fast_tsne.py

from __future__ import absolute_import, division, print_function

import warnings
from typing import Optional
from typing_extensions import Literal

import numpy as np
from odin.utils.crypto import md5_checksum
from odin.utils.mpi import MPI, cpu_count
from sklearn.decomposition import PCA

_cached_values = {}

__all__ = ['fast_tsne']


# ===========================================================================
# auto-select best TSNE
# ===========================================================================
def _create_key(framework, kwargs, md5):
  key = dict(kwargs)
  del key['verbose']
  key['md5'] = md5
  return framework + str(list(sorted(key.items(), key=lambda x: x[0])))


def fast_tsne(
    *X,
    n_components: int = 2,
    max_samples: Optional[int] = None,
    perplexity: float = 30.0,
    early_exaggeration: float = 12.0,
    learning_rate: float = 200.0,
    n_iter: int = 1000,
    n_iter_without_progress: int = 300,
    exaggeration_iter: int = 250,
    perplexity_max_iter: int = 100,
    min_grad_norm: float = 1e-7,
    method: str = 'barnes_hut',
    metric: str = "euclidean",
    init: str = "random",
    angle: float = 0.5,
    n_jobs: Optional[int] = 4,
    merge_inputs: bool = True,
    pca_preprocessing: bool = True,
    return_model: bool = False,
    random_state: int = 1,
    verbose: int = 0,
    framework: Literal['auto', 'sklearn', 'cuml'] = 'auto',
):
  """ t-Stochastic Nearest Neighbors.
  If the algorithm take unexpected long time for running, lower the
  `exaggeration_iter`, or reduce the amount of samples by downsampling
  the dataset.

  Parameters
  ----------
  n_components : int, optional (default: 2)
      Dimension of the embedded space.
  max_samples : {int, None}
      if given, downsampling the data to given number of sample
  perplexity : float, optional (default: 30)
      The perplexity is related to the number of nearest neighbors that
      is used in other manifold learning algorithms. Larger datasets
      usually require a larger perplexity. Consider selecting a value
      between 5 and 50. The choice is not extremely critical since t-SNE
      is quite insensitive to this parameter.
  early_exaggeration : float, optional (default: 8.0)
      Controls how tight natural clusters in the original space are in
      the embedded space and how much space will be between them. For
      larger values, the space between natural clusters will be larger
      in the embedded space. Again, the choice of this parameter is not
      very critical. If the cost function increases during initial
      optimization, the early exaggeration factor or the learning rate
      might be too high.
  learning_rate : float, optional (default: 200.0)
      The learning rate for t-SNE is usually in the range [10.0, 1000.0]. If
      the learning rate is too high, the data may look like a 'ball' with any
      point approximately equidistant from its nearest neighbours. If the
      learning rate is too low, most points may look compressed in a dense
      cloud with few outliers. If the cost function gets stuck in a bad local
      minimum increasing the learning rate may help.
  n_iter : int, optional (default: 1000)
      Maximum number of iterations for the optimization. Should be at
      least 250.
  n_iter_without_progress : int, optional (default: 300)
      Maximum number of iterations without progress before we abort the
      optimization, used after 250 initial iterations with early
      exaggeration. Note that progress is only checked every 50 iterations so
      this value is rounded to the next multiple of 50.
  perplexity_max_iter : int, (default 100)
      The number of epochs the best gaussian bands are found for.
  exaggeration_iter : int, (default 250)
      To promote the growth of clusters, set this higher.
  min_grad_norm : float, optional (default: 1e-7)
      If the gradient norm is below this threshold, the optimization will
      be stopped.
  metric : string or callable, optional
      The metric to use when calculating distance between instances in a
      feature array. If metric is a string, it must be one of the options
      allowed by scipy.spatial.distance.pdist for its metric parameter, or
      a metric listed in pairwise.PAIRWISE_DISTANCE_FUNCTIONS.
      If metric is "precomputed", X is assumed to be a distance matrix.
      Alternatively, if metric is a callable function, it is called on each
      pair of instances (rows) and the resulting value recorded. The callable
      should take two arrays from X as input and return a value indicating
      the distance between them. The default is "euclidean" which is
      interpreted as squared euclidean distance.
  init : string or numpy array, optional (default: "random")
      Initialization of embedding. Possible options are 'random', 'pca',
      and a numpy array of shape (n_samples, n_components).
      PCA initialization cannot be used with precomputed distances and is
      usually more globally stable than random initialization.
  verbose : int, optional (default: 0)
      Verbosity level, a number from 0 to 6.
  random_state : int, RandomState instance or None, optional (default: None)
      If int, random_state is the seed used by the random number generator;
      If RandomState instance, random_state is the random number generator;
      If None, the random number generator is the RandomState instance used
      by `np.random`.  Note that different initializations might result in
      different local minima of the cost function.
  method : string (default: 'barnes_hut')
      By default the gradient calculation algorithm uses Barnes-Hut
      approximation running in O(NlogN) time. method='exact'
      will run on the slower, but exact, algorithm in O(N^2) time. The
      exact algorithm should be used when nearest-neighbor errors need
      to be better than 3%. However, the exact method cannot scale to
      millions of examples.
  angle : float (default: 0.5)
      Only used if method='barnes_hut'
      This is the trade-off between speed and accuracy for Barnes-Hut T-SNE.
      'angle' is the angular size (referred to as theta in [3]) of a distant
      node as measured from a point. If this size is below 'angle' then it is
      used as a summary node of all points contained within it.
      This method is not very sensitive to changes in this parameter
      in the range of 0.2 - 0.8. Angle less than 0.2 has quickly increasing
      computation time and angle greater 0.8 has quickly increasing error.
  return_model : a Boolean, if `True`,
      return the trained t-SNE model
  merge_inputs : a Boolean, if `True`,
      merge all arrays into a single array
      for training t-SNE.
  """
  assert len(X) > 0, "No input is given!"
  if isinstance(X[0], (tuple, list)):
    X = X[0]
  if not all(isinstance(x, np.ndarray) for x in X):
    raise ValueError("`X` can only be list of numpy.ndarray or numpy.ndarray")
  # ====== kwarg for creating T-SNE class ====== #
  kwargs = dict(locals())
  del kwargs['X']
  kwargs.pop('merge_inputs')
  kwargs.pop('return_model')
  kwargs.pop('max_samples')
  kwargs.pop('framework')
  kwargs.pop('pca_preprocessing')
  # ====== downsampling ====== #
  if max_samples is not None:
    max_samples = int(max_samples)
    assert max_samples > 0
    new_X = []
    rand = random_state if isinstance(random_state, np.random.RandomState) else \
    np.random.RandomState(seed=random_state)
    for x in X:
      if x.shape[0] > max_samples:
        ids = rand.permutation(x.shape[0])[:max_samples]
        x = x[ids]
      new_X.append(x)
    X = new_X
  # ====== import proper T-SNE ====== #
  tsne_version = None
  if framework != 'sklearn':
    try:
      from cuml.manifold import TSNE
      tsne_version = 'cuda'
    except ImportError:
      warnings.warn("Install RAPIDSAI cuML GPUs-accelerated t-SNE")
      try:
        from MulticoreTSNE import MulticoreTSNE as TSNE
        tsne_version = 'multicore'
      except ImportError:
        warnings.warn("pip install "
                      "git+https://github.com/DmitryUlyanov/Multicore-TSNE.git")
  if tsne_version is None:
    from sklearn.manifold import TSNE
    tsne_version = 'sklearn'
  # ====== modify kwargs ====== #
  if tsne_version == 'cuda':
    del kwargs['n_jobs']
  elif tsne_version == 'multicore':
    del kwargs['perplexity_max_iter']
    del kwargs['exaggeration_iter']
  else:
    del kwargs['n_jobs']
    del kwargs['perplexity_max_iter']
    del kwargs['exaggeration_iter']
  # ====== getting cached values ====== #
  results = []
  X_new = []
  X_size = []
  if merge_inputs:
    X_size = [x.shape[0] for x in X]
    x = np.vstack(X) if len(X) > 1 else X[0]
    md5 = md5_checksum(x)
    key = _create_key(tsne_version, kwargs, md5)
    if key in _cached_values:
      results.append((0, _cached_values[key]))
    else:
      X_new.append((0, md5, x))
  else:
    for i, x in enumerate(X):
      md5 = md5_checksum(x)
      key = _create_key(tsne_version, kwargs, md5)
      if key in _cached_values:
        results.append((i, _cached_values[key]))
      else:
        X_new.append((i, md5, x))

  # ====== perform T-SNE ====== #
  def apply_tsne(j):
    idx, md5, x = j
    if pca_preprocessing:
      x = PCA(n_components=None, random_state=random_state).fit_transform(x)
    tsne = TSNE(**kwargs)
    return (idx, md5, tsne.fit_transform(x), tsne if return_model else None)

  # only 1 X, no need for MPI
  if len(X_new) == 1 or tsne_version in ('cuda', 'multicore'):
    for x in X_new:
      idx, md5, x, model = apply_tsne(x)
      results.append((idx, x))
      _cached_values[_create_key(tsne_version, kwargs, md5)] = x
  else:
    mpi = MPI(jobs=X_new,
              func=apply_tsne,
              batch=1,
              ncpu=min(len(X_new),
                       cpu_count() - 1))
    model = []
    for idx, md5, x, m in mpi:
      results.append((idx, x))
      _cached_values[_create_key(tsne_version, kwargs, md5)] = x
      model.append(m)
  # ====== return and clean ====== #
  if merge_inputs and len(X_size) > 1:
    indices = [0] + np.cumsum(X_size).tolist()
    results = [results[0][1][s:e] for s, e in zip(indices, indices[1:])]
  else:
    results = sorted(results, key=lambda a: a[0])
    results = [r[1] for r in results]
  results = results[0] if len(results) == 1 else results
  if return_model:
    return results, model
  del model
  return results