def _create_model(self, model_func, client, workers, n_estimators,
base_seed, **kwargs):
self.client = get_client(client)
self.workers = self.client.scheduler_info()['workers'].keys()
self.local_model = None
self.n_estimators_per_worker = \
self._estimators_per_worker(n_estimators)
if base_seed is None:
base_seed = 0
seeds = [base_seed]
for i in range(1, len(self.n_estimators_per_worker)):
sd = self.n_estimators_per_worker[i - 1] + seeds[i - 1]
seeds.append(sd)
self.rfs = {
worker: self.client.submit(
model_func,
n_estimators=self.n_estimators_per_worker[n],
seed=seeds[n],
**kwargs,
pure=False,
workers=[worker],
)
for n, worker in enumerate(self.workers)
}
wait_and_raise_from_futures(list(self.rfs.values()))
def _create_model(self, model_func, client, workers, n_estimators,
base_seed, ignore_empty_partitions, **kwargs):
self.client = get_client(client)
if workers is None:
# Default to all workers
workers = self.client.scheduler_info()['workers'].keys()
self.workers = workers
self._set_internal_model(None)
self.active_workers = list()
self.ignore_empty_partitions = ignore_empty_partitions
self.n_estimators = n_estimators
self.n_estimators_per_worker = \
self._estimators_per_worker(n_estimators)
if base_seed is None:
base_seed = 0
seeds = [base_seed]
for i in range(1, len(self.n_estimators_per_worker)):
sd = self.n_estimators_per_worker[i - 1] + seeds[i - 1]
seeds.append(sd)
self.rfs = {
worker: self.client.submit(
model_func,
n_estimators=self.n_estimators_per_worker[n],
random_state=seeds[n],
**kwargs,
pure=False,
workers=[worker],
)
for n, worker in enumerate(self.workers)
}
wait_and_raise_from_futures(list(self.rfs.values()))
def __init__(self, *, client=None, verbose=False, **kwargs):
"""
Constructor for distributed estimators.
"""
self.client = get_client(client)
self.verbose = verbose
self.kwargs = kwargs
self.internal_model = None
def reduce(futures, func, client=None):
"""
Performs a cluster-wide reduction by first
running function on worker->host->cluster. This
function takes locality into account by first
reducing partitions local to each worker before
reducing partitions on each host and, finally,
reducing the partitions across the cluster into
a single reduced partition.
Parameters
----------
futures : array-like of dask.Future futures to reduce
func : Python reduction function accepting list
of objects to reduce and returning a single
reduced object.
client : dask.distributed.Client to use for scheduling
Returns
-------
output : dask.Future a future containing the final reduce
object.
"""
client = get_client(client)
# Make sure input futures have been assigned to worker(s)
wait(futures)
for local_reduction_func in [workers_to_parts, hosts_to_parts]:
who_has = client.who_has(futures)
workers = [(first(who_has[m.key]), m) for m in futures]
worker_parts = local_reduction_func(workers)
# Short circuit when all parts already have preferred
# locality
if len(worker_parts) > 1:
# Local tree reduction for scalability
futures = client.compute(
[tree_reduce(p, func) for w, p in worker_parts.items()])
wait(futures)
# Merge across workers
ret = client.compute(tree_reduce(futures, func))
wait(ret)
return ret
def __init__(self, gpu_futures=None, workers=None,
datatype=None, multiple=False, client=None):
self.client = get_client(client)
self.gpu_futures = gpu_futures
self.worker_to_parts = _workers_to_parts(gpu_futures)
self.workers = workers
self.datatype = datatype
self.multiple = multiple
self.worker_info = None
self.total_rows = None
self.ranks = None
self.parts_to_sizes = None
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 confusion_matrix(y_true,
y_pred,
labels=None,
normalize=None,
sample_weight=None,
client=None):
"""Compute confusion matrix to evaluate the accuracy of a classification.
Parameters
----------
y_true : dask.Array (device or host) shape = (n_samples,)
or (n_samples, n_outputs)
Ground truth (correct) target values.
y_pred : dask.Array (device or host) shape = (n_samples,)
or (n_samples, n_outputs)
Estimated target values.
labels : array-like (device or host) shape = (n_classes,), optional
List of labels to index the matrix. This may be used to reorder or
select a subset of labels. If None is given, those that appear at least
once in y_true or y_pred are used in sorted order.
sample_weight : dask.Array (device or host) shape = (n_samples,), optional
Sample weights.
normalize : string in [‘true’, ‘pred’, ‘all’]
Normalizes confusion matrix over the true (rows), predicted (columns)
conditions or all the population. If None, confusion matrix will not be
normalized.
client : dask.distributed.Client, optional
Dask client to use. Will use the default client if None.
Returns
-------
C : array-like (device or host) shape = (n_classes, n_classes)
Confusion matrix.
"""
client = get_client(client)
if labels is None:
labels = sorted_unique_labels(y_true, y_pred)
if normalize not in ['true', 'pred', 'all', None]:
msg = "normalize must be one of " \
f"{{'true', 'pred', 'all', None}}, got {normalize}."
raise ValueError(msg)
use_sample_weight = bool(sample_weight is not None)
dask_arrays = [y_true, y_pred, sample_weight] if use_sample_weight else \
[y_true, y_pred]
# run cm computation on each partition.
data = DistributedDataHandler.create(dask_arrays, client=client)
cms = [
client.submit(_local_cm, p, labels, use_sample_weight,
workers=[w]).result() for w, p in data.gpu_futures
]
# reduce each partition's result into one cupy matrix
cm = sum(cms)
with np.errstate(all='ignore'):
if normalize == 'true':
cm = cm / cm.sum(axis=1, keepdims=True)
elif normalize == 'pred':
cm = cm / cm.sum(axis=0, keepdims=True)
elif normalize == 'all':
cm = cm / cm.sum()
cm = np.nan_to_num(cm)
return cm
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',
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
def make_regression(n_samples=100,
n_features=100,
n_informative=10,
n_targets=1,
bias=0.0,
effective_rank=None,
tail_strength=0.5,
noise=0.0,
shuffle=False,
coef=False,
random_state=None,
n_parts=1,
n_samples_per_part=None,
order='F',
dtype='float32',
client=None,
use_full_low_rank=True):
"""
Generate a random regression problem.
The input set can either be well conditioned (by default) or have a low
rank-fat tail singular profile.
The output is generated by applying a (potentially biased) random linear
regression model with "n_informative" nonzero regressors to the previously
generated input and some gaussian centered noise with some adjustable
scale.
Parameters
----------
n_samples : int, optional (default=100)
The number of samples.
n_features : int, optional (default=100)
The number of features.
n_informative : int, optional (default=10)
The number of informative features, i.e., the number of features used
to build the linear model used to generate the output.
n_targets : int, optional (default=1)
The number of regression targets, i.e., the dimension of the y output
vector associated with a sample. By default, the output is a scalar.
bias : float, optional (default=0.0)
The bias term in the underlying linear model.
effective_rank : int or None, optional (default=None)
if not None:
The approximate number of singular vectors required to explain most
of the input data by linear combinations. Using this kind of
singular spectrum in the input allows the generator to reproduce
the correlations often observed in practice.
if None:
The input set is well conditioned, centered and gaussian with
unit variance.
tail_strength : float between 0.0 and 1.0, optional (default=0.5)
The relative importance of the fat noisy tail of the singular values
profile if "effective_rank" is not None.
noise : float, optional (default=0.0)
The standard deviation of the gaussian noise applied to the output.
shuffle : boolean, optional (default=False)
Shuffle the samples and the features.
coef : boolean, optional (default=False)
If True, the coefficients of the underlying linear model are returned.
random_state : int, CuPy RandomState instance, Dask RandomState instance \
or None (default)
Determines random number generation for dataset creation. Pass an int
for reproducible output across multiple function calls.
n_parts : int, optional (default=1)
The number of parts of work.
order : str, optional (default='F')
Row-major or Col-major
dtype: str, optional (default='float32')
dtype of generated data
use_full_low_rank : boolean (default=True)
Whether to use the entire dataset to generate the low rank matrix.
If False, it creates a low rank covariance and uses the
corresponding covariance to generate a multivariate normal
distribution on the remaining chunks
Returns
-------
X : Dask-CuPy array of shape [n_samples, n_features]
The input samples.
y : Dask-CuPy array of shape [n_samples] or [n_samples, n_targets]
The output values.
coef : Dask-CuPy array of shape [n_features] \
or [n_features, n_targets], optional
The coefficient of the underlying linear model. It is returned only if
coef is True.
Notes
-----
Known Performance Limitations:
1. When `effective_rank` is set and `use_full_low_rank` is True, \
we cannot generate order `F` by construction, and an explicit \
transpose is performed on each part. This may cause memory to spike \
(other parameters make order `F` by construction)
2. When `n_targets > 1` and `order = 'F'` as above, we have to \
explicity transpose the `y` array. If `coef = True`, then we also \
explicity transpose the `ground_truth` array
3. When `shuffle = True` and `order = F`, there are memory spikes to \
shuffle the `F` order arrays
.. note:: If out-of-memory errors are encountered in any of the above
configurations, try increasing the `n_parts` parameter.
"""
client = get_client(client=client)
n_informative = min(n_features, n_informative)
rs = _create_rs_generator(random_state)
if n_samples_per_part is None:
n_samples_per_part = max(1, int(n_samples / n_parts))
data_chunksizes = [n_samples_per_part] * n_parts
data_chunksizes[-1] += (n_samples % n_parts)
data_chunksizes = tuple(data_chunksizes)
if effective_rank is None:
# Randomly generate a well conditioned input set
if order == 'F':
X = _f_order_standard_normal(client, rs, data_chunksizes,
n_features, dtype)
elif order == 'C':
X = rs.standard_normal((n_samples, n_features),
chunks=(data_chunksizes, -1),
dtype=dtype)
else:
# Randomly generate a low rank, fat tail input set
if use_full_low_rank:
X = make_low_rank_matrix(n_samples=n_samples,
n_features=n_features,
effective_rank=effective_rank,
tail_strength=tail_strength,
random_state=rs,
n_parts=n_parts,
n_samples_per_part=n_samples_per_part,
dtype=dtype)
X = X.rechunk({0: data_chunksizes, 1: -1})
else:
seed = int(rs.randint(n_samples).compute())
covar = _make_low_rank_covariance(client, n_features,
effective_rank, tail_strength,
seed, n_parts,
n_samples_per_part, dtype)
X = _data_from_multivariate_normal(client, rs, covar,
data_chunksizes, n_features,
dtype)
X = _convert_to_order(client, X, data_chunksizes, order, n_features,
dtype)
# Generate a ground truth model with only n_informative features being non
# zeros (the other features are not correlated to y and should be ignored
# by a sparsifying regularizers such as L1 or elastic net)
ground_truth = 100.0 * rs.standard_normal((n_informative, n_targets),
chunks=(n_samples_per_part, -1),
dtype=dtype)
y = da.dot(X[:, :n_informative], ground_truth) + bias
if n_informative != n_features:
zeroes = 0.0 * rs.standard_normal(
(n_features - n_informative, n_targets), dtype=dtype)
ground_truth = da.concatenate([ground_truth, zeroes], axis=0)
ground_truth = ground_truth.rechunk(-1)
# Add noise
if noise > 0.0:
y += rs.normal(scale=noise, size=y.shape, dtype=dtype)
# Randomly permute samples and features
if shuffle:
features_indices = np.random.permutation(n_features)
X, y = _shuffle(client, rs, X, y, data_chunksizes, n_features,
features_indices, n_targets, dtype)
ground_truth = ground_truth[features_indices, :]
y = da.squeeze(y)
if order == 'F' and n_targets > 1:
y = _convert_to_order(client, y, y.chunks[0], order, n_targets, dtype)
if coef:
ground_truth = _convert_to_order(client, ground_truth,
ground_truth.chunks[0], order,
n_targets, dtype)
if coef:
ground_truth = da.squeeze(ground_truth)
return X, y, ground_truth
else:
return X, y
