at the construction of the sampler.

Returns

  • neg_heads (torch.Tensor, dtype: torch.long, shape: (batch_size)) – Tensor containing the integer key of negatively sampled heads of the relations in the current batch.

  • neg_tails (torch.Tensor, dtype: torch.long, shape: (batch_size)) – Tensor containing the integer key of negatively sampled tails of the relations in the current batch.

evaluate_probabilities()[source]

Evaluate the Bernoulli probabilities for negative sampling as in the TransH original paper by Wang et al. (2014).

Positional negative sampler

class torchkge.sampling.PositionalNegativeSampler(kg, kg_val=None, kg_test=None)[source]

Positional negative sampler as presented in 2011 paper by Socher et al.. Either the head or the tail of a triplet is replaced by another entity chosen among entities that have already appeared at the same place in a triplet (involving the same relation). It is not clear in the paper how the choice of head/tail is done. We chose to use Bernoulli sampling as in 2014 paper by Wang et al. as we believe it serves the same purpose as the original paper. This class inherits from the torchkge.sampling.BernouilliNegativeSampler class seen as an interface. It then has its attributes as well.

References

Parameters
possible_heads

keys : relations, values : list of possible heads for each relation.

Type

dict

possible_tails

keys : relations, values : list of possible tails for each relation.

Type

dict

n_poss_heads

List of number of possible heads for each relation.

Type

list

n_poss_tails

List of number of possible tails for each relation.

Type

list

corrupt_batch(heads, tails, relations, n_neg=None)[source]

For each true triplet, produce a corrupted one not different from any other golden triplet. If heads and tails are cuda objects, then the returned tensors are on the GPU.

Parameters

  • heads (torch.Tensor, dtype: torch.long, shape: (batch_size)) – Tensor containing the integer key of heads of the relations in the current batch.

  • tails (torch.Tensor, dtype: torch.long, shape: (batch_size)) – Tensor containing the integer key of tails of the relations in the current batch.

  • relations (torch.Tensor, dtype: torch.long, shape: (batch_size)) – Tensor containing the integer key of relations in the current batch. This is optional here and mainly present because of the interface with other NegativeSampler objects.

Returns

  • neg_heads (torch.Tensor, dtype: torch.long, shape: (batch_size)) – Tensor containing the integer key of negatively sampled heads of the relations in the current batch.

  • neg_tails (torch.Tensor, dtype: torch.long, shape: (batch_size)) – Tensor containing the integer key of negatively sampled tails of the relations in the current batch.

find_possibilities()[source]

For each relation of the knowledge graph (and possibly the validation graph but not the test graph) find all the possible heads and tails in the sense of Wang et al., e.g. all entities that occupy once this position in another triplet.

Returns

  • possible_heads (dict) – keys : relation index, values : list of possible heads

  • possible tails (dict) – keys : relation index, values : list of possible tails

  • n_poss_heads (torch.Tensor, dtype: torch.long, shape: (n_relations)) – Number of possible heads for each relation.

  • n_poss_tails (torch.Tensor, dtype: torch.long, shape: (n_relations)) – Number of possible tails for each relation.