Libraries/Microsoft.ML.KMeansClustering.xml
|
<?xml version="1.0"?>
<doc> <assembly> <name>Microsoft.ML.KMeansClustering</name> </assembly> <members> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansPlusPlusInit.Initialize(Microsoft.ML.Runtime.IHost,System.Int32,Microsoft.ML.Runtime.IChannel,Microsoft.ML.Runtime.Training.FeatureFloatVectorCursor.Factory,System.Int32,System.Int32,Microsoft.ML.Runtime.Data.VBuffer{System.Single}[],System.Int64@,System.Int64@,System.Boolean)"> <summary> Initialize starting centroids via KMeans++ algorithm. This algorithm will always run single-threaded, regardless of the value of <paramref name="numThreads" />. </summary> </member> <member name="T:Microsoft.ML.Runtime.KMeans.KMeansAcceleratedRowMap"> <summary> An instance of this class is used by SharedStates in YinYangTrainer and KMeansBarBarInitialization. It effectively bounds MaxInstancesToAccelerate and initializes RowIndexGetter. </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansAcceleratedRowMap.BuildParallelIndexLookup(Microsoft.ML.Runtime.Training.FeatureFloatVectorCursor.Factory)"> <summary> Initializes the parallel index lookup HashArray using a sequential RowCursor. We preinitialize the HashArray so we can perform lock-free lookup operations during the primary KMeans pass. </summary> </member> <member name="T:Microsoft.ML.Runtime.KMeans.KMeansBarBarInitialization.SharedState"> <summary> Data for optimizing KMeans|| initialization. Very similar to SharedState class For every instance, there is a space for the best weight and best cluster computed. In this class, new clusters mean the clusters that were added to the cluster set in the previous round of KMeans|| and old clusters are the rest of them (the ones that were added in the rounds before the previous one). In every round of KMeans||, numSamplesPerRound new clusters are added to the set of clusters. There are 'numRounds' number of rounds. We compute and store the distance of each new cluster from every round to all of the previous clusters and use it to avoid unnecessary computation by applying the triangle inequality. </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansBarBarInitialization.SharedState.SetInstanceCluster(System.Int32,System.Single,System.Int32)"> <summary> When assigning an accelerated row to a cluster, we store away the weight to its closest cluster, as well as the identity of the new closest cluster. Note that bestWeight can be negative since it is corresponding to the weight of a distance which does not have the L2 norm of the point itself. </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansBarBarInitialization.SharedState.SetClusterDistance(System.Int32,Microsoft.ML.Runtime.Data.VBuffer{System.Single}@,System.Single,System.Int32,Microsoft.ML.Runtime.Data.VBuffer{System.Single}@,System.Single)"> <summary> Computes and stores the distance of a new cluster to an old cluster <paramref name="newClusterFeatures"/> must be between 0..numSamplesPerRound-1. </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansBarBarInitialization.SharedState.CanWeightComputationBeAvoided(System.Single,System.Int32,System.Int32)"> <summary> This function is the key to use triangle inequality. Given an instance x distance to the best old cluster, cOld, and distance of a new cluster, cNew, to cOld, this function evaluates whether the distance computation of dist(x,cNew) can be avoided. </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansBarBarInitialization.FindBestCluster(Microsoft.ML.Runtime.Data.VBuffer{System.Single}@,System.Int32,Microsoft.ML.Runtime.KMeans.KMeansBarBarInitialization.SharedState,System.Int32,System.Int32,Microsoft.ML.Runtime.Data.VBuffer{System.Single}[],System.Single[],System.Boolean,System.Boolean,System.Single@,System.Int32@)"> <summary> This function finds the best cluster and the best weight for an instance using smart triangle inequality to avoid unnecessary weight computations. Note that <paramref name="needToStoreWeight"/> is used to avoid the storing the new cluster in final round. After the final round, best cluster information will be ignored. </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansBarBarInitialization.ComputeAccelerationMemoryRequirement(System.Int64,System.Int32,System.Int32,System.Boolean,System.Int64@,System.Int64@)"> <summary> This method computes the memory requirement for _clusterDistances in SharedState (clusterBytes) and the maximum number of instances whose weight to the closest cluster can be memorized in order to avoid recomputation later. </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansBarBarInitialization.Initialize(Microsoft.ML.Runtime.IHost,System.Int32,Microsoft.ML.Runtime.IChannel,Microsoft.ML.Runtime.Training.FeatureFloatVectorCursor.Factory,System.Int32,System.Int32,Microsoft.ML.Runtime.Data.VBuffer{System.Single}[],System.Int64,System.Int64@,System.Int64@)"> <summary> KMeans|| Implementation, see http://theory.stanford.edu/~sergei/papers/vldb12-kmpar.pdf This algorithm will require: - (k * overSampleFactor * rounds * diminsionality * 4) bytes for the final sampled clusters. - (k * overSampleFactor * numThreads * diminsionality * 4) bytes for the per-round sampling. Uses memory in initializationState to cache distances and avoids unnecessary distance computations akin to YinYang-KMeans paper. Everywhere in this function, weight of an instance x from a cluster c means weight(x,c) = dist(x,c)^2-norm(x)^2. We store weight in most cases to avoid unnecessary computation of norm(x). </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansRandomInit.Initialize(Microsoft.ML.Runtime.IHost,System.Int32,Microsoft.ML.Runtime.IChannel,Microsoft.ML.Runtime.Training.FeatureFloatVectorCursor.Factory,System.Int32,Microsoft.ML.Runtime.Data.VBuffer{System.Single}[],System.Int64@,System.Int64@)"> <summary> Initialize starting centroids via reservoir sampling. </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansLloydsYinYangTrain.WorkChunkStateBase.Reduce(Microsoft.ML.Runtime.KMeans.KMeansLloydsYinYangTrain.WorkChunkState[],Microsoft.ML.Runtime.KMeans.KMeansLloydsYinYangTrain.ReducedWorkChunkState)"> <summary> Reduces the array of work chunks into this chunk, coalescing the results from multiple worker threads partitioned over a parallel cursor set and clearing their values to prepare them for the next iteration. </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansLloydsYinYangTrain.ReducedWorkChunkState.UpdateClusters(Microsoft.ML.Runtime.Data.VBuffer{System.Single}[],System.Single[],System.Single[],System.Single@)"> <summary> Updates all the passed in variables with the results of the most recent iteration of cluster assignment. It is assumed that centroids will contain the previous results of this call. </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansLloydsYinYangTrain.SharedState.SetYinYangCluster(System.Int32,Microsoft.ML.Runtime.Data.VBuffer{System.Single}@,System.Single,System.Int32,System.Single)"> <summary> When assigning an accelerated row to a cluster, we store away the distance to its closer and second closed cluster, as well as the identity of the new closest cluster. This method returns the last known closest cluster. </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansLloydsYinYangTrain.SharedState.UpdateYinYangBounds(System.Int32)"> <summary> Updates the known YinYang bounds for the given row using the centroid position deltas from the previous iteration. </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansLloydsYinYangTrain.SharedState.IsYinYangGloballyBound(System.Int32)"> <summary> Determines if the triangle distance inequality still applies to the given row, allowing us to avoid per-cluster distance computation. </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansLloydsYinYangTrain.ProcessChunk(Microsoft.ML.Runtime.Training.FeatureFloatVectorCursor,Microsoft.ML.Runtime.KMeans.KMeansLloydsYinYangTrain.SharedState,Microsoft.ML.Runtime.KMeans.KMeansLloydsYinYangTrain.WorkChunkStateBase,System.Int32,Microsoft.ML.Runtime.Data.VBuffer{System.Single}[],System.Single[])"> <summary> Performs the 'update' step of KMeans. This method is passed a WorkChunkState. In the parallel version this chunk will be one of _numThreads chunks and the RowCursor will be part of a RowCursorSet. In the unthreaded version, this chunk will be the final chunk and hold state for the entire data set. </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansUtils.ParallelWeightedReservoirSample(Microsoft.ML.Runtime.IHost,System.Int32,System.Int32,Microsoft.ML.Runtime.Training.FeatureFloatVectorCursor.Factory,Microsoft.ML.Runtime.KMeans.KMeansUtils.WeightFunc,Microsoft.ML.Runtime.KMeans.KMeansUtils.RowIndexGetter,Microsoft.ML.Runtime.Data.VBuffer{System.Single}[]@,Microsoft.ML.Runtime.Internal.Utilities.Heap{Microsoft.ML.Runtime.KMeans.KMeansUtils.WeightedPoint}[]@)"> <summary> Performs a multithreaded version of weighted reservior sampling, returning an array of numSamples, where each sample has been selected from the data set with a probability of numSamples/N * weight/(sum(weight)). Buffer is sized to the number of threads plus one and stores the minheaps needed to perform the per-thread reservior samples. This method assumes that the numSamples is much smaller than the full dataset as it expects to be able to sample numSamples * numThreads. This is based on the 'A-Res' algorithm in 'Weighted Random Sampling', 2005; Efraimidis, Spirakis: http://utopia.duth.gr/~pefraimi/research/data/2007EncOfAlg.pdf </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansUtils.ParallelMapReduce``2(System.Int32,Microsoft.ML.Runtime.IHost,Microsoft.ML.Runtime.Training.FeatureFloatVectorCursor.Factory,Microsoft.ML.Runtime.KMeans.KMeansUtils.RowIndexGetter,Microsoft.ML.Runtime.KMeans.KMeansUtils.InitAction{``0},Microsoft.ML.Runtime.KMeans.KMeansUtils.MapAction{``0},Microsoft.ML.Runtime.KMeans.KMeansUtils.ReduceAction{``0,``1},``0[]@,``1@)"> <summary> Takes a data cursor and perform an in-memory parallel aggregation operation on it. This helper wraps some of the behavior common to parallel operations over a IRowCursor set, including building the set, creating separate IRandom instances, and IRowCursor disposal. </summary> <typeparam name="TPartitionState">The type that each parallel cursor will be expected to aggregate to.</typeparam> <typeparam name="TGlobalState">The type of the final output from combining each per-thread instance of TInterAgg.</typeparam> <param name="numThreads"></param> <param name="baseHost"></param> <param name="factory"></param> <param name="rowIndexGetter"></param> <param name="initChunk">Initializes an instance of TInterAgg, or prepares/clears it if it is already allocated.</param> <param name="mapper">Invoked for every row, should update TInterAgg using row cursor data.</param> <param name="reducer">Invoked after all row cursors have completed, combines the entire array of TInterAgg instances into a final TAgg result.</param> <param name="buffer">A reusable buffer array of TInterAgg.</param> <param name="result">A reusable reference to the final result.</param> <returns></returns> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansUtils.FindBestCluster(Microsoft.ML.Runtime.Data.VBuffer{System.Single}@,Microsoft.ML.Runtime.Data.VBuffer{System.Single}[],System.Single[],System.Int32,System.Boolean,System.Single@,System.Int32@,System.Single@,System.Int32@)"> <summary> Given a point and a set of centroids this method will determine the closest centroid using L2 distance. It will return a value equivalent to that distance, the index of the closest cluster, and a value equivalent to the distance to the second-nearest cluster. </summary> <param name="features"></param> <param name="centroids"></param> <param name="centroidL2s">The L2 norms of the centroids. Used for efficiency and expected to be computed up front.</param> <param name="centroidCount">The number of centroids. Must be less than or equal to the length of the centroid array.</param> <param name="needRealDistance">Whether to return a real L2 distance, or a value missing the L2 norm of <paramref name="features"/>.</param> <param name="minDistance">The distance between <paramref name="features"/> and the nearest centroid in <paramref name="centroids" />.</param> <param name="cluster">The index of the nearest centroid.</param> <param name="secMinDistance">The second nearest distance, or PosInf if <paramref name="centroids" /> only contains a single point.</param> <param name="secCluster">The index of the second nearest centroid, or -1 if <paramref name="centroids" /> only contains a single point.</param> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansUtils.VerifyModelConsistency(Microsoft.ML.Runtime.Data.VBuffer{System.Single}[])"> <summary> Checks that all coordinates of all centroids are finite, and throws otherwise </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansPredictor.GetVersionInfo"> <summary> Version information to be saved in binary format </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansPredictor.#ctor(Microsoft.ML.Runtime.IHostEnvironment,System.Int32,Microsoft.ML.Runtime.Data.VBuffer{System.Single}[],System.Boolean)"> <summary> Initialize predictor with a trained model. </summary> <param name="env">The host environment</param> <param name="k">Number of centroids</param> <param name="centroids">Coordinates of the centroids</param> <param name="copyIn">If true then the <paramref name="centroids"/> vectors will be subject to a deep copy, if false then this constructor will take ownership of the passed in centroid vectors. If false then the caller must take care to not use or modify the input vectors once this object is constructed, and should probably remove all references.</param> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansPredictor.#ctor(Microsoft.ML.Runtime.IHostEnvironment,Microsoft.ML.Runtime.Model.ModelLoadContext)"> <summary> Initialize predictor from a binary file. </summary> <param name="ctx">The load context</param> <param name="env">The host environment</param> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansPredictor.SaveCore(Microsoft.ML.Runtime.Model.ModelSaveContext)"> <summary> Save the predictor in binary format. </summary> <param name="ctx">The context to save to</param> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansPredictor.Create(Microsoft.ML.Runtime.IHostEnvironment,Microsoft.ML.Runtime.Model.ModelLoadContext)"> <summary> This method is called by reflection to instantiate a predictor. </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansPredictor.InitPredictor"> <summary> Initialize internal parameters: L2 norms of the _centroids. </summary> </member> <member name="M:Microsoft.ML.Runtime.KMeans.KMeansPredictor.GetClusterCentroids(Microsoft.ML.Runtime.Data.VBuffer{System.Single}[]@,System.Int32@)"> <summary> Copies the centroids to a set of provided buffers. </summary> <param name="centroids">The buffer to which to copy. Will be extended to an appropriate length, if necessary.</param> <param name="k">The number of clusters, corresponding to the logical size of <paramref name="centroids"/>.</param> </member> </members> </doc> |