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)
distances = random_state.randn(n_samples, n_features)
distances = distances.astype(np.float32)
distances = distances.dot(distances.T)
np.fill_diagonal(distances, 0.0)
params = random_state.randn(n_samples, n_components)
P = _joint_probabilities(distances, perplexity, False)
kl, gradex = _kl_divergence(params, P, degrees_of_freedom, n_samples,
n_components)
k = n_samples - 1
bt = BallTree(distances)
distances_nn, neighbors_nn = bt.query(distances, k=k + 1)
neighbors_nn = neighbors_nn[:, 1:]
Pbh = _joint_probabilities_nn(distances, neighbors_nn,
perplexity, False)
kl, gradbh = _kl_divergence_bh(params, Pbh, neighbors_nn,
degrees_of_freedom, n_samples,
n_components, angle=angle,
skip_num_points=0, verbose=False)
assert_array_almost_equal(Pbh, P, decimal=5)
assert_array_almost_equal(gradex, gradbh, decimal=5)
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.
k = 10
n_samples = 100
random_state = check_random_state(0)
distances = random_state.randn(n_samples, 2).astype(np.float32)
# Distances shouldn't be negative
distances = np.abs(distances.dot(distances.T))
np.fill_diagonal(distances, 0.0)
last_P = None
neighbors_nn = np.argsort(distances, axis=1)[:, :k].astype(np.int64)
for _ in range(100):
P = _binary_search_perplexity(distances.copy(), neighbors_nn.copy(),
3, verbose=0)
P1 = _joint_probabilities_nn(distances, neighbors_nn, 3, 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)
distances = random_state.randn(n_samples, n_features)
distances = distances.astype(np.float32)
distances = distances.dot(distances.T)
np.fill_diagonal(distances, 0.0)
params = random_state.randn(n_samples, n_components)
P = _joint_probabilities(distances, perplexity, False)
kl, gradex = _kl_divergence(params, P, degrees_of_freedom, n_samples,
n_components)
k = n_samples - 1
bt = BallTree(distances)
distances_nn, neighbors_nn = bt.query(distances, k=k + 1)
neighbors_nn = neighbors_nn[:, 1:]
Pbh = _joint_probabilities_nn(distances, neighbors_nn,
perplexity, False)
kl, gradbh = _kl_divergence_bh(params, Pbh, neighbors_nn,
degrees_of_freedom, n_samples,
n_components, angle=angle,
skip_num_points=0, verbose=False)
assert_array_almost_equal(Pbh, P, decimal=5)
assert_array_almost_equal(gradex, gradbh, decimal=5)
def compute_P(X, perplexity):
""" utils function to calculate P matrix in high dim
/opt/anaconda3/lib/python3.6/site-packages/sklearn/manifold/t_sne.py
TODO: try to cache
"""
n_samples = X.shape[0]
# 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.
k = min(n_samples - 1, int(3. * perplexity + 1))
# Find the nearest neighbors for every point
knn = NearestNeighbors(algorithm='auto', n_neighbors=k)
knn.fit(X)
distances_nn, neighbors_nn = knn.kneighbors(None, n_neighbors=k)
# Free the memory used by the ball_tree
del knn
# 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 **= 2
# compute the joint probability distribution for the input space
P = _joint_probabilities_nn(
distances_nn, neighbors_nn, perplexity, verbose=0)
# return P.todense()
return np.maximum(P.todense(), MACHINE_EPSILON)
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 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)
distances = random_state.randn(n_samples, n_features)
distances = distances.astype(np.float32)
distances = abs(distances.dot(distances.T))
np.fill_diagonal(distances, 0.0)
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)
k = n_samples - 1
bt = BallTree(distances)
distances_nn, neighbors_nn = bt.query(distances, k=k + 1)
neighbors_nn = neighbors_nn[:, 1:]
distances_nn = np.array(
[distances[i, neighbors_nn[i]] for i in range(n_samples)])
assert np.all(distances[0, neighbors_nn[0]] == distances_nn[0]),\
abs(distances[0, neighbors_nn[0]] - distances_nn[0])
P_bh = _joint_probabilities_nn(distances_nn,
neighbors_nn,
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 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)
distances = random_state.randn(n_samples, n_features)
distances = distances.astype(np.float32)
distances = abs(distances.dot(distances.T))
np.fill_diagonal(distances, 0.0)
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)
k = n_samples - 1
bt = BallTree(distances)
distances_nn, neighbors_nn = bt.query(distances, k=k + 1)
neighbors_nn = neighbors_nn[:, 1:]
distances_nn = np.array([distances[i, neighbors_nn[i]]
for i in range(n_samples)])
assert np.all(distances[0, neighbors_nn[0]] == distances_nn[0]),\
abs(distances[0, neighbors_nn[0]] - distances_nn[0])
P_bh = _joint_probabilities_nn(distances_nn, neighbors_nn,
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):
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 self.metric == "precomputed":
if isinstance(self.init, string_types) 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")
if np.any(X < 0):
raise ValueError("All distances should be positive, the "
"precomputed distances given as X is not "
"correct")
if self.method == 'barnes_hut' and sp.issparse(X):
raise TypeError('A sparse matrix was passed, but dense '
'data is required for method="barnes_hut". Use '
'X.toarray() to convert to a dense numpy array if '
'the array is small enough for it to fit in '
'memory. Otherwise consider dimensionality '
'reduction techniques (e.g. TruncatedSVD)')
else:
X = check_array(X,
accept_sparse=['csr', 'csc', 'coo'],
dtype=[np.float32, np.float64])
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":
if self.metric == "precomputed":
distances = X
else:
if self.verbose:
print("[t-SNE] Computing pairwise distances...")
if self.metric == "euclidean":
distances = pairwise_distances(
X,
metric=self.metric,
squared=True,
**self.metric_params) # <=ADDED
else:
distances = pairwise_distances(
X, metric=self.metric, **self.metric_params) # <=ADDED
if np.any(distances < 0):
raise ValueError("All distances should be positive, the "
"metric given is not correct")
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:
k = min(n_samples - 1, int(3. * self.perplexity + 1))
if self.verbose:
print("[t-SNE] Computing {} nearest neighbors...".format(k))
knn = NearestNeighbors(algorithm='auto',
n_neighbors=k,
metric=self.metric,
metric_params=self.metric_params) # <=ADDED
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, neighbors_nn = knn.kneighbors(None, n_neighbors=k)
duration = time() - t0
if self.verbose:
print(
"[t-SNE] Computed neighbors for {} samples in {:.3f}s...".
format(n_samples, duration))
del knn
if self.metric == "euclidean":
distances_nn **= 2
P = _joint_probabilities_nn(distances_nn, neighbors_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':
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 = max(self.n_components - 1.0, 1)
return self._tsne(P,
degrees_of_freedom,
n_samples,
X_embedded=X_embedded,
neighbors=neighbors_nn,
skip_num_points=skip_num_points)
def compute_joint_probabilities(X, perplexity=30, metric='euclidean', method='exact', adj=None, verbose=0):
"""
Computes the joint probability matrix P from a feature matrix X of size n x f
Adapted from sklearn.manifold.t_sne
"""
# Compute pairwise distances
if verbose > 0: print('Computing pairwise distances...')
if method == 'exact':
if metric == 'precomputed':
D = X
elif metric == 'euclidean':
D = pairwise_distances(X, metric=metric, squared=True)
elif metric == 'cosine':
D = pairwise_distances(X, metric=metric)
elif metric == 'shortest_path':
assert adj is not None
D = get_shortest_path_matrix(adj, verbose=verbose)
P = _joint_probabilities(D, desired_perplexity=perplexity, verbose=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")
P = squareform(P)
else:
# Cpmpute 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_samples = X.shape[0]
k = min(n_samples - 1, int(3. * perplexity + 1))
# Find the nearest neighbors for every point
knn = NearestNeighbors(algorithm='auto', n_neighbors=k,
metric=metric)
t0 = time()
knn.fit(X)
duration = time() - t0
if verbose:
print("[t-SNE] Indexed {} samples in {:.3f}s...".format(
n_samples, duration))
t0 = time()
distances_nn, neighbors_nn = knn.kneighbors(
None, n_neighbors=k)
duration = time() - t0
if 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 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 **= 2
# compute the joint probability distribution for the input space
P = _joint_probabilities_nn(distances_nn, neighbors_nn,
perplexity, verbose)
P = P.toarray()
# Convert to torch tensor
P = torch.from_numpy(P).type(dtypeFloat)
return P
def _fit(self, X, skip_num_points=0):
"""Fit the model using X as training data.
Note that sparse arrays can only be handled by method='exact'.
It is recommended that you convert your sparse array to dense
(e.g. `X.toarray()`) if it fits in memory, or otherwise using a
dimensionality reduction technique (e.g. TruncatedSVD).
Parameters
----------
X : array, shape (n_samples, n_features) or (n_samples, n_samples)
If the metric is 'precomputed' X must be a square distance
matrix. Otherwise it contains a sample per row. Note that this
when method='barnes_hut', X cannot be a sparse array and if need be
will be converted to a 32 bit float array. Method='exact' allows
sparse arrays and 64bit floating point inputs.
skip_num_points : int (optional, default:0)
This does not compute the gradient for points with indices below
`skip_num_points`. This is useful when computing transforms of new
data where you'd like to keep the old data fixed.
"""
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 self.method == 'barnes_hut' and sp.issparse(X):
raise TypeError('A sparse matrix was passed, but dense '
'data is required for method="barnes_hut". Use '
'X.toarray() to convert to a dense numpy array if '
'the array is small enough for it to fit in '
'memory. Otherwise consider dimensionality '
'reduction techniques (e.g. TruncatedSVD)')
else:
X = check_array(X,
accept_sparse=['csr', 'csc', 'coo'],
dtype=np.float64)
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 "
"%f" % self.early_exaggeration)
if self.n_iter < 200:
raise ValueError("n_iter should be at least 200")
if self.metric == "precomputed":
if isinstance(self.init, string_types) 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")
distances = X
else:
if self.verbose:
print("[t-SNE] Computing pairwise distances...")
if self.metric == "euclidean":
distances = pairwise_distances(X,
metric=self.metric,
squared=True)
else:
distances = pairwise_distances(X, metric=self.metric)
if not np.all(distances >= 0):
raise ValueError("All distances should be positive, either "
"the metric or precomputed distances given "
"as X are not correct")
# 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 - 1.0, 1)
n_samples = X.shape[0]
# the number of nearest neighbors to find
k = min(n_samples - 1, int(3. * self.perplexity + 1))
neighbors_nn = None
if self.method == 'barnes_hut':
if self.verbose:
print("[t-SNE] Computing %i nearest neighbors..." % k)
if self.metric == 'precomputed':
# Use the precomputed distances to find
# the k nearest neighbors and their distances
neighbors_nn = np.argsort(distances, axis=1)[:, :k]
elif self.rho >= 1:
# Find the nearest neighbors for every point
bt = BallTree(X)
# LvdM uses 3 * perplexity as the number of neighbors
# And we add one to not count the data point itself
# In the event that we have very small # of points
# set the neighbors to n - 1
distances_nn, neighbors_nn = bt.query(X, k=k + 1)
neighbors_nn = neighbors_nn[:, 1:]
elif self.rho < 1:
# Use pyFLANN to find the nearest neighbors
myflann = FLANN()
testset = X
params = myflann.build_index(testset,
algorithm="autotuned",
target_precision=self.rho,
log_level='info')
neighbors_nn, distances = myflann.nn_index(
testset, k + 1, checks=params["checks"])
neighbors_nn = neighbors_nn[:, 1:]
P = _joint_probabilities_nn(distances, neighbors_nn,
self.perplexity, self.verbose)
else:
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 zero or positive"
assert np.all(P <= 1), ("All probabilities should be less "
"or then equal to one")
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)
elif self.init == 'random':
X_embedded = None
else:
raise ValueError("Unsupported initialization scheme: %s" %
self.init)
return self._tsne(P,
degrees_of_freedom,
n_samples,
random_state,
X_embedded=X_embedded,
neighbors=neighbors_nn,
skip_num_points=skip_num_points)
