def make_blobs(n_samples=100, n_features=2, centers=None, cluster_std=1.0,
n_parts=None, center_box=(-10, 10), shuffle=True,
random_state=None, return_centers=False,
verbosity=logger.LEVEL_INFO, order='F', dtype='float32',
client=None):
"""
Makes labeled Dask-Cupy arrays containing blobs
for a randomly generated set of centroids.
This function calls `make_blobs` from `cuml.datasets` on each Dask worker
and aggregates them into a single Dask Dataframe.
For more information on Scikit-learn's `make_blobs:
<https://scikit-learn.org/stable/modules/generated/sklearn.datasets.make_blobs.html>`_.
Parameters
----------
n_samples : int
number of rows
n_features : int
number of features
centers : int or array of shape [n_centers, n_features],
optional (default=None) The number of centers to generate, or the fixed
center locations. If n_samples is an int and centers is None, 3 centers
are generated. If n_samples is array-like, centers must be either None
or an array of length equal to the length of n_samples.
cluster_std : float (default = 1.0)
standard deviation of points around centroid
n_parts : int (default = None)
number of partitions to generate (this can be greater
than the number of workers)
center_box : tuple (int, int) (default = (-10, 10))
the bounding box which constrains all the centroids
random_state : int (default = None)
sets random seed (or use None to reinitialize each time)
return_centers : bool, optional (default=False)
If True, then return the centers of each cluster
verbosity : int (default = cuml.logger.LEVEL_INFO)
Logging level.
shuffle : bool (default=False)
Shuffles the samples on each worker.
order: str, optional (default='F')
The order of the generated samples
dtype : str, optional (default='float32')
Dtype of the generated samples
client : dask.distributed.Client (optional)
Dask client to use
Returns
-------
X : dask.array backed by CuPy array of shape [n_samples, n_features]
The input samples.
y : dask.array backed by CuPy array of shape [n_samples]
The output values.
centers : dask.array backed by CuPy array of shape
[n_centers, n_features], optional
The centers of the underlying blobs. It is returned only if
return_centers is True.
"""
client = get_client(client=client)
generator = _create_rs_generator(random_state=random_state)
workers = list(client.scheduler_info()['workers'].keys())
n_parts = n_parts if n_parts is not None else len(workers)
parts_workers = (workers * n_parts)[:n_parts]
centers, n_centers = _get_centers(generator, centers, center_box,
n_samples, n_features,
dtype)
rows_per_part = max(1, int(n_samples / n_parts))
worker_rows = [rows_per_part] * n_parts
if rows_per_part == 1:
worker_rows[-1] += n_samples % n_parts
else:
worker_rows[-1] += n_samples % rows_per_part
worker_rows = tuple(worker_rows)
logger.debug("Generating %d samples across %d partitions on "
"%d workers (total=%d samples)" %
(math.ceil(n_samples / len(workers)),
n_parts, len(workers), n_samples))
seeds = generator.randint(n_samples, size=len(parts_workers))
parts = [client.submit(_create_local_data,
part_rows,
n_features,
centers,
cluster_std,
shuffle,
int(seeds[idx]),
order,
dtype,
pure=False,
workers=[parts_workers[idx]])
for idx, part_rows in enumerate(worker_rows)]
X = [client.submit(_get_X, f, pure=False)
for idx, f in enumerate(parts)]
y = [client.submit(_get_labels, f, pure=False)
for idx, f in enumerate(parts)]
X_del = _create_delayed(X, dtype, worker_rows, n_features)
y_del = _create_delayed(y, dtype, worker_rows)
X_final = da.concatenate(X_del, axis=0)
y_final = da.concatenate(y_del, axis=0)
if return_centers:
return X_final, y_final, centers
else:
return X_final, y_final
def make_blobs(n_samples=100,
n_features=2,
centers=None,
cluster_std=1.0,
center_box=(-10.0, 10.0),
shuffle=True,
random_state=None,
return_centers=False,
order='F',
dtype='float32'):
"""Generate isotropic Gaussian blobs for clustering.
Parameters
----------
n_samples : int or array-like, optional (default=100)
If int, it is the total number of points equally divided among
clusters.
If array-like, each element of the sequence indicates
the number of samples per cluster.
n_features : int, optional (default=2)
The number of features for each sample.
centers : int or array of shape [n_centers, n_features], optional
(default=None)
The number of centers to generate, or the fixed center locations.
If n_samples is an int and centers is None, 3 centers are generated.
If n_samples is array-like, centers must be
either None or an array of length equal to the length of n_samples.
cluster_std : float or sequence of floats, optional (default=1.0)
The standard deviation of the clusters.
center_box : pair of floats (min, max), optional (default=(-10.0, 10.0))
The bounding box for each cluster center when centers are
generated at random.
shuffle : boolean, optional (default=True)
Shuffle the samples.
random_state : int, RandomState instance, default=None
Determines random number generation for dataset creation. Pass an int
for reproducible output across multiple function calls.
return_centers : bool, optional (default=False)
If True, then return the centers of each cluster
order: str, optional (default='F')
The order of the generated samples
dtype : str, optional (default='float32')
Dtype of the generated samples
Returns
-------
X : device array of shape [n_samples, n_features]
The generated samples.
y : device array of shape [n_samples]
The integer labels for cluster membership of each sample.
centers : device array, shape [n_centers, n_features]
The centers of each cluster. Only returned if
``return_centers=True``.
Examples
--------
>>> from sklearn.datasets import make_blobs
>>> X, y = make_blobs(n_samples=10, centers=3, n_features=2,
... random_state=0)
>>> print(X.shape)
(10, 2)
>>> y
array([0, 0, 1, 0, 2, 2, 2, 1, 1, 0])
>>> X, y = make_blobs(n_samples=[3, 3, 4], centers=None, n_features=2,
... random_state=0)
>>> print(X.shape)
(10, 2)
>>> y
array([0, 1, 2, 0, 2, 2, 2, 1, 1, 0])
See also
--------
make_classification: a more intricate variant
"""
generator = _create_rs_generator(random_state=random_state)
centers, n_centers = _get_centers(generator, centers, center_box,
n_samples, n_features, dtype)
# stds: if cluster_std is given as list, it must be consistent
# with the n_centers
if (hasattr(cluster_std, "__len__") and len(cluster_std) != n_centers):
raise ValueError("Length of `clusters_std` not consistent with "
"number of centers. Got centers = {} "
"and cluster_std = {}".format(centers, cluster_std))
if isinstance(cluster_std, numbers.Real):
cluster_std = cp.full(len(centers), cluster_std)
if isinstance(n_samples, Iterable):
n_samples_per_center = n_samples
else:
n_samples_per_center = [int(n_samples // n_centers)] * n_centers
for i in range(n_samples % n_centers):
n_samples_per_center[i] += 1
X = cp.zeros(n_samples * n_features, dtype=dtype)
X = X.reshape((n_samples, n_features), order=order)
y = cp.zeros(n_samples, dtype=dtype)
if shuffle:
proba_samples_per_center = np.array(n_samples_per_center) / np.sum(
n_samples_per_center)
np_seed = int(generator.randint(n_samples, size=1))
np.random.seed(np_seed)
shuffled_sample_indices = cp.array(
np.random.choice(n_centers,
n_samples,
replace=True,
p=proba_samples_per_center))
for i, (n, std) in enumerate(zip(n_samples_per_center, cluster_std)):
center_indices = cp.where(shuffled_sample_indices == i)
y[center_indices[0]] = i
X_k = generator.normal(scale=std,
size=(len(center_indices[0]), n_features),
dtype=dtype)
# NOTE: Adding the loc explicitly as cupy has a bug
# when calling generator.normal with an array for loc.
# cupy.random.normal, however, works with the same
# arguments
cp.add(X_k, centers[i], out=X_k)
X[center_indices[0], :] = X_k
else:
stop = 0
for i, (n, std) in enumerate(zip(n_samples_per_center, cluster_std)):
start, stop = stop, stop + n_samples_per_center[i]
y[start:stop] = i
X_k = generator.normal(scale=std,
size=(n, n_features),
dtype=dtype)
cp.add(X_k, centers[i], out=X_k)
X[start:stop, :] = X_k
if return_centers:
return X, y, centers
else:
return X, y
def make_classification(n_samples=100, n_features=20, n_informative=2,
n_redundant=2, n_repeated=0, n_classes=2,
n_clusters_per_class=2, weights=None, flip_y=0.01,
class_sep=1.0, hypercube=True, shift=0.0, scale=1.0,
shuffle=True, random_state=None, order='F',
dtype='float32', n_parts=None, client=None):
"""
Generate a random n-class classification problem.
This initially creates clusters of points normally distributed (std=1)
about vertices of an `n_informative`-dimensional hypercube with sides of
length ``2 * class_sep`` and assigns an equal number of clusters to each
class. It introduces interdependence between these features and adds
various types of further noise to the data.
Without shuffling, ``X`` horizontally stacks features in the following
order: the primary `n_informative` features, followed by `n_redundant`
linear combinations of the informative features, followed by `n_repeated`
duplicates, drawn randomly with replacement from the informative and
redundant features. The remaining features are filled with random noise.
Thus, without shuffling, all useful features are contained in the columns
``X[:, :n_informative + n_redundant + n_repeated]``.
Examples
--------
.. code-block:: python
from dask.distributed import Client
from dask_cuda import LocalCUDACluster
from cuml.dask.datasets.classification import make_classification
cluster = LocalCUDACluster()
client = Client(cluster)
X, y = make_classification(n_samples=10, n_features=4,
n_informative=2, n_classes=2)
print("X:")
print(X.compute())
print("y:")
print(y.compute())
Output:
.. code-block:: python
X:
[[-1.6990056 -0.8241044 -0.06997631 0.45107925]
[-1.8105277 1.7829906 0.492909 0.05390119]
[-0.18290454 -0.6155432 0.6667889 -1.0053712 ]
[-2.7530136 -0.888528 -0.5023055 1.3983376 ]
[-0.9788184 -0.89851004 0.10802134 -0.10021686]
[-0.76883423 -1.0689086 0.01249526 -0.1404741 ]
[-1.5676656 -0.83082974 -0.03072987 0.34499463]
[-0.9381793 -1.0971068 -0.07465998 0.02618019]
[-1.3021476 -0.87076336 0.02249984 0.15187258]
[ 1.1820307 1.7524253 1.5087451 -2.4626074 ]]
y:
[0 1 0 0 0 0 0 0 0 1]
Parameters
----------
n_samples : int, optional (default=100)
The number of samples.
n_features : int, optional (default=20)
The total number of features. These comprise `n_informative`
informative features, `n_redundant` redundant features,
`n_repeated` duplicated features and
``n_features-n_informative-n_redundant-n_repeated`` useless features
drawn at random.
n_informative : int, optional (default=2)
The number of informative features. Each class is composed of a number
of gaussian clusters each located around the vertices of a hypercube
in a subspace of dimension `n_informative`. For each cluster,
informative features are drawn independently from N(0, 1) and then
randomly linearly combined within each cluster in order to add
covariance. The clusters are then placed on the vertices of the
hypercube.
n_redundant : int, optional (default=2)
The number of redundant features. These features are generated as
random linear combinations of the informative features.
n_repeated : int, optional (default=0)
The number of duplicated features, drawn randomly from the informative
and the redundant features.
n_classes : int, optional (default=2)
The number of classes (or labels) of the classification problem.
n_clusters_per_class : int, optional (default=2)
The number of clusters per class.
weights : array-like of shape ``(n_classes,)`` or ``(n_classes - 1,)``, \
(default=None)
The proportions of samples assigned to each class. If None, then
classes are balanced. Note that if ``len(weights) == n_classes - 1``,
then the last class weight is automatically inferred.
More than `n_samples` samples may be returned if the sum of
`weights` exceeds 1.
flip_y : float, optional (default=0.01)
The fraction of samples whose class is assigned randomly. Larger
values introduce noise in the labels and make the classification
task harder.
class_sep : float, optional (default=1.0)
The factor multiplying the hypercube size. Larger values spread
out the clusters/classes and make the classification task easier.
hypercube : boolean, optional (default=True)
If True, the clusters are put on the vertices of a hypercube. If
False, the clusters are put on the vertices of a random polytope.
shift : float, array of shape [n_features] or None, optional (default=0.0)
Shift features by the specified value. If None, then features
are shifted by a random value drawn in [-class_sep, class_sep].
scale : float, array of shape [n_features] or None, optional (default=1.0)
Multiply features by the specified value. If None, then features
are scaled by a random value drawn in [1, 100]. Note that scaling
happens after shifting.
shuffle : boolean, optional (default=True)
Shuffle the samples and the features.
random_state : int, RandomState instance or None (default)
Determines random number generation for dataset creation. Pass an int
for reproducible output across multiple function calls.
See :term:`Glossary <random_state>`.
order: str, optional (default='F')
The order of the generated samples
dtype : str, optional (default='float32')
Dtype of the generated samples
n_parts : int (default = None)
number of partitions to generate (this can be greater
than the number of workers)
Returns
-------
X : dask.array backed by CuPy array of shape [n_samples, n_features]
The generated samples.
y : dask.array backed by CuPy array of shape [n_samples]
The integer labels for class membership of each sample.
Notes
-----
How we extended the dask MNMG version from the single GPU version:
1. We generate centroids of shape ``(n_centroids, n_informative)``
2. We generate an informative covariance of shape \
``(n_centroids, n_informative, n_informative)``
3. We generate a redundant covariance of shape \
``(n_informative, n_redundant)``
4. We generate the indices for the repeated features \
We pass along the references to the futures of the above arrays \
with each part to the single GPU \
`cuml.datasets.classification.make_classification` so that each \
part (and worker) has access to the correct values to generate \
data from the same covariances
"""
client = get_client(client=client)
rs = _create_rs_generator(random_state)
workers = list(client.scheduler_info()['workers'].keys())
n_parts = n_parts if n_parts is not None else len(workers)
parts_workers = (workers * n_parts)[:n_parts]
n_clusters = n_classes * n_clusters_per_class
# create centroids
centroids = cp.array(_generate_hypercube(n_clusters, n_informative,
rs)).astype(dtype, copy=False)
covariance_seeds = rs.randint(n_features, size=2)
informative_covariance = client.submit(_create_covariance,
(n_clusters, n_informative,
n_informative),
int(covariance_seeds[0]),
pure=False)
redundant_covariance = client.submit(_create_covariance,
(n_informative,
n_redundant),
int(covariance_seeds[1]),
pure=False)
# repeated indices
n = n_informative + n_redundant
repeated_indices = ((n - 1) * rs.rand(n_repeated, dtype=dtype)
+ 0.5).astype(np.intp)
# scale and shift
if shift is None:
shift = (2 * rs.rand(n_features, dtype=dtype) - 1) * class_sep
if scale is None:
scale = 1 + 100 * rs.rand(n_features, dtype=dtype)
# Create arrays on each worker (gpu)
rows_per_part = max(1, int(n_samples / n_parts))
worker_rows = [rows_per_part] * n_parts
worker_rows[-1] += (n_samples % n_parts)
worker_rows = tuple(worker_rows)
part_seeds = rs.permutation(n_parts)
parts = [client.submit(sg_make_classification, worker_rows[i], n_features,
n_informative, n_redundant, n_repeated, n_classes,
n_clusters_per_class, weights, flip_y, class_sep,
hypercube, shift, scale, shuffle,
int(part_seeds[i]), order, dtype, centroids,
informative_covariance, redundant_covariance,
repeated_indices, pure=False,
workers=[parts_workers[i]])
for i in range(len(parts_workers))]
X_parts = [client.submit(_get_X, f, pure=False)
for idx, f in enumerate(parts)]
y_parts = [client.submit(_get_labels, f, pure=False)
for idx, f in enumerate(parts)]
X_dela = _create_delayed(X_parts, dtype, worker_rows, n_features)
y_dela = _create_delayed(y_parts, dtype, worker_rows)
X = da.concatenate(X_dela)
y = da.concatenate(y_dela)
return X, y
def make_classification(n_samples=100,
n_features=20,
n_informative=2,
n_redundant=2,
n_repeated=0,
n_classes=2,
n_clusters_per_class=2,
weights=None,
flip_y=0.01,
class_sep=1.0,
hypercube=True,
shift=0.0,
scale=1.0,
shuffle=True,
random_state=None,
order='F',
dtype='float32',
_centroids=None,
_informative_covariance=None,
_redundant_covariance=None,
_repeated_indices=None):
"""Generate a random n-class classification problem.
This initially creates clusters of points normally distributed (std=1)
about vertices of an ``n_informative``-dimensional hypercube with sides of
length ``2*class_sep`` and assigns an equal number of clusters to each
class. It introduces interdependence between these features and adds
various types of further noise to the data.
Without shuffling, ``X`` horizontally stacks features in the following
order: the primary ``n_informative`` features, followed by ``n_redundant``
linear combinations of the informative features, followed by ``n_repeated``
duplicates, drawn randomly with replacement from the informative and
redundant features. The remaining features are filled with random noise.
Thus, without shuffling, all useful features are contained in the columns
``X[:, :n_informative + n_redundant + n_repeated]``.
Examples
--------
.. code-block:: python
from cuml.datasets.classification import make_classification
X, y = make_classification(n_samples=10, n_features=4,
n_informative=2, n_classes=2)
print("X:")
print(X)
print("y:")
print(y)
Output:
.. code-block:: python
X:
[[-2.3249989 -0.8679415 -1.1511791 1.3525577 ]
[ 2.2933831 1.3743551 0.63128835 -0.84648645]
[ 1.6361488 -1.3233329 0.807027 -0.894092 ]
[-1.0093077 -0.9990691 -0.00808992 0.00950443]
[ 0.99803793 2.068382 0.49570698 -0.8462848 ]
[-1.2750955 -0.9725835 -0.2390058 0.28081596]
[-1.3635055 -0.9637669 -0.31582272 0.37106958]
[ 1.1893625 2.227583 0.48750278 -0.8737561 ]
[-0.05753583 -1.0939395 0.8188342 -0.9620734 ]
[ 0.47910076 0.7648213 -0.17165393 0.26144698]]
y:
[0 1 0 0 1 0 0 1 0 1]
Parameters
----------
n_samples : int, optional (default=100)
The number of samples.
n_features : int, optional (default=20)
The total number of features. These comprise ``n_informative``
informative features, ``n_redundant`` redundant features,
``n_repeated`` duplicated features and
``n_features-n_informative-n_redundant-n_repeated`` useless features
drawn at random.
n_informative : int, optional (default=2)
The number of informative features. Each class is composed of a number
of gaussian clusters each located around the vertices of a hypercube
in a subspace of dimension ``n_informative``. For each cluster,
informative features are drawn independently from N(0, 1) and then
randomly linearly combined within each cluster in order to add
covariance. The clusters are then placed on the vertices of the
hypercube.
n_redundant : int, optional (default=2)
The number of redundant features. These features are generated as
random linear combinations of the informative features.
n_repeated : int, optional (default=0)
The number of duplicated features, drawn randomly from the informative
and the redundant features.
n_classes : int, optional (default=2)
The number of classes (or labels) of the classification problem.
n_clusters_per_class : int, optional (default=2)
The number of clusters per class.
weights : array-like of shape (n_classes,) or (n_classes - 1,),\
(default=None)
The proportions of samples assigned to each class. If None, then
classes are balanced. Note that if ``len(weights) == n_classes - 1``,
then the last class weight is automatically inferred.
More than ``n_samples`` samples may be returned if the sum of
``weights`` exceeds 1.
flip_y : float, optional (default=0.01)
The fraction of samples whose class is assigned randomly. Larger
values introduce noise in the labels and make the classification
task harder.
class_sep : float, optional (default=1.0)
The factor multiplying the hypercube size. Larger values spread
out the clusters/classes and make the classification task easier.
hypercube : boolean, optional (default=True)
If True, the clusters are put on the vertices of a hypercube. If
False, the clusters are put on the vertices of a random polytope.
shift : float, array of shape [n_features] or None, optional (default=0.0)
Shift features by the specified value. If None, then features
are shifted by a random value drawn in [-class_sep, class_sep].
scale : float, array of shape [n_features] or None, optional (default=1.0)
Multiply features by the specified value. If None, then features
are scaled by a random value drawn in [1, 100]. Note that scaling
happens after shifting.
shuffle : boolean, optional (default=True)
Shuffle the samples and the features.
random_state : int, RandomState instance or None (default)
Determines random number generation for dataset creation. Pass an int
for reproducible output across multiple function calls.
See :term:`Glossary <random_state>`.
order: str, optional (default='F')
The order of the generated samples
dtype : str, optional (default='float32')
Dtype of the generated samples
_centroids: array of centroids of shape (n_clusters, n_informative)
_informative_covariance: array for covariance between informative features
of shape (n_clusters, n_informative, n_informative)
_redundant_covariance: array for covariance between redundant features
of shape (n_informative, n_redundant)
_repeated_indices: array of indices for the repeated features
of shape (n_repeated, )
Returns
-------
X : device array of shape [n_samples, n_features]
The generated samples.
y : device array of shape [n_samples]
The integer labels for class membership of each sample.
Notes
-----
The algorithm is adapted from Guyon [1] and was designed to generate
the "Madelon" dataset. How we optimized for GPUs:
1. Firstly, we generate X from a standard univariate instead of zeros.
This saves memory as we don't need to generate univariates each
time for each feature class (informative, repeated, etc.) while
also providing the added speedup of generating a big matrix
on GPU
2. We generate `order=F` construction. We exploit the
fact that X is a generated from a univariate normal, and
covariance is introduced with matrix multiplications. Which means,
we can generate X as a 1D array and just reshape it to the
desired order, which only updates the metadata and eliminates
copies
3. Lastly, we also shuffle by construction. Centroid indices are
permuted for each sample, and then we construct the data for
each centroid. This shuffle works for both `order=C` and
`order=F` and eliminates any need for secondary copies
References
----------
.. [1] I. Guyon, "Design of experiments for the NIPS 2003 variable
selection benchmark", 2003.
"""
generator = _create_rs_generator(random_state)
np_seed = int(generator.randint(n_samples, size=1))
np.random.seed(np_seed)
# Count features, clusters and samples
if n_informative + n_redundant + n_repeated > n_features:
raise ValueError("Number of informative, redundant and repeated "
"features must sum to less than the number of total"
" features")
# Use log2 to avoid overflow errors
if n_informative < np.log2(n_classes * n_clusters_per_class):
msg = "n_classes({}) * n_clusters_per_class({}) must be"
msg += " smaller or equal 2**n_informative({})={}"
raise ValueError(
msg.format(n_classes, n_clusters_per_class, n_informative,
2**n_informative))
if weights is not None:
if len(weights) not in [n_classes, n_classes - 1]:
raise ValueError("Weights specified but incompatible with number "
"of classes.")
if len(weights) == n_classes - 1:
if isinstance(weights, list):
weights = weights + [1.0 - sum(weights)]
else:
weights = np.resize(weights, n_classes)
weights[-1] = 1.0 - sum(weights[:-1])
else:
weights = [1.0 / n_classes] * n_classes
n_clusters = n_classes * n_clusters_per_class
# Distribute samples among clusters by weight
n_samples_per_cluster = [
int(n_samples * weights[k % n_classes] / n_clusters_per_class)
for k in range(n_clusters)
]
for i in range(n_samples - sum(n_samples_per_cluster)):
n_samples_per_cluster[i % n_clusters] += 1
# Initialize X and y
X = generator.randn(n_samples * n_features, dtype=dtype)
X = X.reshape((n_samples, n_features), order=order)
y = cp.zeros(n_samples, dtype=np.int64)
# Build the polytope whose vertices become cluster centroids
if _centroids is None:
centroids = cp.array(
_generate_hypercube(n_clusters, n_informative,
generator)).astype(dtype, copy=False)
else:
centroids = _centroids
centroids *= 2 * class_sep
centroids -= class_sep
if not hypercube:
centroids *= generator.rand(n_clusters, 1, dtype=dtype)
centroids *= generator.rand(1, n_informative, dtype=dtype)
# Create redundant features
if n_redundant > 0:
if _redundant_covariance is None:
B = 2 * generator.rand(n_informative, n_redundant, dtype=dtype) - 1
else:
B = _redundant_covariance
# Create each cluster; a variant of make_blobs
if shuffle:
proba_samples_per_cluster = np.array(n_samples_per_cluster) / np.sum(
n_samples_per_cluster)
shuffled_sample_indices = cp.array(
np.random.choice(n_clusters,
n_samples,
replace=True,
p=proba_samples_per_cluster))
for k, centroid in enumerate(centroids):
centroid_indices = cp.where(shuffled_sample_indices == k)
y[centroid_indices[0]] = k % n_classes
X_k = X[centroid_indices[0], :n_informative]
if _informative_covariance is None:
A = 2 * generator.rand(
n_informative, n_informative, dtype=dtype) - 1
else:
A = _informative_covariance[k]
X_k = cp.dot(X_k, A)
# NOTE: This could be done outside the loop, but a current
# cupy bug does not allow that
# https://github.com/cupy/cupy/issues/3284
if n_redundant > 0:
X[centroid_indices[0],
n_informative:n_informative + n_redundant] = cp.dot(X_k, B)
X_k += centroid # shift the cluster to a vertex
X[centroid_indices[0], :n_informative] = X_k
else:
stop = 0
for k, centroid in enumerate(centroids):
start, stop = stop, stop + n_samples_per_cluster[k]
y[start:stop] = k % n_classes # assign labels
X_k = X[start:stop, :n_informative] # slice a view of the cluster
if _informative_covariance is None:
A = 2 * generator.rand(
n_informative, n_informative, dtype=dtype) - 1
else:
A = _informative_covariance[k]
X_k = cp.dot(X_k, A) # introduce random covariance
if n_redundant > 0:
X[start:stop, n_informative:n_informative + n_redundant] = \
cp.dot(X_k, B)
X_k += centroid # shift the cluster to a vertex
X[start:stop, :n_informative] = X_k
# Repeat some features
if n_repeated > 0:
n = n_informative + n_redundant
if _repeated_indices is None:
indices = ((n - 1) * generator.rand(n_repeated, dtype=dtype) +
0.5).astype(np.intp)
else:
indices = _repeated_indices
X[:, n:n + n_repeated] = X[:, indices]
# Randomly replace labels
if flip_y >= 0.0:
flip_mask = generator.rand(n_samples, dtype=dtype) < flip_y
y[flip_mask] = generator.randint(n_classes, size=int(flip_mask.sum()))
# Randomly shift and scale
if shift is None:
shift = (2 * generator.rand(n_features, dtype=dtype) - 1) * class_sep
X += shift
if scale is None:
scale = 1 + 100 * generator.rand(n_features, dtype=dtype)
X *= scale
return X, y
def make_classification(n_samples=100,
n_features=20,
n_informative=2,
n_redundant=2,
n_repeated=0,
n_classes=2,
n_clusters_per_class=2,
weights=None,
flip_y=0.01,
class_sep=1.0,
hypercube=True,
shift=0.0,
scale=1.0,
shuffle=True,
random_state=None,
order='F',
dtype='float32',
n_parts=None,
client=None):
"""
Generate a random n-class classification problem.
This initially creates clusters of points normally distributed (std=1)
about vertices of an `n_informative`-dimensional hypercube with sides of
length :py:`2 * class_sep` and assigns an equal number of clusters to each
class. It introduces interdependence between these features and adds
various types of further noise to the data.
Without shuffling, `X` horizontally stacks features in the following
order: the primary `n_informative` features, followed by `n_redundant`
linear combinations of the informative features, followed by `n_repeated`
duplicates, drawn randomly with replacement from the informative and
redundant features. The remaining features are filled with random noise.
Thus, without shuffling, all useful features are contained in the columns
:py:`X[:, :n_informative + n_redundant + n_repeated]`.
Examples
--------
.. code-block:: python
>>> from dask.distributed import Client
>>> from dask_cuda import LocalCUDACluster
>>> from cuml.dask.datasets.classification import make_classification
>>> cluster = LocalCUDACluster()
>>> client = Client(cluster)
>>> X, y = make_classification(n_samples=10, n_features=4,
... random_state=1, n_informative=2,
... n_classes=2)
>>> print(X.compute()) # doctest: +SKIP
[[-1.1273878 1.2844919 -0.32349187 0.1595734 ]
[ 0.80521786 -0.65946865 -0.40753683 0.15538901]
[ 1.0404129 -1.481386 1.4241115 1.2664981 ]
[-0.92821544 -0.6805706 -0.26001272 0.36004275]
[-1.0392245 -1.1977317 0.16345565 -0.21848428]
[ 1.2273135 -0.529214 2.4799604 0.44108105]
[-1.9163864 -0.39505136 -1.9588828 -1.8881643 ]
[-0.9788184 -0.89851004 -0.08339313 0.1130247 ]
[-1.0549078 -0.8993015 -0.11921967 0.04821599]
[-1.8388828 -1.4063598 -0.02838472 -1.0874642 ]]
>>> print(y.compute()) # doctest: +SKIP
[1 0 0 0 0 1 0 0 0 0]
>>> client.close()
>>> cluster.close()
Parameters
----------
n_samples : int, optional (default=100)
The number of samples.
n_features : int, optional (default=20)
The total number of features. These comprise `n_informative`
informative features, `n_redundant` redundant features,
`n_repeated` duplicated features and
:py:`n_features-n_informative-n_redundant-n_repeated` useless features
drawn at random.
n_informative : int, optional (default=2)
The number of informative features. Each class is composed of a number
of gaussian clusters each located around the vertices of a hypercube
in a subspace of dimension `n_informative`. For each cluster,
informative features are drawn independently from N(0, 1) and then
randomly linearly combined within each cluster in order to add
covariance. The clusters are then placed on the vertices of the
hypercube.
n_redundant : int, optional (default=2)
The number of redundant features. These features are generated as
random linear combinations of the informative features.
n_repeated : int, optional (default=0)
The number of duplicated features, drawn randomly from the informative
and the redundant features.
n_classes : int, optional (default=2)
The number of classes (or labels) of the classification problem.
n_clusters_per_class : int, optional (default=2)
The number of clusters per class.
weights : array-like of shape :py:`(n_classes,)` or :py:`(n_classes - 1,)`\
, (default=None)
The proportions of samples assigned to each class. If None, then
classes are balanced. Note that if :py:`len(weights) == n_classes - 1`,
then the last class weight is automatically inferred.
More than `n_samples` samples may be returned if the sum of
`weights` exceeds 1.
flip_y : float, optional (default=0.01)
The fraction of samples whose class is assigned randomly. Larger
values introduce noise in the labels and make the classification
task harder.
class_sep : float, optional (default=1.0)
The factor multiplying the hypercube size. Larger values spread
out the clusters/classes and make the classification task easier.
hypercube : boolean, optional (default=True)
If True, the clusters are put on the vertices of a hypercube. If
False, the clusters are put on the vertices of a random polytope.
shift : float, array of shape [n_features] or None, optional (default=0.0)
Shift features by the specified value. If None, then features
are shifted by a random value drawn in [-class_sep, class_sep].
scale : float, array of shape [n_features] or None, optional (default=1.0)
Multiply features by the specified value. If None, then features
are scaled by a random value drawn in [1, 100]. Note that scaling
happens after shifting.
shuffle : boolean, optional (default=True)
Shuffle the samples and the features.
random_state : int, RandomState instance or None (default)
Determines random number generation for dataset creation. Pass an int
for reproducible output across multiple function calls.
See :term:`Glossary <random_state>`.
order: str, optional (default='F')
The order of the generated samples
dtype : str, optional (default='float32')
Dtype of the generated samples
n_parts : int (default = None)
number of partitions to generate (this can be greater
than the number of workers)
Returns
-------
X : dask.array backed by CuPy array of shape [n_samples, n_features]
The generated samples.
y : dask.array backed by CuPy array of shape [n_samples]
The integer labels for class membership of each sample.
Notes
-----
How we extended the dask MNMG version from the single GPU version:
1. We generate centroids of shape ``(n_centroids, n_informative)``
2. We generate an informative covariance of shape \
``(n_centroids, n_informative, n_informative)``
3. We generate a redundant covariance of shape \
``(n_informative, n_redundant)``
4. We generate the indices for the repeated features \
We pass along the references to the futures of the above arrays \
with each part to the single GPU \
`cuml.datasets.classification.make_classification` so that each \
part (and worker) has access to the correct values to generate \
data from the same covariances
"""
client = get_client(client=client)
rs = _create_rs_generator(random_state)
workers = list(client.scheduler_info()['workers'].keys())
n_parts = n_parts if n_parts is not None else len(workers)
parts_workers = (workers * n_parts)[:n_parts]
n_clusters = n_classes * n_clusters_per_class
# create centroids
centroids = cp.array(_generate_hypercube(n_clusters, n_informative,
rs)).astype(dtype, copy=False)
covariance_seeds = rs.randint(n_features, size=2)
informative_covariance = client.submit(
_create_covariance, (n_clusters, n_informative, n_informative),
int(covariance_seeds[0]),
pure=False)
redundant_covariance = client.submit(_create_covariance,
(n_informative, n_redundant),
int(covariance_seeds[1]),
pure=False)
# repeated indices
n = n_informative + n_redundant
repeated_indices = ((n - 1) * rs.rand(n_repeated, dtype=dtype) +
0.5).astype(np.intp)
# scale and shift
if shift is None:
shift = (2 * rs.rand(n_features, dtype=dtype) - 1) * class_sep
if scale is None:
scale = 1 + 100 * rs.rand(n_features, dtype=dtype)
# Create arrays on each worker (gpu)
rows_per_part = max(1, int(n_samples / n_parts))
worker_rows = [rows_per_part] * n_parts
worker_rows[-1] += (n_samples % n_parts)
worker_rows = tuple(worker_rows)
part_seeds = rs.permutation(n_parts)
parts = [
client.submit(sg_make_classification,
worker_rows[i],
n_features,
n_informative,
n_redundant,
n_repeated,
n_classes,
n_clusters_per_class,
weights,
flip_y,
class_sep,
hypercube,
shift,
scale,
shuffle,
int(part_seeds[i]),
order,
dtype,
centroids,
informative_covariance,
redundant_covariance,
repeated_indices,
pure=False,
workers=[parts_workers[i]])
for i in range(len(parts_workers))
]
X_parts = [
client.submit(_get_X, f, pure=False) for idx, f in enumerate(parts)
]
y_parts = [
client.submit(_get_labels, f, pure=False)
for idx, f in enumerate(parts)
]
X_dela = _create_delayed(X_parts, dtype, worker_rows, n_features)
y_dela = _create_delayed(y_parts, np.int64, worker_rows)
X = da.concatenate(X_dela)
y = da.concatenate(y_dela)
return X, y
