package carls.memory_store

Get desktop application:
View/edit binary Protocol Buffers messages

A cluster of the data represented by its (mean, variance).

Used in: GaussianMemoryCheckpointMetaData.ClusterData, MemoryLookupResult

• optional EmbeddingVectorProto mean = 1

  Center/Average of the cluster.

• optional EmbeddingVectorProto variance = 2

  Element-wise variance of the cluster.

MetaData for restoring the state of a GaussianMemory.

• repeated GaussianMemoryCheckpointMetaData.ClusterData cluster_data = 1

  ClusterData for different clusters.

ClusterData corresponds to in-memory

Used in: GaussianMemoryCheckpointMetaData

• optional GaussianCluster gaussian_cluster = 1

  Mean and variance of the Gaussian cluster.

• repeated EmbeddingVectorProto activation_instance = 2

  Sampled activation instances stored in the memory.

Config for a Gaussian Memory.

• int32 per_cluster_buffer_size = 1

  Maximal number of data points for each cluster to keep in its buffer. Required: must > 0. Usually set this number to be at least mini-batch size.

• float distance_to_cluster_threshold = 2

  A positive value for deciding if a new cluster should be created during training. For example, if the distance is measured by CWISE_MEAN_GAUSSIAN, a point that is smaller than the given threshold should fork a new cluster; If the distance is measured by SQUARED_L2 distance to the cluster center, a point that is larger than the given threshold should fork a new cluster, etc.

• int32 max_num_clusters = 3

  Optional, maximal allowed number of clusters. If specified, must > 0.

• int32 bootstrap_steps = 4

  Number of update steps that must be ran before memory growth can happen.

• float min_variance = 5

  The minimal variance as a prior for each dimension. Required: must > 0.

• MemoryDistanceConfig.DistanceType distance_type = 6

  Distance type for closest gaussian cluster lookup.

Configuration for the distance measurement in memory items.

(message has no fields)

Distance types used for computing the distance between an input point and a cluster.

Used in: GaussianMemoryConfig

• DEFAULT_UNKNOWN = 0

• CWISE_MEAN_GAUSSIAN = 1

  Coefficient-wise or element-wise mean Gaussian distance, defined as exp(-dist^2/2) where dist = \sum_i (x_i - mean_i)/sqrt(var_i) / n, with i indexing each dimension of input x and n is the dimension number.

• SQUARED_L2 = 2

  The squared L2 distance from the given point to the center of a cluster. Note that this distance ignores data variance inside each cluster.

Representation of the lookup result for a single input.

Used in: MemoryLookupResponse

• optional GaussianCluster gaussian_cluster = 1

  The closest gaussian cluster. Expect this to morph into a oneof field once new memory types are added.

• float distance_to_cluster = 2

  Distance to the cluster center.

• int32 cluster_index = 3

  Index of the cluster in the memory.

Configuration for a MemoryStore.

Used in: DynamicEmbeddingConfig

• optional google.protobuf.Any extension = 1000

  Implementation is encoded in extension.