def test_gradient_bh_multithread_match_sequential():
# check that the bh gradient with different num_threads gives the same
# results
n_features = 10
n_samples = 30
n_components = 2
degrees_of_freedom = 1
angle = 3
perplexity = 5
random_state = check_random_state(0)
data = random_state.randn(n_samples, n_features).astype(np.float32)
params = random_state.randn(n_samples, n_components)
n_neighbors = n_samples - 1
distances_csr = NearestNeighbors().fit(data).kneighbors_graph(
n_neighbors=n_neighbors, mode='distance')
P_bh = _joint_probabilities_nn(distances_csr, perplexity, verbose=0)
kl_sequential, grad_sequential = _kl_divergence_bh(
params, P_bh, degrees_of_freedom, n_samples, n_components,
angle=angle, skip_num_points=0, verbose=0, num_threads=1)
for num_threads in [2, 4]:
kl_multithread, grad_multithread = _kl_divergence_bh(
params, P_bh, degrees_of_freedom, n_samples, n_components,
angle=angle, skip_num_points=0, verbose=0, num_threads=num_threads)
assert_allclose(kl_multithread, kl_sequential, rtol=1e-6)
assert_allclose(grad_multithread, grad_multithread)
def test_binary_perplexity_stability():
# Binary perplexity search should be stable.
# The binary_search_perplexity had a bug wherein the P array
# was uninitialized, leading to sporadically failing tests.
n_neighbors = 10
n_samples = 100
random_state = check_random_state(0)
data = random_state.randn(n_samples, 5)
nn = NearestNeighbors().fit(data)
distance_graph = nn.kneighbors_graph(n_neighbors=n_neighbors,
mode='distance')
distances = distance_graph.data.astype(np.float32, copy=False)
distances = distances.reshape(n_samples, n_neighbors)
last_P = None
desired_perplexity = 3
for _ in range(100):
P = _binary_search_perplexity(distances.copy(), desired_perplexity,
verbose=0)
P1 = _joint_probabilities_nn(distance_graph, desired_perplexity,
verbose=0)
# Convert the sparse matrix to a dense one for testing
P1 = P1.toarray()
if last_P is None:
last_P = P
last_P1 = P1
else:
assert_array_almost_equal(P, last_P, decimal=4)
assert_array_almost_equal(P1, last_P1, decimal=4)
def test_barnes_hut_angle():
# When Barnes-Hut's angle=0 this corresponds to the exact method.
angle = 0.0
perplexity = 10
n_samples = 100
for n_components in [2, 3]:
n_features = 5
degrees_of_freedom = float(n_components - 1.0)
random_state = check_random_state(0)
data = random_state.randn(n_samples, n_features)
distances = pairwise_distances(data)
params = random_state.randn(n_samples, n_components)
P = _joint_probabilities(distances, perplexity, verbose=0)
kl_exact, grad_exact = _kl_divergence(params, P, degrees_of_freedom,
n_samples, n_components)
n_neighbors = n_samples - 1
distances_csr = NearestNeighbors().fit(data).kneighbors_graph(
n_neighbors=n_neighbors, mode='distance')
P_bh = _joint_probabilities_nn(distances_csr, perplexity, verbose=0)
kl_bh, grad_bh = _kl_divergence_bh(params, P_bh, degrees_of_freedom,
n_samples, n_components,
angle=angle, skip_num_points=0,
verbose=0)
P = squareform(P)
P_bh = P_bh.toarray()
assert_array_almost_equal(P_bh, P, decimal=5)
assert_almost_equal(kl_exact, kl_bh, decimal=3)
def _fit(self, X, skip_num_points=0):
"""Private function to fit the model using X as training data."""
if self.method not in ['barnes_hut', 'exact']:
raise ValueError("'method' must be 'barnes_hut' or 'exact'")
if self.angle < 0.0 or self.angle > 1.0:
raise ValueError("'angle' must be between 0.0 - 1.0")
if hasattr(self, 'square_distances'):
if self.square_distances not in [True, 'legacy']:
raise ValueError(
"'square_distances' must be True or 'legacy'.")
if self.metric != "euclidean" and self.square_distances is not True:
warnings.warn(
("'square_distances' has been introduced in 0.24"
"to help phase out legacy squaring behavior. The "
"'legacy' setting will be removed in 0.26, and the "
"default setting will be changed to True. In 0.28, "
"'square_distances' will be removed altogether,"
"and distances will be squared by default. Set "
"'square_distances'=True to silence this warning."),
FutureWarning)
if self.method == 'barnes_hut':
if sklearn_check_version('0.23'):
X = self._validate_data(X,
accept_sparse=['csr'],
ensure_min_samples=2,
dtype=[np.float32, np.float64])
else:
X = check_array(X,
accept_sparse=['csr'],
ensure_min_samples=2,
dtype=[np.float32, np.float64])
else:
if sklearn_check_version('0.23'):
X = self._validate_data(X,
accept_sparse=['csr', 'csc', 'coo'],
dtype=[np.float32, np.float64])
else:
X = check_array(X,
accept_sparse=['csr', 'csc', 'coo'],
dtype=[np.float32, np.float64])
if self.metric == "precomputed":
if isinstance(self.init, str) and self.init == 'pca':
raise ValueError("The parameter init=\"pca\" cannot be "
"used with metric=\"precomputed\".")
if X.shape[0] != X.shape[1]:
raise ValueError("X should be a square distance matrix")
check_non_negative(
X, "TSNE.fit(). With metric='precomputed', X "
"should contain positive distances.")
if self.method == "exact" and issparse(X):
raise TypeError(
'TSNE with method="exact" does not accept sparse '
'precomputed distance matrix. Use method="barnes_hut" '
'or provide the dense distance matrix.')
if self.method == 'barnes_hut' and self.n_components > 3:
raise ValueError("'n_components' should be inferior to 4 for the "
"barnes_hut algorithm as it relies on "
"quad-tree or oct-tree.")
random_state = check_random_state(self.random_state)
if self.early_exaggeration < 1.0:
raise ValueError(
"early_exaggeration must be at least 1, but is {}".format(
self.early_exaggeration))
if self.n_iter < 250:
raise ValueError("n_iter should be at least 250")
n_samples = X.shape[0]
neighbors_nn = None
if self.method == "exact":
# Retrieve the distance matrix, either using the precomputed one or
# computing it.
if self.metric == "precomputed":
distances = X
else:
if self.verbose:
print("[t-SNE] Computing pairwise distances...")
if self.metric == "euclidean":
# Euclidean is squared here, rather than using **= 2,
# because euclidean_distances already calculates
# squared distances, and returns np.sqrt(dist) for
# squared=False.
# Also, Euclidean is slower for n_jobs>1, so don't set here
distances = pairwise_distances(X,
metric=self.metric,
squared=True)
else:
distances = pairwise_distances(X,
metric=self.metric,
n_jobs=self.n_jobs)
if np.any(distances < 0):
raise ValueError("All distances should be positive, the "
"metric given is not correct")
if self.metric != "euclidean" and \
getattr(self, 'square_distances', True) is True:
distances **= 2
# compute the joint probability distribution for the input space
P = _joint_probabilities(distances, self.perplexity, self.verbose)
assert np.all(np.isfinite(P)), "All probabilities should be finite"
assert np.all(P >= 0), "All probabilities should be non-negative"
assert np.all(P <= 1), ("All probabilities should be less "
"or then equal to one")
else:
# Compute the number of nearest neighbors to find.
# LvdM uses 3 * perplexity as the number of neighbors.
# In the event that we have very small # of points
# set the neighbors to n - 1.
n_neighbors = min(n_samples - 1, int(3. * self.perplexity + 1))
if self.verbose:
print("[t-SNE] Computing {} nearest neighbors...".format(
n_neighbors))
# Find the nearest neighbors for every point
knn = NearestNeighbors(algorithm='auto',
n_jobs=self.n_jobs,
n_neighbors=n_neighbors,
metric=self.metric)
t0 = time()
knn.fit(X)
duration = time() - t0
if self.verbose:
print("[t-SNE] Indexed {} samples in {:.3f}s...".format(
n_samples, duration))
t0 = time()
distances_nn = knn.kneighbors_graph(mode='distance')
duration = time() - t0
if self.verbose:
print("[t-SNE] Computed neighbors for {} samples "
"in {:.3f}s...".format(n_samples, duration))
# Free the memory used by the ball_tree
del knn
if getattr(self, 'square_distances', True) is True or \
self.metric == "euclidean":
# knn return the euclidean distance but we need it squared
# to be consistent with the 'exact' method. Note that the
# the method was derived using the euclidean method as in the
# input space. Not sure of the implication of using a different
# metric.
distances_nn.data **= 2
# compute the joint probability distribution for the input space
P = _joint_probabilities_nn(distances_nn, self.perplexity,
self.verbose)
if isinstance(self.init, np.ndarray):
X_embedded = self.init
elif self.init == 'pca':
pca = PCA(n_components=self.n_components,
svd_solver='randomized',
random_state=random_state)
X_embedded = pca.fit_transform(X).astype(np.float32, copy=False)
elif self.init == 'random':
# The embedding is initialized with iid samples from Gaussians with
# standard deviation 1e-4.
X_embedded = 1e-4 * random_state.randn(
n_samples, self.n_components).astype(np.float32)
else:
raise ValueError("'init' must be 'pca', 'random', or "
"a numpy array")
# Degrees of freedom of the Student's t-distribution. The suggestion
# degrees_of_freedom = n_components - 1 comes from
# "Learning a Parametric Embedding by Preserving Local Structure"
# Laurens van der Maaten, 2009.
degrees_of_freedom = max(self.n_components - 1, 1)
return self._tsne(P,
degrees_of_freedom,
n_samples,
X_embedded=X_embedded,
neighbors=neighbors_nn,
skip_num_points=skip_num_points)