package carls.candidate_sampling

Get desktop application:
View/edit binary Protocol Buffers messages

A brute force implementation of top-k sampling, which traverses all existing embeddings in a KnowledgeBank and compares with given input.

• SimilarityType similarity_type = 1

  The type of distance used for top-k.

The base config for a candidate sampler.

Used in: DynamicEmbeddingConfig

• optional google.protobuf.Any extension = 1000

  Implementation is encoded in extension.

A sampler used in the sampled softmax/logistic layer. See https://www.tensorflow.org/extras/candidate_sampling.pdf for more details.

• bool unique = 1

  If the sampled key should be unique or not.

• NegativeSamplerConfig.Sampler sampler = 2

Used in: NegativeSamplerConfig

• UNKNOWN = 0

• UNIFORM = 1

  Sample uniformly among both positive and negative data.

• LOG_UNIFORM = 2

  Sample log uniformly among both positive and negative data.

A single sample for negative sampling. See https://www.tensorflow.org/extras/candidate_sampling.pdf for details.

Used in: SampledResult

• string key = 1

  Key of the sampled result.

• optional EmbeddingVectorProto embedding = 2

  Embedding of the sampled result.

• bool is_positive = 3

  Indicates if this is a positive sample (from the given keys) or negative.

• float expected_count = 4

  The sampling probability (or expected count) of Q(key). Usually used in computing sampled logistic/softmax loss.

Context information for one sample input.

Used in: SampleRequest

• repeated string positive_key = 1

  List of positive samples. Usually used in negative sampling.

• optional EmbeddingVectorProto activation = 2

  Activation used to compute the similarity. Usually used in top-k sampling.

A union of the results returned by various samplers.

Used in: SampleResponse.Samples

• oneof sampled_result

  • NegativeSamplingResult negative_sampling_result = 1

  • TopkSamplingResult topk_sampling_result = 2

Types of similarity between two embeddings.

Used in: BruteForceTopkSamplerConfig

• UNKNOWN = 0

• DOT_PRODUCT = 1

• COSINE = 2

A single sample for top-k sampling.

Used in: SampledResult

• string key = 1

  Key of the sampled result.

• optional EmbeddingVectorProto embedding = 2

  Embedding of the sampled result.

• float similarity = 3

  Used in top k sample to indicate its similarity with the input.