tsne.h

Go to the documentation of this file.

/*
 * Copyright (c) 2019-2024, NVIDIA CORPORATION.
 *
 * Licensed under the Apache License, Version 2.0 (the "License");
 * you may not use this file except in compliance with the License.
 * You may obtain a copy of the License at
 *
 *     http://www.apache.org/licenses/LICENSE-2.0
 *
 * Unless required by applicable law or agreed to in writing, software
 * distributed under the License is distributed on an "AS IS" BASIS,
 * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 * See the License for the specific language governing permissions and
 * limitations under the License.
 */
 
#pragma once
 
#include <cuml/common/logger.hpp>
 
#include <cuvs/distance/distance.hpp>
 
namespace raft {
class handle_t;
}
 
namespace ML {
 
enum TSNE_ALGORITHM { EXACT, BARNES_HUT, FFT };
 
enum TSNE_INIT { RANDOM, PCA };
 
struct TSNEParams {
  // Number of output dimensions for embeddings Y.
  int dim = 2;
 
  // Number of nearest neighbors used.
  int n_neighbors = 1023;
 
  // Float between 0 and 1. Tradeoff for speed (0) vs accuracy (1).
  // (Barnes-Hut only.)
  float theta = 0.5f;
 
  // A tiny jitter to promote numerical stability. (Barnes-Hut only.)
  float epssq = 0.0025;
 
  // How many nearest neighbors are used during construction of Pij.
  float perplexity = 50.0f;
 
  // Number of iterations used to construct Pij.
  int perplexity_max_iter = 100;
 
  // The small tolerance used for Pij to ensure numerical stability.
  float perplexity_tol = 1e-5;
 
  // How much pressure to apply to clusters to spread out
  // during the exaggeration phase.
  float early_exaggeration = 12.0f;
 
  // How much pressure to apply to clusters to
  // spread out after the exaggeration phase. (FIT-SNE only)
  float late_exaggeration = 1.0f;
 
  // How many iterations you want the early pressure to run for.
  // If late exaggeration is used, it will be applied to all iterations
  // that remain after this number of iterations.
  int exaggeration_iter = 250;
 
  // Rounds up small gradient updates. (Barnes-Hut and Exact only.)
  float min_gain = 0.01f;
 
  // The learning rate during exaggeration phase.
  float pre_learning_rate = 200.0f;
 
  // The learning rate after exaggeration phase.
  float post_learning_rate = 500.0f;
 
  // The maximum number of iterations TSNE should run for.
  int max_iter = 1000;
 
  // The smallest gradient norm TSNE should terminate on.
  // (Exact only; ignored for others.)
  float min_grad_norm = 1e-7;
 
  // The momentum used during the exaggeration phase.
  float pre_momentum = 0.5;
 
  // The momentum used after the exaggeration phase.
  float post_momentum = 0.8;
 
  // Set this to -1 for pure random initializations or >= 0 for
  // reproducible outputs. This sets random seed correctly, but there
  // may still be some variance due to the parallel nature of this algorithm.
  long long random_state = -1;
 
  // verbosity level for logging messages during execution
  int verbosity = CUML_LEVEL_INFO;
 
  // Embedding initializer algorithm
  TSNE_INIT init = TSNE_INIT::RANDOM;
 
  // When this is set to true, the distances from the knn graph will
  // always be squared before computing conditional probabilities, even if
  // the knn graph is passed in explicitly. This is to better match the
  // behavior of Scikit-learn's T-SNE.
  bool square_distances = true;
 
  // Distance metric to use.
  cuvs::distance::DistanceType metric = cuvs::distance::DistanceType::L2SqrtExpanded;
 
  // Value of p for Minkowski distance
  float p = 2.0;
 
  // Which implementation algorithm to use.
  TSNE_ALGORITHM algorithm = TSNE_ALGORITHM::FFT;
};
 
void TSNE_fit(const raft::handle_t& handle,
              float* X,
              float* Y,
              int n,
              int p,
              int64_t* knn_indices,
              float* knn_dists,
              TSNEParams& params,
              float* kl_div = nullptr);
 
void TSNE_fit_sparse(const raft::handle_t& handle,
                     int* indptr,
                     int* indices,
                     float* data,
                     float* Y,
                     int nnz,
                     int n,
                     int p,
                     int* knn_indices,
                     float* knn_dists,
                     TSNEParams& params,
                     float* kl_div = nullptr);
 
}  // namespace ML