cuML API Reference¶
Module Configuration¶
Output Data Type Configuration¶
memory_utils.
set_global_output_type
(output_type)¶Method to set cuML’s single GPU estimators global output type. It will be used by all estimators unless overriden in their initialization with their own output_type parameter. Can also be overriden by the context manager method using_output_type
 Parameters
 output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’} (default = ‘input’)
Desired output type of results and attributes of the estimators.
‘input’ will mean that the parameters and methods will mirror the format of the data sent to the estimators/methods as much as possible. Specifically:
Input type > Output type
cuDF DataFrame or Series > cuDF DataFrame or Series
NumPy arrays > NumPy arrays
Pandas DataFrame or Series > NumPy arrays
Numba device arrays > Numba device arrays
CuPy arrays > CuPy arrays
Other __cuda_array_interface__ objects > CuPy arrays
‘cudf’ will return cuDF Series for single dimensional results and DataFrames for the rest.
‘cupy’ will return CuPy arrays.
‘numpy’ will return NumPy arrays.
Notes
‘cupy’ and ‘numba’ options (as well as ‘input’ when using Numba and CuPy ndarrays for input) have the least overhead. cuDF add memory consumption and processing time needed to build the Series and DataFrames. ‘numpy’ has the biggest overhead due to the need to transfer data to CPU memory.
Examples
import cuml import cupy as cp ary = [[1.0, 4.0, 4.0], [2.0, 2.0, 2.0], [5.0, 1.0, 1.0]] ary = cp.asarray(ary) cuml.set_global_output_type('cudf'): dbscan_float = cuml.DBSCAN(eps=1.0, min_samples=1) dbscan_float.fit(ary) print("cuML output type") print(dbscan_float.labels_) print(type(dbscan_float.labels_))Output:
cuML output type 0 0 1 1 2 2 dtype: int32 <class 'cudf.core.series.Series'>
memory_utils.
using_output_type
(output_type)¶Context manager method to set cuML’s global output type inside a with statement. It gets reset to the prior value it had once the with code block is executer.
 Parameters
 output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’} (default = ‘input’)
Desired output type of results and attributes of the estimators.
‘input’ will mean that the parameters and methods will mirror the format of the data sent to the estimators/methods as much as possible. Specifically:
Input type > Output type
cuDF DataFrame or Series > cuDF DataFrame or Series
NumPy arrays > NumPy arrays
Pandas DataFrame or Series > NumPy arrays
Numba device arrays > Numba device arrays
CuPy arrays > CuPy arrays
Other __cuda_array_interface__ objects > CuPy arrays
‘cudf’ will return cuDF Series for single dimensional results and DataFrames for the rest.
‘cupy’ will return CuPy arrays.
‘numpy’ will return NumPy arrays.
Examples
import cuml import cupy as cp ary = [[1.0, 4.0, 4.0], [2.0, 2.0, 2.0], [5.0, 1.0, 1.0]] ary = cp.asarray(ary) with cuml.using_output_type('cudf'): dbscan_float = cuml.DBSCAN(eps=1.0, min_samples=1) dbscan_float.fit(ary) print("cuML output inside `with` context") print(dbscan_float.labels_) print(type(dbscan_float.labels_)) # use cuml again outside the context manager dbscan_float2 = cuml.DBSCAN(eps=1.0, min_samples=1) dbscan_float2.fit(ary) print("cuML default output") print(dbscan_float2.labels_) print(type(dbscan_float2.labels_))Output:
cuML output inside `with` context 0 0 1 1 2 2 dtype: int32 <class 'cudf.core.series.Series'> cuML default output [0 1 2] <class 'cupy.core.core.ndarray'>
Verbosity Levels¶
cuML follows a verbosity model similar to Scikitlearn’s: The verbose parameter can be a boolean, or a numeric value, and higher numeric values mean more verbosity. The exact values can be set directly, or through the cuml.common.logger module, and they are:
Numeric value 
cuml.common.logger value 
Verbosity level 

0 
cuml.common.logger.level_off 
Disables all log messages 
1 
cuml.common.logger.level_critical 
Enables only critical messages 
2 
cuml.common.logger.level_error 
Enables all messages up to and including errors. 
3 
cuml.common.logger.level_warn 
Enables all messages up to and including warnings. 
4 or False 
cuml.common.logger.level_info 
Enables all messages up to and including information messages. 
5 or True 
cuml.common.logger.level_debug 
Enables all messages up to and including debug messages. 
6 
cuml.common.logger.level_trace 
Enables all messages up to and including trace messages. 
Preprocessing, Metrics, and Utilities¶
Model Selection and Data Splitting¶
model_selection.
train_test_split
(X, y, test_size: Union[float, int] = None, train_size: Union[float, int] = None, shuffle: bool = True, random_state: Union[int, cupy.random.generator.RandomState, numpy.random.mtrand.RandomState] = None, seed: Union[int, cupy.random.generator.RandomState, numpy.random.mtrand.RandomState] = None)¶Partitions device data into four collated objects, mimicking Scikitlearn’s train_test_split
 Parameters
 Xcudf.DataFrame or cuda_array_interface compliant device array
Data to split, has shape (n_samples, n_features)
 ystr, cudf.Series or cuda_array_interface compliant device array
Set of labels for the data, either a series of shape (n_samples) or the string label of a column in X (if it is a cuDF DataFrame) containing the labels
 train_sizefloat or int, optional
If float, represents the proportion [0, 1] of the data to be assigned to the training set. If an int, represents the number of instances to be assigned to the training set. Defaults to 0.8
 shufflebool, optional
Whether or not to shuffle inputs before splitting
 random_stateint, CuPy RandomState or NumPy RandomState optional
If shuffle is true, seeds the generator. Unseeded by default
 seed: random_stateint, CuPy RandomState or NumPy RandomState optional
Deprecated in favor of random_state. If shuffle is true, seeds the generator. Unseeded by default
 Returns
 X_train, X_test, y_train, y_testcudf.DataFrame or arraylike objects
Partitioned dataframes if X and y were cuDF objects. If y was provided as a column name, the column was dropped from the `X`s Partitioned numba device arrays if X and y were Numba device arrays. Partitioned CuPy arrays for any other input.
Examples
import cudf from cuml.preprocessing.model_selection import train_test_split # Generate some sample data df = cudf.DataFrame({'x': range(10), 'y': [0, 1] * 5}) print(f'Original data: {df.shape[0]} elements') # Suppose we want an 80/20 split X_train, X_test, y_train, y_test = train_test_split(df, 'y', train_size=0.8) print(f'X_train: {X_train.shape[0]} elements') print(f'X_test: {X_test.shape[0]} elements') print(f'y_train: {y_train.shape[0]} elements') print(f'y_test: {y_test.shape[0]} elements') # Alternatively, if our labels are stored separately labels = df['y'] df = df.drop(['y']) # we can also do X_train, X_test, y_train, y_test = train_test_split(df, labels, train_size=0.8)Output:
Original data: 10 elements X_train: 8 elements X_test: 2 elements y_train: 8 elements y_test: 2 elements
Feature and Label Encoding (SingleGPU)¶
 class
cuml.preprocessing.
LabelEncoder
(handle_unknown='error')¶An nvcategory based implementation of ordinal label encoding
 Parameters
 handle_unknown{‘error’, ‘ignore’}, default=’error’
Whether to raise an error or ignore if an unknown categorical feature is present during transform (default is to raise). When this parameter is set to ‘ignore’ and an unknown category is encountered during transform or inverse transform, the resulting encoding will be null.
Examples
Converting a categorical implementation to a numerical one
from cudf import DataFrame, Series data = DataFrame({'category': ['a', 'b', 'c', 'd']}) # There are two functionally equivalent ways to do this le = LabelEncoder() le.fit(data.category) # le = le.fit(data.category) also works encoded = le.transform(data.category) print(encoded) # This method is preferred le = LabelEncoder() encoded = le.fit_transform(data.category) print(encoded) # We can assign this to a new column data = data.assign(encoded=encoded) print(data.head()) # We can also encode more data test_data = Series(['c', 'a']) encoded = le.transform(test_data) print(encoded) # After train, ordinal label can be inverse_transform() back to # string labels ord_label = cudf.Series([0, 0, 1, 2, 1]) ord_label = dask_cudf.from_cudf(data, npartitions=2) str_label = le.inverse_transform(ord_label) print(str_label)Output:
0 0 1 1 2 2 3 3 dtype: int64 0 0 1 1 2 2 3 3 dtype: int32 category encoded 0 a 0 1 b 1 2 c 2 3 d 3 0 2 1 0 dtype: int64 0 a 1 a 2 b 3 c 4 b dtype: objectMethods
fit
(self, y)¶Fit a LabelEncoder (nvcategory) instance to a set of categories
 Parameters
 ycudf.Series
Series containing the categories to be encoded. It’s elements may or may not be unique
 Returns
 selfLabelEncoder
A fitted instance of itself to allow method chaining
fit_transform
(self, y: cudf.core.series.Series) → cudf.core.series.Series¶Simultaneously fit and transform an input
This is functionally equivalent to (but faster than) LabelEncoder().fit(y).transform(y)
inverse_transform
(self, y: cudf.core.series.Series) → cudf.core.series.Series¶Revert ordinal label to original label
 Parameters
 ycudf.Series, dtype=int32
Ordinal labels to be reverted
 Returns
 revertedcudf.Series
Reverted labels
transform
(self, y: cudf.core.series.Series) → cudf.core.series.Series¶Transform an input into its categorical keys.
This is intended for use with small inputs relative to the size of the dataset. For fitting and transforming an entire dataset, prefer fit_transform.
 Parameters
 ycudf.Series
Input keys to be transformed. Its values should match the categories given to fit
 Returns
 encodedcudf.Series
The ordinally encoded input series
 Raises
 KeyError
if a category appears that was not seen in fit
 class
cuml.preprocessing.
LabelBinarizer
(neg_label=0, pos_label=1, sparse_output=False)¶A multiclass dummy encoder for labels.
Examples
Create an array with labels and dummy encode them
import cupy as cp from cuml.preprocessing import LabelBinarizer labels = cp.asarray([0, 5, 10, 7, 2, 4, 1, 0, 0, 4, 3, 2, 1], dtype=cp.int32) lb = LabelBinarizer() encoded = lb.fit_transform(labels) print(str(encoded) decoded = lb.inverse_transform(encoded) print(str(decoded)Output:
[[1 0 0 0 0 0 0 0] [0 0 0 0 0 1 0 0] [0 0 0 0 0 0 0 1] [0 0 0 0 0 0 1 0] [0 0 1 0 0 0 0 0] [0 0 0 0 1 0 0 0] [0 1 0 0 0 0 0 0] [1 0 0 0 0 0 0 0] [1 0 0 0 0 0 0 0] [0 0 0 0 1 0 0 0] [0 0 0 1 0 0 0 0] [0 0 1 0 0 0 0 0] [0 1 0 0 0 0 0 0]] [ 0 5 10 7 2 4 1 0 0 4 3 2 1]Methods
fit
(self, y)Fit label binarizer
fit_transform
(self, y)Fit label binarizer and transform multiclass labels to their dummyencoded representation.
inverse_transform
(self, y[, threshold])Transform binary labels back to original multiclass labels
transform
(self, y)Transform multiclass labels to their dummyencoded representation labels.
fit
(self, y)¶Fit label binarizer
 Parameters
 yarray of shape [n_samples,] or [n_samples, n_classes]
Target values. The 2d matrix should only contain 0 and 1, represents multilabel classification.
 Returns
 selfreturns an instance of self.
fit_transform
(self, y)¶Fit label binarizer and transform multiclass labels to their dummyencoded representation.
 Parameters
 yarray of shape [n_samples,] or [n_samples, n_classes]
 Returns
 arrarray with encoded labels
inverse_transform
(self, y, threshold=None)¶Transform binary labels back to original multiclass labels
 Parameters
 yarray of shape [n_samples, n_classes]
 thresholdfloat this value is currently ignored
 Returns
 arrarray with original labels
transform
(self, y)¶Transform multiclass labels to their dummyencoded representation labels.
 Parameters
 yarray of shape [n_samples,] or [n_samples, n_classes]
 Returns
 arrarray with encoded labels
preprocessing.
label_binarize
(y, classes, neg_label=0, pos_label=1, sparse_output=False)¶A stateless helper function to dummy encode multiclass labels.
 Parameters
 yarraylike of size [n_samples,] or [n_samples, n_classes]
 classesthe set of unique classes in the input
 neg_labelinteger the negative value for transformed output
 pos_labelinteger the positive value for transformed output
 sparse_outputbool whether to return sparse array
 class
cuml.preprocessing.
OneHotEncoder
(categories='auto', drop=None, sparse=True, dtype=<class 'float'>, handle_unknown='error')¶Encode categorical features as a onehot numeric array. The input to this estimator should be a cuDF.DataFrame or a cupy.ndarray, denoting the unique values taken on by categorical (discrete) features. The features are encoded using a onehot (aka ‘oneofK’ or ‘dummy’) encoding scheme. This creates a binary column for each category and returns a sparse matrix or dense array (depending on the
sparse
parameter). By default, the encoder derives the categories based on the unique values in each feature. Alternatively, you can also specify the categories manually. Note: a onehot encoding of y labels should use a LabelBinarizer instead.
 Parameters
 categories‘auto’ an cupy.ndarray or a cudf.DataFrame, default=’auto’
Categories (unique values) per feature:
‘auto’ : Determine categories automatically from the training data.
DataFrame/ndarray :
categories[col]
holds the categories expected in the feature col. drop‘first’, None, a dict or a list, default=None
Specifies a methodology to use to drop one of the categories per feature. This is useful in situations where perfectly collinear features cause problems, such as when feeding the resulting data into a neural network or an unregularized regression.
None : retain all features (the default).
‘first’ : drop the first category in each feature. If only one category is present, the feature will be dropped entirely.
dict/list :
drop[col]
is the category in feature col that should be dropped. sparsebool, default=False
This feature was deactivated and will give an exception when True. The reason is because sparse matrix are not fully supported by cupy yet, causing incorrect values when computing one hot encodings. See https://github.com/cupy/cupy/issues/3223
 dtypenumber type, default=np.float
Desired datatype of transform’s output.
 handle_unknown{‘error’, ‘ignore’}, default=’error’
Whether to raise an error or ignore if an unknown categorical feature is present during transform (default is to raise). When this parameter is set to ‘ignore’ and an unknown category is encountered during transform, the resulting onehot encoded columns for this feature will be all zeros. In the inverse transform, an unknown category will be denoted as None.
 Attributes
 drop_idx_array of shape (n_features,)
drop_idx_[i]
is the index incategories_[i]
of the category to be dropped for each feature. None if all the transformed features will be retained.Methods
fit
(self, X)Fit OneHotEncoder to X.
fit_transform
(self, X)Fit OneHotEncoder to X, then transform X.
inverse_transform
(self, X)Convert the data back to the original representation.
transform
(self, X)Transform X using onehot encoding.
 property
categories_
¶Returns categories used for the one hot encoding in the correct order.
fit
(self, X)¶Fit OneHotEncoder to X.
 Parameters
 XcuDF.DataFrame or cupy.ndarray, shape = (n_samples, n_features)
The data to determine the categories of each feature.
 Returns
 self
fit_transform
(self, X)¶Fit OneHotEncoder to X, then transform X. Equivalent to fit(X).transform(X).
 Parameters
 Xcudf.DataFrame or cupy.ndarray, shape = (n_samples, n_features)
The data to encode.
 Returns
 X_outsparse matrix if sparse=True else a 2d array
Transformed input.
inverse_transform
(self, X)¶Convert the data back to the original representation. In case unknown categories are encountered (all zeros in the onehot encoding),
None
is used to represent this category.The return type is the same as the type of the input used by the first call to fit on this estimator instance.
 Parameters
 Xarraylike or sparse matrix, shape [n_samples, n_encoded_features]
The transformed data.
 Returns
 X_trcudf.DataFrame or cupy.ndarray
Inverse transformed array.
transform
(self, X)¶Transform X using onehot encoding.
 Parameters
 Xcudf.DataFrame or cupy.ndarray
The data to encode.
 Returns
 X_outsparse matrix if sparse=True else a 2d array
Transformed input.
Feature and Label Encoding (Daskbased MultiGPU)¶
 class
cuml.dask.preprocessing.
LabelBinarizer
(client=None, **kwargs)¶A distributed version of LabelBinarizer for onehot encoding a collection of labels.
Examples
Create an array with labels and dummy encode them
import cupy as cp from cuml.dask.preprocessing import LabelBinarizer from dask_cuda import LocalCUDACluster from dask.distributed import Client import dask cluster = LocalCUDACluster() client = Client(cluster) labels = cp.asarray([0, 5, 10, 7, 2, 4, 1, 0, 0, 4, 3, 2, 1], dtype=cp.int32) labels = dask.array.from_array(labels) lb = LabelBinarizer() encoded = lb.fit_transform(labels) print(str(encoded.compute()) decoded = lb.inverse_transform(encoded) print(str(decoded.compute())Output:
[[1 0 0 0 0 0 0 0] [0 0 0 0 0 1 0 0] [0 0 0 0 0 0 0 1] [0 0 0 0 0 0 1 0] [0 0 1 0 0 0 0 0] [0 0 0 0 1 0 0 0] [0 1 0 0 0 0 0 0] [1 0 0 0 0 0 0 0] [1 0 0 0 0 0 0 0] [0 0 0 0 1 0 0 0] [0 0 0 1 0 0 0 0] [0 0 1 0 0 0 0 0] [0 1 0 0 0 0 0 0]] [ 0 5 10 7 2 4 1 0 0 4 3 2 1]Methods
fit
(self, y)Fit label binarizer
fit_transform
(self, y)Fit the label encoder and return transformed labels
inverse_transform
(self, y[, threshold])Invert a set of encoded labels back to original labels
transform
(self, y)Transform and return encoded labels
fit
(self, y)¶Fit label binarizer
 Parameters
 yDask.Array of shape [n_samples,] or [n_samples, n_classes]
chunked by row. Target values. The 2d matrix should only contain 0 and 1, represents multilabel classification.
 Returns
 selfreturns an instance of self.
fit_transform
(self, y)¶Fit the label encoder and return transformed labels
 Parameters
 yDask.Array of shape [n_samples,] or [n_samples, n_classes]
target values. The 2d matrix should only contain 0 and 1, represents multilabel classification.
 Returns
 arrDask.Array backed by CuPy arrays containing encoded labels
inverse_transform
(self, y, threshold=None)¶Invert a set of encoded labels back to original labels
 Parameters
 yDask.Array of shape [n_samples, n_classes] containing encoded
labels
 thresholdfloat This value is currently ignored
 Returns
 arrDask.Array backed by CuPy arrays containing original labels
transform
(self, y)¶Transform and return encoded labels
 Parameters
 yDask.Array of shape [n_samples,] or [n_samples, n_classes]
 Returns
 arrDask.Array backed by CuPy arrays containing encoded labels
 class
cuml.dask.preprocessing.
OneHotEncoder
(client=None, verbose=False, **kwargs)¶Encode categorical features as a onehot numeric array. The input to this transformer should be a dask_cuDF.DataFrame or cupy dask.Array, denoting the values taken on by categorical features. The features are encoded using a onehot (aka ‘oneofK’ or ‘dummy’) encoding scheme. This creates a binary column for each category and returns a sparse matrix or dense array (depending on the
sparse
parameter). By default, the encoder derives the categories based on the unique values in each feature. Alternatively, you can also specify the categories manually.
 Parameters
 categories‘auto’, cupy.ndarray or cudf.DataFrame, default=’auto’
Categories (unique values) per feature. All categories are expected to fit on one GPU.
‘auto’ : Determine categories automatically from the training data.
DataFrame/ndarray :
categories[col]
holds the categories expected in the feature col. drop‘first’, None or a dict, default=None
Specifies a methodology to use to drop one of the categories per feature. This is useful in situations where perfectly collinear features cause problems, such as when feeding the resulting data into a neural network or an unregularized regression.
None : retain all features (the default).
‘first’ : drop the first category in each feature. If only one category is present, the feature will be dropped entirely.
Dict :
drop[col]
is the category in feature col that should be dropped. sparsebool, default=False
This feature was deactivated and will give an exception when True. The reason is because sparse matrix are not fully supported by cupy yet, causing incorrect values when computing one hot encodings. See https://github.com/cupy/cupy/issues/3223
 dtypenumber type, default=np.float
Desired datatype of transform’s output.
 handle_unknown{‘error’, ‘ignore’}, default=’error’
Whether to raise an error or ignore if an unknown categorical feature is present during transform (default is to raise). When this parameter is set to ‘ignore’ and an unknown category is encountered during transform, the resulting onehot encoded columns for this feature will be all zeros. In the inverse transform, an unknown category will be denoted as None.
Methods
fit
(self, X)Fit a multinode multigpu OneHotEncoder to X.
fit_transform
(self, X[, delayed])Fit OneHotEncoder to X, then transform X.
inverse_transform
(self, X[, delayed])Convert the data back to the original representation.
transform
(self, X[, delayed])Transform X using onehot encoding.
fit
(self, X)¶Fit a multinode multigpu OneHotEncoder to X.
 Parameters
 XDask cuDF DataFrame or CuPy backed Dask Array
The data to determine the categories of each feature.
 Returns
 self
fit_transform
(self, X, delayed=True)¶Fit OneHotEncoder to X, then transform X. Equivalent to fit(X).transform(X).
 Parameters
 XDask cuDF DataFrame or CuPy backed Dask Array
The data to encode.
 delayedbool (default = True)
Whether to execute as a delayed task or eager.
 Returns
 outDask cuDF DataFrame or CuPy backed Dask Array
Distributed object containing the transformed data
inverse_transform
(self, X, delayed=True)¶Convert the data back to the original representation. In case unknown categories are encountered (all zeros in the onehot encoding),
None
is used to represent this category.
 Parameters
 XCuPy backed Dask Array, shape [n_samples, n_encoded_features]
The transformed data.
 delayedbool (default = True)
Whether to execute as a delayed task or eager.
 Returns
 X_trDask cuDF DataFrame or CuPy backed Dask Array
Distributed object containing the inverse transformed array.
transform
(self, X, delayed=True)¶Transform X using onehot encoding.
 Parameters
 XDask cuDF DataFrame or CuPy backed Dask Array
The data to encode.
 delayedbool (default = True)
Whether to execute as a delayed task or eager.
 Returns
 outDask cuDF DataFrame or CuPy backed Dask Array
Distributed object containing the transformed input.
Dataset Generation (SingleGPU)¶
datasets.
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_samplesint or arraylike, optional (default=100)
If int, it is the total number of points equally divided among clusters. If arraylike, each element of the sequence indicates the number of samples per cluster.
 n_featuresint, optional (default=2)
The number of features for each sample.
 centersint 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 arraylike, centers must be either None or an array of length equal to the length of n_samples.
 cluster_stdfloat or sequence of floats, optional (default=1.0)
The standard deviation of the clusters.
 center_boxpair of floats (min, max), optional (default=(10.0, 10.0))
The bounding box for each cluster center when centers are generated at random.
 shuffleboolean, optional (default=True)
Shuffle the samples.
 random_stateint, RandomState instance, default=None
Determines random number generation for dataset creation. Pass an int for reproducible output across multiple function calls.
 return_centersbool, optional (default=False)
If True, then return the centers of each cluster
 order: str, optional (default=’F’)
The order of the generated samples
 dtypestr, optional (default=’float32’)
Dtype of the generated samples
 Returns
 Xdevice array of shape [n_samples, n_features]
The generated samples.
 ydevice array of shape [n_samples]
The integer labels for cluster membership of each sample.
 centersdevice array, shape [n_centers, n_features]
The centers of each cluster. Only returned if
return_centers=True
.See also
make_classification
a more intricate variant
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])
datasets.
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 nclass classification problem. This initially creates clusters of points normally distributed (std=1) about vertices of an
n_informative
dimensional hypercube with sides of length2*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 primaryn_informative
features, followed byn_redundant
linear combinations of the informative features, followed byn_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 columnsX[:, :n_informative + n_redundant + n_repeated]
.
 Parameters
 n_samplesint, optional (default=100)
The number of samples.
 n_featuresint, optional (default=20)
The total number of features. These comprise
n_informative
informative features,n_redundant
redundant features,n_repeated
duplicated features andn_featuresn_informativen_redundantn_repeated
useless features drawn at random. n_informativeint, 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_redundantint, optional (default=2)
The number of redundant features. These features are generated as random linear combinations of the informative features.
 n_repeatedint, optional (default=0)
The number of duplicated features, drawn randomly from the informative and the redundant features.
 n_classesint, optional (default=2)
The number of classes (or labels) of the classification problem.
 n_clusters_per_classint, optional (default=2)
The number of clusters per class.
 weightsarraylike 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 thann_samples
samples may be returned if the sum ofweights
exceeds 1. flip_yfloat, 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_sepfloat, optional (default=1.0)
The factor multiplying the hypercube size. Larger values spread out the clusters/classes and make the classification task easier.
 hypercubeboolean, 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.
 shiftfloat, 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].
 scalefloat, 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.
 shuffleboolean, optional (default=True)
Shuffle the samples and the features.
 random_stateint, RandomState instance or None (default)
Determines random number generation for dataset creation. Pass an int for reproducible output across multiple function calls. See Glossary.
 order: str, optional (default=’F’)
The order of the generated samples
 dtypestr, 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
 Xdevice array of shape [n_samples, n_features]
The generated samples.
 ydevice 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:
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
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
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.
Examples
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:
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]
datasets.
make_regression
()¶Generate a random regression problem.
See https://scikitlearn.org/stable/modules/generated/sklearn.datasets.make_regression.html
 Parameters
 n_samplesint, optional (default=100)
The number of samples.
 n_featuresint, optional (default=2)
The number of features.
 n_informativeint, optional (default=2)
The number of informative features, i.e., the number of features used to build the linear model used to generate the output.
 n_targetsint, 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.
 biasfloat, optional (default=0.0)
The bias term in the underlying linear model.
 effective_rankint 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_strengthfloat 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.
 noisefloat, optional (default=0.0)
The standard deviation of the gaussian noise applied to the output.
 shuffleboolean, optional (default=True)
Shuffle the samples and the features.
 coefboolean, optional (default=False)
If True, the coefficients of the underlying linear model are returned.
 random_stateint, RandomState instance or None (default)
Seed for the random number generator for dataset creation.
 dtype: string or numpy dtype (default: ‘single’)
Type of the data. Possible values: float32, float64, ‘single’, ‘float’ or ‘double’.
 handle: cuml.Handle
If it is None, a new one is created just for this function call
 Returns
 outdevice array of shape [n_samples, n_features]
The input samples.
 valuesdevice array of shape [n_samples, n_targets]
The output values.
 coefdevice array of shape [n_features, n_targets], optional
The coefficient of the underlying linear model. It is returned only if coef is True.
datasets.
make_arima
()¶Generates a dataset of time series by simulating an ARIMA process of a given order.
 Parameters
 batch_size: int
Number of time series to generate
 n_obs: int
Number of observations per series
 orderTuple[int, int, int]
Order (p, d, q) of the simulated ARIMA process
 seasonal_order: Tuple[int, int, int, int]
Seasonal ARIMA order (P, D, Q, s) of the simulated ARIMA process
 intercept: bool or int
Whether to include a constant trend mu in the simulated ARIMA process
 random_state: int, RandomState instance or None (default)
Seed for the random number generator for dataset creation.
 dtype: string or numpy dtype (default: ‘single’)
Type of the data. Possible values: float32, float64, ‘single’, ‘float’ or ‘double’
 output_type: {‘cudf’, ‘cupy’, ‘numpy’}
Type of the returned dataset
 handle: cuml.Handle
If it is None, a new one is created just for this function call
Dataset Generation (Daskbased MultiGPU)¶
cuml.dask.datasets.blobs.
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, verbose=False, order='F', dtype='float32', client=None)¶Makes labeled DaskCupy 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 Scikitlearn’s make_blobs:.
 Parameters
 n_samplesint
number of rows
 n_featuresint
number of features
 centersint 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 arraylike, centers must be either None or an array of length equal to the length of n_samples.
 cluster_stdfloat (default = 1.0)
standard deviation of points around centroid
 n_partsint (default = None)
number of partitions to generate (this can be greater than the number of workers)
 center_boxtuple (int, int) (default = (10, 10))
the bounding box which constrains all the centroids
 random_stateint (default = None)
sets random seed (or use None to reinitialize each time)
 return_centersbool, optional (default=False)
If True, then return the centers of each cluster
 verboseint or boolean (default = False)
Logging level.
 shufflebool (default=False)
Shuffles the samples on each worker.
 order: str, optional (default=’F’)
The order of the generated samples
 dtypestr, optional (default=’float32’)
Dtype of the generated samples
 clientdask.distributed.Client (optional)
Dask client to use
 Returns
 Xdask.array backed by CuPy array of shape [n_samples, n_features]
The input samples.
 ydask.array backed by CuPy array of shape [n_samples]
The output values.
 centersdask.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.
cuml.dask.datasets.classification.
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 nclass classification problem.
This initially creates clusters of points normally distributed (std=1) about vertices of an
n_informative
dimensional hypercube with sides of length2*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 primaryn_informative
features, followed byn_redundant
linear combinations of the informative features, followed byn_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 columnsX[:, :n_informative + n_redundant + n_repeated]
.
 Parameters
 n_samplesint, optional (default=100)
The number of samples.
 n_featuresint, optional (default=20)
The total number of features. These comprise
n_informative
informative features,n_redundant
redundant features,n_repeated
duplicated features andn_featuresn_informativen_redundantn_repeated
useless features drawn at random. n_informativeint, 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_redundantint, optional (default=2)
The number of redundant features. These features are generated as random linear combinations of the informative features.
 n_repeatedint, optional (default=0)
The number of duplicated features, drawn randomly from the informative and the redundant features.
 n_classesint, optional (default=2)
The number of classes (or labels) of the classification problem.
 n_clusters_per_classint, optional (default=2)
The number of clusters per class.
 weightsarraylike 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 thann_samples
samples may be returned if the sum ofweights
exceeds 1. flip_yfloat, 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_sepfloat, optional (default=1.0)
The factor multiplying the hypercube size. Larger values spread out the clusters/classes and make the classification task easier.
 hypercubeboolean, 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.
 shiftfloat, 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].
 scalefloat, 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.
 shuffleboolean, optional (default=True)
Shuffle the samples and the features.
 random_stateint, RandomState instance or None (default)
Determines random number generation for dataset creation. Pass an int for reproducible output across multiple function calls. See Glossary.
 order: str, optional (default=’F’)
The order of the generated samples
 dtypestr, optional (default=’float32’)
Dtype of the generated samples
 n_partsint (default = None)
number of partitions to generate (this can be greater than the number of workers)
 Returns
 Xdask.array backed by CuPy array of shape [n_samples, n_features]
The generated samples.
 ydask.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:
We generate centroids of shape (n_centroids, n_informative)
We generate an informative covariance of shape (n_centroids, n_informative, n_informative)
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
Examples
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:
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]
cuml.dask.datasets.regression.
make_low_rank_matrix
(n_samples=100, n_features=100, effective_rank=10, tail_strength=0.5, random_state=None, n_parts=1, n_samples_per_part=None, dtype='float32')¶Generate a mostly low rank matrix with bellshaped singular values
 Parameters
 n_samplesint, optional (default=100)
The number of samples.
 n_featuresint, optional (default=100)
The number of features.
 effective_rankint, optional (default=10)
The approximate number of singular vectors required to explain most of the data by linear combinations.
 tail_strengthfloat between 0.0 and 1.0, optional (default=0.5)
The relative importance of the fat noisy tail of the singular values profile.
 random_stateint, 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_partsint, optional (default=1)
The number of parts of work.
 dtype: str, optional (default=’float32’)
dtype of generated data
 Returns
 XDaskCuPy array of shape [n_samples, n_features]
The matrix.
cuml.dask.datasets.regression.
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 rankfat 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_samplesint, optional (default=100)
The number of samples.
 n_featuresint, optional (default=100)
The number of features.
 n_informativeint, 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_targetsint, 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.
 biasfloat, optional (default=0.0)
The bias term in the underlying linear model.
 effective_rankint 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_strengthfloat 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.
 noisefloat, optional (default=0.0)
The standard deviation of the gaussian noise applied to the output.
 shuffleboolean, optional (default=False)
Shuffle the samples and the features.
 coefboolean, optional (default=False)
If True, the coefficients of the underlying linear model are returned.
 random_stateint, 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_partsint, optional (default=1)
The number of parts of work.
 orderstr, optional (default=’F’)
Rowmajor or Colmajor
 dtype: str, optional (default=’float32’)
dtype of generated data
 use_full_low_rankboolean (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
 XDaskCuPy array of shape [n_samples, n_features]
The input samples.
 yDaskCuPy array of shape [n_samples] or [n_samples, n_targets]
The output values.
 coefDaskCuPy 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:
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)
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
When shuffle = True and order = F, there are memory spikes to shuffle the F order arrays
NOTE: If outofmemory errors are encountered in any of the above configurations, try increasing the n_parts parameter.
Metrics¶
cuml.metrics.regression.
mean_absolute_error
(*args, **kwargs)¶Mean absolute error regression loss
Be careful when using this metric with float32 inputs as the result can be slightly incorrect because of floating point precision if the input is large enough. float64 will have lower numerical error.
 Parameters
 y_truearraylike (device or host) shape = (n_samples,)
or (n_samples, n_outputs) Ground truth (correct) target values.
 y_predarraylike (device or host) shape = (n_samples,)
or (n_samples, n_outputs) Estimated target values.
 sample_weightarraylike (device or host) shape = (n_samples,), optional
Sample weights.
 multioutputstring in [‘raw_values’, ‘uniform_average’]
or arraylike of shape (n_outputs) Defines aggregating of multiple output values. Arraylike value defines weights used to average errors. ‘raw_values’ : Returns a full set of errors in case of multioutput input. ‘uniform_average’ : Errors of all outputs are averaged with uniform weight.
 Returns
 lossfloat or ndarray of floats
If multioutput is ‘raw_values’, then mean absolute error is returned for each output separately. If multioutput is ‘uniform_average’ or an ndarray of weights, then the weighted average of all output errors is returned.
MAE output is nonnegative floating point. The best value is 0.0.
cuml.metrics.regression.
mean_squared_error
(*args, **kwargs)¶Mean squared error regression loss
Be careful when using this metric with float32 inputs as the result can be slightly incorrect because of floating point precision if the input is large enough. float64 will have lower numerical error.
 Parameters
 y_truearraylike (device or host) shape = (n_samples,)
or (n_samples, n_outputs) Ground truth (correct) target values.
 y_predarraylike (device or host) shape = (n_samples,)
or (n_samples, n_outputs) Estimated target values.
 sample_weightarraylike (device or host) shape = (n_samples,), optional
Sample weights.
 multioutputstring in [‘raw_values’, ‘uniform_average’]
or arraylike of shape (n_outputs) Defines aggregating of multiple output values. Arraylike value defines weights used to average errors. ‘raw_values’ : Returns a full set of errors in case of multioutput input. ‘uniform_average’ : Errors of all outputs are averaged with uniform weight.
 squaredboolean value, optional (default = True)
If True returns MSE value, if False returns RMSE value.
 Returns
 lossfloat or ndarray of floats
A nonnegative floating point value (the best value is 0.0), or an array of floating point values, one for each individual target.
cuml.metrics.regression.
mean_squared_log_error
(*args, **kwargs)¶Mean squared log error regression loss
Be careful when using this metric with float32 inputs as the result can be slightly incorrect because of floating point precision if the input is large enough. float64 will have lower numerical error.
 Parameters
 y_truearraylike (device or host) shape = (n_samples,)
or (n_samples, n_outputs) Ground truth (correct) target values.
 y_predarraylike (device or host) shape = (n_samples,)
or (n_samples, n_outputs) Estimated target values.
 sample_weightarraylike (device or host) shape = (n_samples,), optional
Sample weights.
 multioutputstring in [‘raw_values’, ‘uniform_average’]
or arraylike of shape (n_outputs) Defines aggregating of multiple output values. Arraylike value defines weights used to average errors. ‘raw_values’ : Returns a full set of errors in case of multioutput input. ‘uniform_average’ : Errors of all outputs are averaged with uniform weight.
 squaredboolean value, optional (default = True)
If True returns MSE value, if False returns RMSE value.
 Returns
 lossfloat or ndarray of floats
A nonnegative floating point value (the best value is 0.0), or an array of floating point values, one for each individual target.
cuml.metrics.regression.
r2_score
(y, y_hat, convert_dtype=False, handle=None)¶Calculates r2 score between y and y_hat
 Parameters
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 y_hatarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the fit method will, when necessary, convert y_hat to be the same data type as y if they differ. This will increase memory used for the method.
 Returns
 trustworthiness scoredouble
Trustworthiness of the lowdimensional embedding
cuml.metrics.accuracy.
accuracy_score
(ground_truth, predictions, handle=None, convert_dtype=True)¶Calcuates the accuracy score of a classification model.
 Parameters
 handlecuml.Handle
 predictionNumPy ndarray or Numba device
The labels predicted by the model for the test dataset
 ground_truthNumPy ndarray, Numba device
The ground truth labels of the test dataset
 Returns
 float
The accuracy of the model used for prediction
cuml.metrics.trustworthiness.
trustworthiness
(X, X_embedded, handle=None, n_neighbors=5, metric='euclidean', should_downcast=True, convert_dtype=False, batch_size=512)¶Expresses to what extent the local structure is retained in embedding. The score is defined in the range [0, 1].
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 X_embeddedarraylike (device or host) shape= (n_samples, n_features)
Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 n_neighborsint, optional (default: 5)
Number of neighbors considered
 convert_dtypebool, optional (default = False)
When set to True, the trustworthiness method will automatically convert the inputs to np.float32.
 Returns
 trustworthiness scoredouble
Trustworthiness of the lowdimensional embedding
cuml.metrics.cluster.adjustedrandindex.
adjusted_rand_score
(labels_true, labels_pred, handle=None, convert_dtype=True)¶Adjusted_rand_score is a clustering similarity metric based on the Rand index and is corrected for chance.
 Parameters
 labels_trueGround truth labels to be used as a reference
labels_pred : Array of predicted labels used to evaluate the model
handle : cuml.Handle
 Returns
 float
The adjusted rand index value between 1.0 and 1.0
cuml.metrics.cluster.entropy.
cython_entropy
(clustering, base=None, handle=None)¶Computes the entropy of a distribution for given probability values.
 Parameters
 clusteringarraylike (device or host) shape = (n_samples,)
Clustering of labels. Probabilities are computed based on occurrences of labels. For instance, to represent a fair coin (2 equally possible outcomes), the clustering could be [0,1]. For a biased coin with 2/3 probability for tail, the clustering could be [0, 0, 1].
 base: float, optional
The logarithmic base to use, defaults to e (natural logarithm).
 handlecuml.Handle
Specifies the cuml.handle that holds internal CUDA state for computations in this model. Most importantly, this specifies the CUDA stream that will be used for the model’s computations, so users can run different models concurrently in different streams by creating handles in several streams. If it is None, a new one is created.
 Returns
 Sfloat
The calculated entropy.
Benchmarking¶
 class
cuml.benchmark.algorithms.
AlgorithmPair
(cpu_class, cuml_class, shared_args, cuml_args={}, cpu_args={}, name=None, accepts_labels=True, cpu_data_prep_hook=None, cuml_data_prep_hook=None, accuracy_function=None, bench_func=<function fit>, setup_cpu_func=None, setup_cuml_func=None)¶Wraps a cuML algorithm and (optionally) a cpubased algorithm (typically scikitlearn, but does not need to be as long as it offers fit and predict or transform methods). Provides mechanisms to run each version with default arguments. If no CPUbased version of the algorithm is available, pass None for the cpu_class when instantiating
 Parameters
 cpu_classclass
Class for CPU version of algorithm. Set to None if not available.
 cuml_classclass
Class for cuML algorithm
 shared_argsdict
Arguments passed to both implementations’s initializer
 cuml_argsdict
Arguments only passed to cuml’s initializer
 cpu_args dict
Arguments only passed to sklearn’s initializer
 accepts_labelsboolean
If True, the fit methods expects both X and y inputs. Otherwise, it expects only an X input.
 data_prep_hookfunction (data > data)
Optional function to run on input data before passing to fit
 accuracy_functionfunction (y_test, y_pred)
Function that returns a scalar representing accuracy
 bench_funccustom function to perform fit/predict/transform
calls.
Methods
run_cpu
(self, data, **override_args)Runs the cpubased algorithm’s fit method on specified data
run_cuml
(self, data, **override_args)Runs the cumlbased algorithm’s fit method on specified data
setup_cpu
setup_cuml
run_cpu
(self, data, **override_args)¶Runs the cpubased algorithm’s fit method on specified data
run_cuml
(self, data, **override_args)¶Runs the cumlbased algorithm’s fit method on specified data
cuml.benchmark.algorithms.
algorithm_by_name
(name)¶Returns the algorithm pair with the name ‘name’ (caseinsensitive)
cuml.benchmark.algorithms.
all_algorithms
()¶Returns all defined AlgorithmPair objects
Wrappers to run ML benchmarks
 class
cuml.benchmark.runners.
AccuracyComparisonRunner
(bench_rows, bench_dims, dataset_name='blobs', input_type='numpy', test_fraction=0.1, n_reps=1)¶Wrapper to run an algorithm with multiple dataset sizes and compute accuracy and speedup of cuml relative to sklearn baseline.
 class
cuml.benchmark.runners.
BenchmarkTimer
(reps=1)¶Provides a context manager that runs a code block reps times and records results to the instance variable timings. Use like:
timer = BenchmarkTimer(rep=5) for _ in timer.benchmark_runs(): ... do something ... print(np.min(timer.timings))Methods
benchmark_runs
 class
cuml.benchmark.runners.
SpeedupComparisonRunner
(bench_rows, bench_dims, dataset_name='blobs', input_type='numpy', n_reps=1)¶Wrapper to run an algorithm with multiple dataset sizes and compute speedup of cuml relative to sklearn baseline.
Methods
run
cuml.benchmark.runners.
run_variations
(algos, dataset_name, bench_rows, bench_dims, param_override_list=[{}], cuml_param_override_list=[{}], cpu_param_override_list=[{}], dataset_param_override_list=[{}], input_type='numpy', test_fraction=0.1, run_cpu=True, raise_on_error=False, n_reps=1)¶Runs each algo in algos once per bench_rows X bench_dims X params_override_list X cuml_param_override_list combination and returns a dataframe containing timing and accuracy data.
 Parameters
 algosstr or list
Name of algorithms to run and evaluate
 dataset_namestr
Name of dataset to use
 bench_rowslist of int
Dataset row counts to test
 bench_dimslist of int
Dataset column counts to test
 param_override_listlist of dict
Dicts containing parameters to pass to __init__. Each dict specifies parameters to override in one run of the algorithm.
 cuml_param_override_listlist of dict
Dicts containing parameters to pass to __init__ of the cuml algo only.
 cpu_param_override_listlist of dict
Dicts containing parameters to pass to __init__ of the cpu algo only.
 dataset_param_override_listdict
Dicts containing parameters to pass to dataset generator function
 test_fractionfloat
The fraction of data to use for testing.
 run_cpuboolean
If True, run the cpubased algorithm for comparison
Data generators for cuML benchmarks
The main entry point for consumers is gen_data, which wraps the underlying data generators.
Notes when writing new generators:
 Each generator is a function that accepts:
n_samples (set to 0 for ‘default’)
n_features (set to 0 for ‘default’)
random_state
(and optional generatorspecific parameters)
The function should return a 2tuple (X, y), where X is a Pandas dataframe and y is a Pandas series. If the generator does not produce labels, it can return (X, None)
A set of helper functions (convert_*) can convert these to alternative formats. Future revisions may support generating cudf dataframes or GPU arrays directly instead.
cuml.benchmark.datagen.
gen_data
(dataset_name, dataset_format, n_samples=0, n_features=0, random_state=42, test_fraction=0.0, **kwargs)¶Returns a tuple of data from the specified generator.
 Parameters
 dataset_namestr
Dataset to use. Can be a synthetic generator (blobs or regression) or a specified dataset (higgs currently, others coming soon)
 dataset_formatstr
Type of data to return. (One of cudf, numpy, pandas, gpuarray)
 n_samplesint
Number of samples to include in training set (regardless of test split)
 test_fractionfloat
Fraction of the dataset to partition randomly into the test set. If this is 0.0, no test set will be created.
cuml.benchmark.datagen.
load_higgs
()¶Returns the Higgs Boson dataset as an X, y tuple of dataframes.
Regression and Classification¶
Linear Regression¶

class
cuml.
LinearRegression
¶ LinearRegression is a simple machine learning model where the response y is modelled by a linear combination of the predictors in X.
cuML’s LinearRegression expects either a cuDF DataFrame or a NumPy matrix and provides 2 algorithms SVD and Eig to fit a linear model. SVD is more stable, but Eig (default) is much faster.
 Parameters
 algorithm‘eig’ or ‘svd’ (default = ‘eig’)
Eig uses a eigendecomposition of the covariance matrix, and is much faster. SVD is slower, but guaranteed to be stable.
 fit_interceptboolean (default = True)
If True, LinearRegression tries to correct for the global mean of y. If False, the model expects that you have centered the data.
 normalizeboolean (default = False)
This parameter is ignored when fit_intercept is set to False. If True, the predictors in X will be normalized by dividing by it’s L2 norm. If False, no scaling will be done.
Notes
LinearRegression suffers from multicollinearity (when columns are correlated with each other), and variance explosions from outliers. Consider using Ridge Regression to fix the multicollinearity problem, and consider maybe first DBSCAN to remove the outliers, or statistical analysis to filter possible outliers.
Applications of LinearRegression
LinearRegression is used in regression tasks where one wants to predict say sales or house prices. It is also used in extrapolation or time series tasks, dynamic systems modelling and many other machine learning tasks. This model should be first tried if the machine learning problem is a regression task (predicting a continuous variable).
For additional information, see scikitlearn’s OLS documentation.
For an additional example see the OLS notebook.
Examples
import numpy as np import cudf # Both import methods supported from cuml import LinearRegression from cuml.linear_model import LinearRegression lr = LinearRegression(fit_intercept = True, normalize = False, algorithm = "eig") X = cudf.DataFrame() X['col1'] = np.array([1,1,2,2], dtype = np.float32) X['col2'] = np.array([1,2,2,3], dtype = np.float32) y = cudf.Series( np.array([6.0, 8.0, 9.0, 11.0], dtype = np.float32) ) reg = lr.fit(X,y) print("Coefficients:") print(reg.coef_) print("Intercept:") print(reg.intercept_) X_new = cudf.DataFrame() X_new['col1'] = np.array([3,2], dtype = np.float32) X_new['col2'] = np.array([5,5], dtype = np.float32) preds = lr.predict(X_new) print("Predictions:") print(preds)
Output:
Coefficients: 0 1.0000001 1 1.9999998 Intercept: 3.0 Predictions: 0 15.999999 1 14.999999
 Attributes
 coef_array, shape (n_features)
The estimated coefficients for the linear regression model.
 intercept_array
The independent term. If fit_intercept_ is False, will be 0.
Methods
fit
(self, X, y[, convert_dtype])Fit the model with X and y.
get_param_names
(self)predict
(self, X[, convert_dtype])Predicts y values for X.

fit
(self, X, y, convert_dtype=False)¶ Fit the model with X and y.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the fit method will, when necessary, convert y to be the same data type as X if they differ. This will increase memory used for the method.

get_param_names
(self)¶

predict
(self, X, convert_dtype=False)¶ Predicts y values for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 y: cuDF DataFrame
Dense vector (floats or doubles) of shape (n_samples, 1)
Logistic Regression¶

class
cuml.
LogisticRegression
¶ LogisticRegression is a linear model that is used to model probability of occurrence of certain events, for example probability of success or fail of an event.
cuML’s LogisticRegression can take arraylike objects, either in host as NumPy arrays or in device (as Numba or __cuda_array_interface__ compliant), in addition to cuDF objects. It provides both singleclass (using sigmoid loss) and multipleclass (using softmax loss) variants, depending on the input variables
Only one solver option is currently available: QuasiNewton (QN) algorithms. Even though it is presented as a single option, this solver resolves to two different algorithms underneath:
OrthantWise Limited Memory QuasiNewton (OWLQN) if there is l1 regularization
Limited Memory BFGS (LBFGS) otherwise.
Note that, just like in Scikitlearn, the bias will not be regularized.
 Parameters
 penalty: ‘none’, ‘l1’, ‘l2’, ‘elasticnet’ (default = ‘l2’)
Used to specify the norm used in the penalization. If ‘none’ or ‘l2’ are selected, then LBFGS solver will be used. If ‘l1’ is selected, solver OWLQN will be used. If ‘elasticnet’ is selected, OWLQN will be used if l1_ratio > 0, otherwise LBFGS will be used.
 tol: float (default = 1e4)
The training process will stop if current_loss > previous_loss  tol
 C: float (default = 1.0)
Inverse of regularization strength; must be a positive float.
 fit_intercept: boolean (default = True)
If True, the model tries to correct for the global mean of y. If False, the model expects that you have centered the data.
 class_weight: None
Custom class weighs are currently not supported.
 max_iter: int (default = 1000)
Maximum number of iterations taken for the solvers to converge.
 linesearch_max_iter: int (default = 50)
Max number of linesearch iterations per outer iteration used in the lbfgs and owl QN solvers.
 verboseint or boolean (default = False)
Controls verbose level of logging.
 l1_ratio: float or None, optional (default=None)
The ElasticNet mixing parameter, with 0 <= l1_ratio <= 1
 solver: ‘qn’, ‘lbfgs’, ‘owl’ (default=’qn’).
Algorithm to use in the optimization problem. Currently only qn is supported, which automatically selects either LBFGS or OWLQN depending on the conditions of the l1 regularization described above. Options ‘lbfgs’ and ‘owl’ are just convenience values that end up using the same solver following the same rules.
Notes
cuML’s LogisticRegression uses a different solver that the equivalent Scikitlearn, except when there is no penalty and solver=lbfgs is used in Scikitlearn. This can cause (smaller) differences in the coefficients and predictions of the model, similar to using different solvers in Scikitlearn.
For additional information, see Scikitlearn’s LogistRegression <https://scikitlearn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html>`_.
Examples
import cudf import numpy as np # Both import methods supported # from cuml import LogisticRegression from cuml.linear_model import LogisticRegression X = cudf.DataFrame() X['col1'] = np.array([1,1,2,2], dtype = np.float32) X['col2'] = np.array([1,2,2,3], dtype = np.float32) y = cudf.Series( np.array([0.0, 0.0, 1.0, 1.0], dtype = np.float32) ) reg = LogisticRegression() reg.fit(X,y) print("Coefficients:") print(reg.coef_.to_output('cupy')) print("Intercept:") print(reg.intercept_.to_output('cupy')) X_new = cudf.DataFrame() X_new['col1'] = np.array([1,5], dtype = np.float32) X_new['col2'] = np.array([2,5], dtype = np.float32) preds = reg.predict(X_new) print("Predictions:") print(preds)
Output:
Coefficients: 0.22309814 0.21012752 Intercept: 0.7548761 Predictions: 0 0.0 1 1.0
 Attributes
 coef_: dev array, dim (n_classes, n_features) or (n_classes, n_features+1)
The estimated coefficients for the linear regression model. Note: this includes the intercept as the last column if fit_intercept is True
 intercept_: device array (n_classes, 1)
The independent term. If fit_intercept_ is False, will be 0.
Methods
decision_function
(self, X[, convert_dtype])Gives confidence score for X
fit
(self, X, y[, convert_dtype])Fit the model with X and y.
get_param_names
(self)predict
(self, X[, convert_dtype])Predicts the y for X.
predict_log_proba
(self, X[, convert_dtype])Predicts the log class probabilities for each class in X
predict_proba
(self, X[, convert_dtype])Predicts the class probabilities for each class in X
score
(self, X, y[, convert_dtype])Calculates the accuracy metric score of the model for X.

decision_function
(self, X, convert_dtype=False)¶ Gives confidence score for X
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 y: arraylike (device)
Dense matrix (floats or doubles) of shape (n_samples, n_classes)

fit
(self, X, y, convert_dtype=False)¶ Fit the model with X and y.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the fit method will, when necessary, convert y to be the same data type as X if they differ. This will increase memory used for the method.

get_param_names
(self)¶

predict
(self, X, convert_dtype=False)¶ Predicts the y for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 y: cuDF DataFrame
Dense vector (floats or doubles) of shape (n_samples, 1)

predict_log_proba
(self, X, convert_dtype=False)¶ Predicts the log class probabilities for each class in X
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 y: arraylike (device)
Dense matrix (floats or doubles) of shape (n_samples, n_classes)

predict_proba
(self, X, convert_dtype=False)¶ Predicts the class probabilities for each class in X
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 y: arraylike (device)
Dense matrix (floats or doubles) of shape (n_samples, n_classes)

score
(self, X, y, convert_dtype=False)¶ Calculates the accuracy metric score of the model for X.
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Observations for which labels score will be calculated. Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Ground truth labels to compare predictions to for the score. Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
Ridge Regression¶

class
cuml.
Ridge
¶ Ridge extends LinearRegression by providing L2 regularization on the coefficients when predicting response y with a linear combination of the predictors in X. It can reduce the variance of the predictors, and improves the conditioning of the problem.
cuML’s Ridge can take arraylike objects, either in host as NumPy arrays or in device (as Numba or __cuda_array_interface__ compliant), in addition to cuDF objects. It provides 3 algorithms: SVD, Eig and CD to fit a linear model. In general SVD uses significantly more memory and is slower than Eig. If using CUDA 10.1, the memory difference is even bigger than in the other supported CUDA versions. However, SVD is more stable than Eig (default). CD uses Coordinate Descent and can be faster when data is large.
 Parameters
 alphafloat (default = 1.0)
Regularization strength  must be a positive float. Larger values specify stronger regularization. Array input will be supported later.
 solver{‘eig’, ‘svd’, ‘cd’} (default = ‘eig’)
Eig uses a eigendecomposition of the covariance matrix, and is much faster. SVD is slower, but guaranteed to be stable. CD or Coordinate Descent is very fast and is suitable for large problems.
 fit_interceptboolean (default = True)
If True, Ridge tries to correct for the global mean of y. If False, the model expects that you have centered the data.
 normalizeboolean (default = False)
If True, the predictors in X will be normalized by dividing by it’s L2 norm. If False, no scaling will be done.
Notes
Ridge provides L2 regularization. This means that the coefficients can shrink to become very small, but not zero. This can cause issues of interpretability on the coefficients. Consider using Lasso, or thresholding small coefficients to zero.
Applications of Ridge
Ridge Regression is used in the same way as LinearRegression, but does not suffer from multicollinearity issues. Ridge is used in insurance premium prediction, stock market analysis and much more.
For additional docs, see Scikitlearn’s Ridge Regression.
Examples
import numpy as np import cudf # Both import methods supported from cuml import Ridge from cuml.linear_model import Ridge alpha = np.array([1e5]) ridge = Ridge(alpha = alpha, fit_intercept = True, normalize = False, solver = "eig") X = cudf.DataFrame() X['col1'] = np.array([1,1,2,2], dtype = np.float32) X['col2'] = np.array([1,2,2,3], dtype = np.float32) y = cudf.Series( np.array([6.0, 8.0, 9.0, 11.0], dtype = np.float32) ) result_ridge = ridge.fit(X, y) print("Coefficients:") print(result_ridge.coef_) print("Intercept:") print(result_ridge.intercept_) X_new = cudf.DataFrame() X_new['col1'] = np.array([3,2], dtype = np.float32) X_new['col2'] = np.array([5,5], dtype = np.float32) preds = result_ridge.predict(X_new) print("Predictions:") print(preds)
Output:
Coefficients: 0 1.0000001 1 1.9999998 Intercept: 3.0 Preds: 0 15.999999 1 14.999999
 Attributes
 coef_array, shape (n_features)
The estimated coefficients for the linear regression model.
 intercept_array
The independent term. If fit_intercept_ is False, will be 0.
Methods
fit
(self, X, y[, convert_dtype])Fit the model with X and y.
get_param_names
(self)predict
(self, X[, convert_dtype])Predicts the y for X.

fit
(self, X, y, convert_dtype=False)¶ Fit the model with X and y.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the fit method will, when necessary, convert y to be the same data type as X if they differ. This will increase memory used for the method.

get_param_names
(self)¶

predict
(self, X, convert_dtype=False)¶ Predicts the y for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 y: cuDF DataFrame
Dense vector (floats or doubles) of shape (n_samples, 1)
Lasso Regression¶

class
cuml.
Lasso
¶ Lasso extends LinearRegression by providing L1 regularization on the coefficients when predicting response y with a linear combination of the predictors in X. It can zero some of the coefficients for feature selection and improves the conditioning of the problem.
cuML’s Lasso can take arraylike objects, either in host as NumPy arrays or in device (as Numba or __cuda_array_interface__ compliant), in addition to cuDF objects. It uses coordinate descent to fit a linear model.
 Parameters
 alphafloat (default = 1.0)
Constant that multiplies the L1 term. alpha = 0 is equivalent to an ordinary least square, solved by the LinearRegression class. For numerical reasons, using alpha = 0 with the Lasso class is not advised. Given this, you should use the LinearRegression class.
 fit_interceptboolean (default = True)
If True, Lasso tries to correct for the global mean of y. If False, the model expects that you have centered the data.
 normalizeboolean (default = False)
If True, the predictors in X will be normalized by dividing by it’s L2 norm. If False, no scaling will be done.
 max_iterint
The maximum number of iterations
 tolfloat (default = 1e3)
The tolerance for the optimization: if the updates are smaller than tol, the optimization code checks the dual gap for optimality and continues until it is smaller than tol.
 selection{‘cyclic’, ‘random’} (default=’cyclic’)
If set to ‘random’, a random coefficient is updated every iteration rather than looping over features sequentially by default. This (setting to ‘random’) often leads to significantly faster convergence especially when tol is higher than 1e4.
 handlecuml.Handle
If it is None, a new one is created just for this class.
Notes
For additional docs, see scikitlearn’s Lasso.
Examples
import numpy as np import cudf from cuml.linear_model import Lasso ls = Lasso(alpha = 0.1) X = cudf.DataFrame() X['col1'] = np.array([0, 1, 2], dtype = np.float32) X['col2'] = np.array([0, 1, 2], dtype = np.float32) y = cudf.Series( np.array([0.0, 1.0, 2.0], dtype = np.float32) ) result_lasso = ls.fit(X, y) print("Coefficients:") print(result_lasso.coef_) print("intercept:") print(result_lasso.intercept_) X_new = cudf.DataFrame() X_new['col1'] = np.array([3,2], dtype = np.float32) X_new['col2'] = np.array([5,5], dtype = np.float32) preds = result_lasso.predict(X_new) print(preds)
Output:
Coefficients: 0 0.85 1 0.0 Intercept: 0.149999 Preds: 0 2.7 1 1.85
 Attributes
 coef_array, shape (n_features)
The estimated coefficients for the linear regression model.
 intercept_array
The independent term. If fit_intercept_ is False, will be 0.
Methods
fit
(self, X, y[, convert_dtype])Fit the model with X and y.
get_params
(self[, deep])Scikitlearn style function that returns the estimator parameters.
predict
(self, X[, convert_dtype])Predicts the y for X.
set_params
(self, **params)Sklearn style set parameter state to dictionary of params.

fit
(self, X, y, convert_dtype=False)¶ Fit the model with X and y.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the transform method will, when necessary, convert y to be the same data type as X if they differ. This will increase memory used for the method.

get_params
(self, deep=True)¶ Scikitlearn style function that returns the estimator parameters.
 Parameters
 deepboolean (default = True)

predict
(self, X, convert_dtype=False)¶ Predicts the y for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 Returns
 y: cuDF DataFrame
Dense vector (floats or doubles) of shape (n_samples, 1)

set_params
(self, **params)¶ Sklearn style set parameter state to dictionary of params.
 Parameters
 paramsdict of new params
ElasticNet Regression¶

class
cuml.
ElasticNet
¶ ElasticNet extends LinearRegression with combined L1 and L2 regularizations on the coefficients when predicting response y with a linear combination of the predictors in X. It can reduce the variance of the predictors, force some coefficients to be small, and improves the conditioning of the problem.
cuML’s ElasticNet an arraylike object or cuDF DataFrame, uses coordinate descent to fit a linear model.
 Parameters
 alphafloat (default = 1.0)
Constant that multiplies the L1 term. alpha = 0 is equivalent to an ordinary least square, solved by the LinearRegression object. For numerical reasons, using alpha = 0 with the Lasso object is not advised. Given this, you should use the LinearRegression object.
 l1_ratio: float (default = 0.5)
The ElasticNet mixing parameter, with 0 <= l1_ratio <= 1. For l1_ratio = 0 the penalty is an L2 penalty. For l1_ratio = 1 it is an L1 penalty. For 0 < l1_ratio < 1, the penalty is a combination of L1 and L2.
 fit_interceptboolean (default = True)
If True, Lasso tries to correct for the global mean of y. If False, the model expects that you have centered the data.
 normalizeboolean (default = False)
If True, the predictors in X will be normalized by dividing by it’s L2 norm. If False, no scaling will be done.
 max_iterint (default = 1000)
The maximum number of iterations
 tolfloat (default = 1e3)
The tolerance for the optimization: if the updates are smaller than tol, the optimization code checks the dual gap for optimality and continues until it is smaller than tol.
 selection{‘cyclic’, ‘random’} (default=’cyclic’)
If set to ‘random’, a random coefficient is updated every iteration rather than looping over features sequentially by default. This (setting to ‘random’) often leads to significantly faster convergence especially when tol is higher than 1e4.
 handlecuml.Handle
If it is None, a new one is created just for this class.
 output_type(optional) {‘input’, ‘cudf’, ‘cupy’, ‘numpy’} default = None
Use it to control output type of the results and attributes. If None it’ll inherit the output type set at the module level, cuml.output_type. If that has not been changed, by default the estimator will mirror the type of the data used for each fit or predict call. If set, the estimator will override the global option for its behavior.
Notes
For additional docs, see scikitlearn’s ElasticNet.
Examples
import numpy as np import cudf from cuml.linear_model import ElasticNet enet = ElasticNet(alpha = 0.1, l1_ratio=0.5) X = cudf.DataFrame() X['col1'] = np.array([0, 1, 2], dtype = np.float32) X['col2'] = np.array([0, 1, 2], dtype = np.float32) y = cudf.Series( np.array([0.0, 1.0, 2.0], dtype = np.float32) ) result_enet = enet.fit(X, y) print("Coefficients:") print(result_enet.coef_) print("intercept:") print(result_enet.intercept_) X_new = cudf.DataFrame() X_new['col1'] = np.array([3,2], dtype = np.float32) X_new['col2'] = np.array([5,5], dtype = np.float32) preds = result_enet.predict(X_new) print(preds)
Output:
Coefficients: 0 0.448408 1 0.443341 Intercept: 0.1082506 Preds: 0 3.67018 1 3.22177
 Attributes
 coef_array, shape (n_features)
The estimated coefficients for the linear regression model.
 intercept_array
The independent term. If fit_intercept_ is False, will be 0.
Methods
fit
(self, X, y[, convert_dtype])Fit the model with X and y.
get_params
(self[, deep])Scikitlearn style function that returns the estimator parameters.
predict
(self, X[, convert_dtype])Predicts the y for X.
set_params
(self, **params)Sklearn style set parameter state to dictionary of params.

fit
(self, X, y, convert_dtype=False)¶ Fit the model with X and y.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the transform method will, when necessary, convert y to be the same data type as X if they differ. This will increase memory used for the method.

get_params
(self, deep=True)¶ Scikitlearn style function that returns the estimator parameters.
 Parameters
 deepboolean (default = True)

predict
(self, X, convert_dtype=False)¶ Predicts the y for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 y: cuDF DataFrame
Dense vector (floats or doubles) of shape (n_samples, 1)

set_params
(self, **params)¶ Sklearn style set parameter state to dictionary of params.
 Parameters
 paramsdict of new params
Mini Batch SGD Classifier¶

class
cuml.
MBSGDClassifier
¶ Linear models (linear SVM, logistic regression, or linear regression) fitted by minimizing a regularized empirical loss with minibatch SGD.
 Parameters
 loss{‘hinge’, ‘log’, ‘squared_loss’} (default = ‘squared_loss’)
‘hinge’ uses linear SVM
‘log’ uses logistic regression
‘squared_loss’ uses linear regression
 penalty: {‘none’, ‘l1’, ‘l2’, ‘elasticnet’} (default = ‘none’)
‘none’ does not perform any regularization
‘l1’ performs L1 norm (Lasso) which minimizes the sum of the abs value of coefficients
‘l2’ performs L2 norm (Ridge) which minimizes the sum of the square of the coefficients
‘elasticnet’ performs Elastic Net regularization which is a weighted average of L1 and L2 norms
 alpha: float (default = 0.0001)
The constant value which decides the degree of regularization
 fit_interceptboolean (default = True)
If True, the model tries to correct for the global mean of y. If False, the model expects that you have centered the data.
 epochsint (default = 1000)
The number of times the model should iterate through the entire dataset during training (default = 1000)
 tolfloat (default = 1e3)
The training process will stop if current_loss > previous_loss  tol
 shuffleboolean (default = True)
True, shuffles the training data after each epoch False, does not shuffle the training data after each epoch
 eta0float (default = 0.001)
Initial learning rate
 power_tfloat (default = 0.5)
The exponent used for calculating the invscaling learning rate
 learning_rate{‘optimal’, ‘constant’, ‘invscaling’, ‘adaptive’}
(default = ‘constant’)
optimal option will be supported in a future version
constant keeps the learning rate constant
adaptive changes the learning rate if the training loss or the validation accuracy does not improve for n_iter_no_change epochs. The old learning rate is generally divided by 5
 n_iter_no_changeint (default = 5)
the number of epochs to train without any imporvement in the model
 output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’}, optional
Variable to control output type of the results and attributes of the estimators. If None, it’ll inherit the output type set at the module level, cuml.output_type. If set, the estimator will override the global option for its behavior.
Notes
For additional docs, see scikitlearn’s SGDClassifier.
Examples
import numpy as np import cudf from cuml.linear_model import MBSGDClassifier as cumlMBSGDClassifier X = cudf.DataFrame() X['col1'] = np.array([1,1,2,2], dtype = np.float32) X['col2'] = np.array([1,2,2,3], dtype = np.float32) y = cudf.Series(np.array([1, 1, 2, 2], dtype=np.float32)) pred_data = cudf.DataFrame() pred_data['col1'] = np.asarray([3, 2], dtype=np.float32) pred_data['col2'] = np.asarray([5, 5], dtype=np.float32) cu_mbsgd_classifier = cumlMBSGClassifier(learning_rate='constant', eta0=0.05, epochs=2000, fit_intercept=True, batch_size=1, tol=0.0, penalty='l2', loss='squared_loss', alpha=0.5) cu_mbsgd_classifier.fit(X, y) cu_pred = cu_mbsgd_classifier.predict(pred_data).to_array() print(" cuML intercept : ", cu_mbsgd_classifier.intercept_) print(" cuML coef : ", cu_mbsgd_classifier.coef_) print("cuML predictions : ", cu_pred)
Output:
cuML intercept : 0.7150013446807861 cuML coef : 0 0.27320495 1 0.1875956 dtype: float32 cuML predictions : [1. 1.]
Methods
fit
(self, X, y[, convert_dtype])Fit the model with X and y.
get_params
(self[, deep])Scikitlearn style function that returns the estimator parameters.
predict
(self, X[, convert_dtype])Predicts the y for X.
set_params
(self, **params)Sklearn style set parameter state to dictionary of params.

fit
(self, X, y, convert_dtype=False)¶ Fit the model with X and y.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the fit method will, when necessary, convert y to be the same data type as X if they differ. This will increase memory used for the method.

get_params
(self, deep=True)¶ Scikitlearn style function that returns the estimator parameters.
 Parameters
 deepboolean (default = True)

predict
(self, X, convert_dtype=False)¶ Predicts the y for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 y: Type specified by output_type
Dense vector (floats or doubles) of shape (n_samples, 1)

set_params
(self, **params)¶ Sklearn style set parameter state to dictionary of params.
 Parameters
 paramsdict of new params
Mini Batch SGD Regressor¶

class
cuml.
MBSGDRegressor
¶ Linear regression model fitted by minimizing a regularized empirical loss with minibatch SGD.
 Parameters
 loss‘squared_loss’ (default = ‘squared_loss’)
‘squared_loss’ uses linear regression
 penalty: ‘none’, ‘l1’, ‘l2’, ‘elasticnet’ (default = ‘none’)
‘none’ does not perform any regularization ‘l1’ performs L1 norm (Lasso) which minimizes the sum of the abs value of coefficients ‘l2’ performs L2 norm (Ridge) which minimizes the sum of the square of the coefficients ‘elasticnet’ performs Elastic Net regularization which is a weighted average of L1 and L2 norms
 alpha: float (default = 0.0001)
The constant value which decides the degree of regularization
 fit_interceptboolean (default = True)
If True, the model tries to correct for the global mean of y. If False, the model expects that you have centered the data.
 epochsint (default = 1000)
The number of times the model should iterate through the entire dataset during training (default = 1000)
 tolfloat (default = 1e3)
The training process will stop if current_loss > previous_loss  tol
 shuffleboolean (default = True)
True, shuffles the training data after each epoch False, does not shuffle the training data after each epoch
 eta0float (default = 0.001)
Initial learning rate
 power_tfloat (default = 0.5)
The exponent used for calculating the invscaling learning rate
 learning_rate{‘optimal’, ‘constant’, ‘invscaling’, ‘adaptive’}
(default = ‘constant’)
optimal option will be supported in a future version
constant keeps the learning rate constant
adaptive changes the learning rate if the training loss or the validation accuracy does not improve for n_iter_no_change epochs. The old learning rate is generally divided by 5
 n_iter_no_changeint (default = 5)
the number of epochs to train without any imporvement in the model
 output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’}, optional
Variable to control output type of the results and attributes of the estimators. If None, it’ll inherit the output type set at the module level, cuml.output_type. If set, the estimator will override the global option for its behavior.
Notes
For additional docs, see scikitlearn’s SGDRegressor.
Examples
import numpy as np import cudf from cuml.linear_model import MBSGDRegressor as cumlMBSGDRegressor X = cudf.DataFrame() X['col1'] = np.array([1,1,2,2], dtype = np.float32) X['col2'] = np.array([1,2,2,3], dtype = np.float32) y = cudf.Series(np.array([1, 1, 2, 2], dtype=np.float32)) pred_data = cudf.DataFrame() pred_data['col1'] = np.asarray([3, 2], dtype=np.float32) pred_data['col2'] = np.asarray([5, 5], dtype=np.float32) cu_mbsgd_regressor = cumlMBSGDRegressor(learning_rate='constant', eta0=0.05, epochs=2000, fit_intercept=True, batch_size=1, tol=0.0, penalty='l2', loss='squared_loss', alpha=0.5) cu_mbsgd_regressor.fit(X, y) cu_pred = cu_mbsgd_regressor.predict(pred_data).to_array() print(" cuML intercept : ", cu_mbsgd_regressor.intercept_) print(" cuML coef : ", cu_mbsgd_regressor.coef_) print("cuML predictions : ", cu_pred)
Output:
cuML intercept : 0.7150013446807861 cuML coef : 0 0.27320495 1 0.1875956 dtype: float32 cuML predictions : [2.4725943 2.1993892]
Methods
fit
(self, X, y[, convert_dtype])Fit the model with X and y.
get_params
(self[, deep])Scikitlearn style function that returns the estimator parameters.
predict
(self, X[, convert_dtype])Predicts the y for X.
set_params
(self, **params)Sklearn style set parameter state to dictionary of params.

fit
(self, X, y, convert_dtype=False)¶ Fit the model with X and y.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the fit method will, when necessary, convert y to be the same data type as X if they differ. This will increase memory used for the method.

get_params
(self, deep=True)¶ Scikitlearn style function that returns the estimator parameters.
 Parameters
 deepboolean (default = True)

predict
(self, X, convert_dtype=False)¶ Predicts the y for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 y: Type specified by output_type
Dense vector (floats or doubles) of shape (n_samples, 1)

set_params
(self, **params)¶ Sklearn style set parameter state to dictionary of params.
 Parameters
 paramsdict of new params
Stochastic Gradient Descent¶

class
cuml.
SGD
¶ Stochastic Gradient Descent is a very common machine learning algorithm where one optimizes some cost function via gradient steps. This makes SGD very attractive for large problems when the exact solution is hard or even impossible to find.
cuML’s SGD algorithm accepts a numpy matrix or a cuDF DataFrame as the input dataset. The SGD algorithm currently works with linear regression, ridge regression and SVM models.
 Parameters
 loss‘hinge’, ‘log’, ‘squared_loss’ (default = ‘squared_loss’)
‘hinge’ uses linear SVM ‘log’ uses logistic regression ‘squared_loss’ uses linear regression
 penalty: ‘none’, ‘l1’, ‘l2’, ‘elasticnet’ (default = ‘none’)
‘none’ does not perform any regularization ‘l1’ performs L1 norm (Lasso) which minimizes the sum of the abs value of coefficients ‘l2’ performs L2 norm (Ridge) which minimizes the sum of the square of the coefficients ‘elasticnet’ performs Elastic Net regularization which is a weighted average of L1 and L2 norms
 alpha: float (default = 0.0001)
The constant value which decides the degree of regularization
 fit_interceptboolean (default = True)
If True, the model tries to correct for the global mean of y. If False, the model expects that you have centered the data.
 epochsint (default = 1000)
The number of times the model should iterate through the entire dataset during training (default = 1000)
 tolfloat (default = 1e3)
The training process will stop if current_loss > previous_loss  tol
 shuffleboolean (default = True)
True, shuffles the training data after each epoch False, does not shuffle the training data after each epoch
 eta0float (default = 0.001)
Initial learning rate
 power_tfloat (default = 0.5)
The exponent used for calculating the invscaling learning rate
 learning_rate‘optimal’, ‘constant’, ‘invscaling’,
‘adaptive’ (default = ‘constant’)
optimal option supported in the next version constant keeps the learning rate constant adaptive changes the learning rate if the training loss or the validation accuracy does not improve for n_iter_no_change epochs. The old learning rate is generally divide by 5
 n_iter_no_changeint (default = 5)
the number of epochs to train without any imporvement in the model
 output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’}, optional
Variable to control output type of the results and attributes of the estimators. If None, it’ll inherit the output type set at the module level, cuml.output_type. If set, the estimator will override the global option for its behavior.
Examples
import numpy as np import cudf from cuml.solvers import SGD as cumlSGD X = cudf.DataFrame() X['col1'] = np.array([1,1,2,2], dtype = np.float32) X['col2'] = np.array([1,2,2,3], dtype = np.float32) y = cudf.Series(np.array([1, 1, 2, 2], dtype=np.float32)) pred_data = cudf.DataFrame() pred_data['col1'] = np.asarray([3, 2], dtype=dtype) pred_data['col2'] = np.asarray([5, 5], dtype=dtype) cu_sgd = cumlSGD(learning_rate=lrate, eta0=0.005, epochs=2000, fit_intercept=True, batch_size=2, tol=0.0, penalty=penalty, loss=loss) cu_sgd.fit(X, y) cu_pred = cu_sgd.predict(pred_data).to_array() print(" cuML intercept : ", cu_sgd.intercept_) print(" cuML coef : ", cu_sgd.coef_) print("cuML predictions : ", cu_pred)
Output:
cuML intercept : 0.004561662673950195 cuML coef : 0 0.9834546 1 0.010128272 dtype: float32 cuML predictions : [3.0055666 2.0221121]
Methods
fit
(self, X, y[, convert_dtype])Fit the model with X and y.
predict
(self, X[, convert_dtype])Predicts the y for X.
predictClass
(self, X[, convert_dtype])Predicts the y for X.

fit
(self, X, y, convert_dtype=False)¶ Fit the model with X and y.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the fit method will, when necessary, convert y to be the same data type as X if they differ. This will increase memory used for the method.

predict
(self, X, convert_dtype=False)¶ Predicts the y for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 y: Type specified in output_type
Dense vector (floats or doubles) of shape (n_samples, 1)

predictClass
(self, X, convert_dtype=False)¶ Predicts the y for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the predictClass method will automatically convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 yType specified in output_type
Dense vector (floats or doubles) of shape (n_samples, 1)
Random Forest¶

class
cuml.ensemble.
RandomForestClassifier
¶ Implements a Random Forest classifier model which fits multiple decision tree classifiers in an ensemble.
Note that the underlying algorithm for tree node splits differs from that used in scikitlearn. By default, the cuML Random Forest uses a histogrambased algorithms to determine splits, rather than an exact count. You can tune the size of the histograms with the n_bins parameter.
Known Limitations: This is an early release of the cuML Random Forest code. It contains a few known limitations:
GPUbased inference is only supported if the model was trained with 32bit (float32) datatypes. CPUbased inference may be used in this case as a slower fallback.
Very deep / very wide models may exhaust available GPU memory. Future versions of cuML will provide an alternative algorithm to reduce memory consumption.
 Parameters
 n_estimatorsint (default = 100)
Number of trees in the forest. (Default changed to 100 in cuML 0.11)
 handlecuml.Handle
If it is None, a new one is created just for this class.
 split_criterionThe criterion used to split nodes.
0 for GINI, 1 for ENTROPY 2 and 3 not valid for classification (default = 0)
 split_algoint (default = 1)
The algorithm to determine how nodes are split in the tree. 0 for HIST and 1 for GLOBAL_QUANTILE. HIST curently uses a slower treebuilding algorithm so GLOBAL_QUANTILE is recommended for most cases.
 bootstrapboolean (default = True)
Control bootstrapping. If True, each tree in the forest is built on a bootstrapped sample with replacement. If False, sampling without replacement is done.
 bootstrap_featuresboolean (default = False)
Control bootstrapping for features. If features are drawn with or without replacement
 rows_samplefloat (default = 1.0)
Ratio of dataset rows used while fitting each tree.
 max_depthint (default = 16)
Maximum tree depth. Unlimited (i.e, until leaves are pure), if 1. Unlimited depth is not supported. Note that this default differs from scikitlearn’s random forest, which defaults to unlimited depth.
 max_leavesint (default = 1)
Maximum leaf nodes per tree. Soft constraint. Unlimited, if 1.
 max_featuresint, float, or string (default = ‘auto’)
Ratio of number of features (columns) to consider per node split. If int then max_features/n_features. If float then max_features is used as a fraction. If ‘auto’ then max_features=1/sqrt(n_features). If ‘sqrt’ then max_features=1/sqrt(n_features). If ‘log2’ then max_features=log2(n_features)/n_features.
 n_binsint (default = 8)
Number of bins used by the split algorithm.
 min_rows_per_nodeint or float (default = 2)
The minimum number of samples (rows) needed to split a node. If int then number of sample rows. If float the min_rows_per_sample*n_rows
 min_impurity_decreasefloat (default = 0.0)
Minimum decrease in impurity requried for node to be spilt.
 quantile_per_treeboolean (default = False)
Whether quantile is computed for individal trees in RF. Only relevant for GLOBAL_QUANTILE split_algo.
 seedint (default = None)
Seed for the random number generator. Unseeded by default.
Examples
import numpy as np from cuml.ensemble import RandomForestClassifier as cuRFC X = np.random.normal(size=(10,4)).astype(np.float32) y = np.asarray([0,1]*5, dtype=np.int32) cuml_model = cuRFC(max_features=1.0, n_bins=8, n_estimators=40) cuml_model.fit(X,y) cuml_predict = cuml_model.predict(X) print("Predicted labels : ", cuml_predict)
Output:
Predicted labels : [0 1 0 1 0 1 0 1 0 1]
Methods
convert_to_fil_model
(self[, output_class, …])Create a Forest Inference (FIL) model from the trained cuML Random Forest model.
Converts the cuML RF model to a Treelite model
fit
(self, X, y[, convert_dtype])Perform Random Forest Classification on the input data
get_params
(self[, deep])Returns the value of all parameters required to configure this estimator as a dictionary.
predict
(self, X[, predict_model, …])Predicts the labels for X.
predict_proba
(self, X[, output_class, …])Predicts class probabilites for X.
print_detailed
(self)Prints the detailed information about the forest used to train and test the Random Forest model
print_summary
(self)Prints the summary of the forest used to train and test the model
score
(self, X, y[, threshold, algo, …])Calculates the accuracy metric score of the model for X.
set_params
(self, **params)Sets the value of parameters required to configure this estimator, it functions similar to the sklearn set_params.

convert_to_fil_model
(self, output_class=True, threshold=0.5, algo='auto', fil_sparse_format='auto')¶ Create a Forest Inference (FIL) model from the trained cuML Random Forest model.
 Parameters
 output_classboolean (default = True)
This is optional and required only while performing the predict operation on the GPU. If true, return a 1 or 0 depending on whether the raw prediction exceeds the threshold. If False, just return the raw prediction.
 algostring (default = ‘auto’)
This is optional and required only while performing the predict operation on the GPU. ‘naive’  simple inference using shared memory ‘tree_reorg’  similar to naive but trees rearranged to be more coalescingfriendly ‘batch_tree_reorg’  similar to tree_reorg but predicting multiple rows per thread block auto  choose the algorithm automatically. Currently ‘batch_tree_reorg’ is used for dense storage and ‘naive’ for sparse storage
 thresholdfloat (default = 0.5)
Threshold used for classification. Optional and required only while performing the predict operation on the GPU. It is applied if output_class == True, else it is ignored
 fil_sparse_formatboolean or string (default = auto)
This variable is used to choose the type of forest that will be created in the Forest Inference Library. It is not required while using predict_model=’CPU’. ‘auto’  choose the storage type automatically (currently True is chosen by auto) False  create a dense forest True  create a sparse forest, requires algo=’naive’ or algo=’auto’
 Returns
 fil_model :
A Forest Inference model which can be used to perform inferencing on the random forest model.

convert_to_treelite_model
(self)¶ Converts the cuML RF model to a Treelite model
 Returns
 tl_to_fil_modelTreelite version of this model

fit
(self, X, y, convert_dtype=False)¶ Perform Random Forest Classification on the input data
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (int32) of shape (n_samples, 1). Acceptable formats: NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy These labels should be contiguous integers from 0 to n_classes.
 convert_dtypebool, optional (default = False)
When set to True, the fit method will, when necessary, convert y to be the same data type as X if they differ. This will increase memory used for the method.

get_params
(self, deep=True)¶ Returns the value of all parameters required to configure this estimator as a dictionary.
 Parameters
 deepboolean (default = True)

predict
(self, X, predict_model='GPU', output_class=True, threshold=0.5, algo='auto', num_classes=2, convert_dtype=True, fil_sparse_format='auto')¶ Predicts the labels for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 predict_modelString (default = ‘GPU’)
‘GPU’ to predict using the GPU, ‘CPU’ otherwise. The ‘GPU’ can only be used if the model was trained on float32 data and X is float32 or convert_dtype is set to True. Also the ‘GPU’ should only be used for binary classification problems.
 output_classboolean (default = True)
This is optional and required only while performing the predict operation on the GPU. If true, return a 1 or 0 depending on whether the raw prediction exceeds the threshold. If False, just return the raw prediction.
 algostring (default = ‘auto’)
This is optional and required only while performing the predict operation on the GPU. ‘naive’  simple inference using shared memory ‘tree_reorg’  similar to naive but trees rearranged to be more coalescingfriendly ‘batch_tree_reorg’  similar to tree_reorg but predicting multiple rows per thread block auto  choose the algorithm automatically. Currently ‘batch_tree_reorg’ is used for dense storage and ‘naive’ for sparse storage
 thresholdfloat (default = 0.5)
Threshold used for classification. Optional and required only while performing the predict operation on the GPU. It is applied if output_class == True, else it is ignored
 num_classesint (default = 2)
number of different classes present in the dataset
 convert_dtypebool, optional (default = True)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 fil_sparse_formatboolean or string (default = auto)
This variable is used to choose the type of forest that will be created in the Forest Inference Library. It is not required while using predict_model=’CPU’. ‘auto’  choose the storage type automatically (currently True is chosen by auto) False  create a dense forest True  create a sparse forest, requires algo=’naive’ or algo=’auto’
 Returns
 yNumPy
Dense vector (int) of shape (n_samples, 1)

predict_proba
(self, X, output_class=True, threshold=0.5, algo='auto', num_classes=2, convert_dtype=True, fil_sparse_format='auto')¶ Predicts class probabilites for X. This function uses the GPU implementation of predict. Therefore, data with ‘dtype = np.float32’ and ‘num_classes = 2’ should be used while using this function. The option to use predict_proba for multi_class classification is not currently implemented. Please check cuml issue #1679 for more information.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 output_class: boolean (default = True)
This is optional and required only while performing the predict operation on the GPU. If true, return a 1 or 0 depending on whether the raw prediction exceeds the threshold. If False, just return the raw prediction.
 algostring (default = ‘auto’)
This is optional and required only while performing the predict operation on the GPU. ‘naive’  simple inference using shared memory ‘tree_reorg’  similar to naive but trees rearranged to be more coalescingfriendly ‘batch_tree_reorg’  similar to tree_reorg but predicting multiple rows per thread block auto  choose the algorithm automatically. Currently ‘batch_tree_reorg’ is used for dense storage and ‘naive’ for sparse storage
 thresholdfloat (default = 0.5)
Threshold used for classification. Optional and required only while performing the predict operation on the GPU. It is applied if output_class == True, else it is ignored
 num_classesint (default = 2)
number of different classes present in the dataset
 convert_dtypebool, optional (default = True)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 fil_sparse_formatboolean or string (default = auto)
This variable is used to choose the type of forest that will be created in the Forest Inference Library. It is not required while using predict_model=’CPU’. ‘auto’  choose the storage type automatically (currently True is chosen by auto) False  create a dense forest True  create a sparse forest, requires algo=’naive’ or algo=’auto’
 Returns
 y(same as the input datatype)
Dense vector (float) of shape (n_samples, 1). The datatype of y depend on the value of ‘output_type’ varaible specified by the user while intializing the model.

print_detailed
(self)¶ Prints the detailed information about the forest used to train and test the Random Forest model

print_summary
(self)¶ Prints the summary of the forest used to train and test the model

score
(self, X, y, threshold=0.5, algo='auto', num_classes=2, predict_model='GPU', convert_dtype=True, fil_sparse_format='auto')¶ Calculates the accuracy metric score of the model for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yNumPy
Dense vector (int) of shape (n_samples, 1)
 algostring (default = ‘auto’)
This is optional and required only while performing the predict operation on the GPU. ‘naive’  simple inference using shared memory ‘tree_reorg’  similar to naive but trees rearranged to be more coalescingfriendly ‘batch_tree_reorg’  similar to tree_reorg but predicting multiple rows per thread block auto  choose the algorithm automatically. Currently ‘batch_tree_reorg’ is used for dense storage and ‘naive’ for sparse storage
 thresholdfloat
threshold is used to for classification This is optional and required only while performing the predict operation on the GPU.
 num_classesinteger
number of different classes present in the dataset
 convert_dtypeboolean, default=True
whether to convert input data to correct dtype automatically
 predict_modelString (default = ‘GPU’)
‘GPU’ to predict using the GPU, ‘CPU’ otherwise. The ‘GPU’ can only be used if the model was trained on float32 data and X is float32 or convert_dtype is set to True. Also the ‘GPU’ should only be used for binary classification problems.
 fil_sparse_formatboolean or string (default = auto)
This variable is used to choose the type of forest that will be created in the Forest Inference Library. It is not required while using predict_model=’CPU’. ‘auto’  choose the storage type automatically (currently True is chosen by auto) False  create a dense forest True  create a sparse forest, requires algo=’naive’ or algo=’auto’
 Returns
 accuracyfloat
Accuracy of the model [0.0  1.0]

set_params
(self, **params)¶ Sets the value of parameters required to configure this estimator, it functions similar to the sklearn set_params.
 Parameters
 paramsdict of new params

class
cuml.ensemble.
RandomForestRegressor
¶ Implements a Random Forest regressor model which fits multiple decision trees in an ensemble. Note that the underlying algorithm for tree node splits differs from that used in scikitlearn. By default, the cuML Random Forest uses a histogrambased algorithm to determine splits, rather than an exact count. You can tune the size of the histograms with the n_bins parameter.
Known Limitations: This is an early release of the cuML Random Forest code. It contains a few known limitations:
GPUbased inference is only supported if the model was trained with 32bit (float32) datatypes. CPUbased inference may be used in this case as a slower fallback.
Very deep / very wide models may exhaust available GPU memory. Future versions of cuML will provide an alternative algorithm to reduce memory consumption.
 Parameters
 n_estimatorsint (default = 100)
Number of trees in the forest. (Default changed to 100 in cuML 0.11)
 handlecuml.Handle
If it is None, a new one is created just for this class.
 split_algoint (default = 1)
The algorithm to determine how nodes are split in the tree. 0 for HIST and 1 for GLOBAL_QUANTILE. HIST curently uses a slower treebuilding algorithm so GLOBAL_QUANTILE is recommended for most cases.
 split_criterionint (default = 2)
The criterion used to split nodes. 0 for GINI, 1 for ENTROPY, 2 for MSE, or 3 for MAE 0 and 1 not valid for regression
 bootstrapboolean (default = True)
 Control bootstrapping.
If True, each tree in the forest is built on a bootstrapped sample with replacement. If False, sampling without replacement is done.
 bootstrap_featuresboolean (default = False)
Control bootstrapping for features. If features are drawn with or without replacement
 rows_samplefloat (default = 1.0)
Ratio of dataset rows used while fitting each tree.
 max_depthint (default = 16)
Maximum tree depth. Unlimited (i.e, until leaves are pure), if 1. Unlimited depth is not supported with split_algo=1. Note that this default differs from scikitlearn’s random forest, which defaults to unlimited depth.
 max_leavesint (default = 1)
Maximum leaf nodes per tree. Soft constraint. Unlimited, if 1.
 max_featuresint, float, or string (default = ‘auto’)
Ratio of number of features (columns) to consider per node split. If int then max_features/n_features. If float then max_features is used as a fraction. If ‘auto’ then max_features=1.0. If ‘sqrt’ then max_features=1/sqrt(n_features). If ‘log2’ then max_features=log2(n_features)/n_features.
 n_binsint (default = 8)
Number of bins used by the split algorithm.
 min_rows_per_nodeint or float (default = 2)
The minimum number of samples (rows) needed to split a node. If int then number of sample rows If float the min_rows_per_sample*n_rows
 min_impurity_decreasefloat (default = 0.0)
The minimum decrease in impurity required for node to be split
 accuracy_metricstring (default = ‘mse’)
Decides the metric used to evaluate the performance of the model. for median of abs error : ‘median_ae’ for mean of abs error : ‘mean_ae’ for mean square error’ : ‘mse’
 quantile_per_treeboolean (default = False)
Whether quantile is computed for individal trees in RF. Only relevant for GLOBAL_QUANTILE split_algo.
 seedint (default = None)
Seed for the random number generator. Unseeded by default. Does not currently fully guarantee the exact same results.
Examples
import numpy as np from cuml.test.utils import get_handle from cuml.ensemble import RandomForestRegressor as curfc from cuml.test.utils import get_handle X = np.asarray([[0,10],[0,20],[0,30],[0,40]], dtype=np.float32) y = np.asarray([0.0,1.0,2.0,3.0], dtype=np.float32) cuml_model = curfc(max_features=1.0, n_bins=8, split_algo=0, min_rows_per_node=2, n_estimators=40, accuracy_metric='mse') cuml_model.fit(X,y) cuml_score = cuml_model.score(X,y) print("MSE score of cuml : ", cuml_score)
Output:
MSE score of cuml : 0.1123437201231765
Methods
convert_to_fil_model
(self[, output_class, …])Create a Forest Inference (FIL) model from the trained cuML Random Forest model.
Converts the cuML RF model to a Treelite model
fit
(self, X, y[, convert_dtype])Perform Random Forest Regression on the input data
get_params
(self[, deep])Returns the value of all parameters required to configure this estimator as a dictionary.
predict
(self, X[, predict_model, algo, …])Predicts the labels for X.
print_detailed
(self)Prints the detailed information about the forest used to train and test the Random Forest model
print_summary
(self)Prints the summary of the forest used to train and test the model
score
(self, X, y[, algo, convert_dtype, …])Calculates the accuracy metric score of the model for X.
set_params
(self, **params)Sets the value of parameters required to configure this estimator, it functions similar to the sklearn set_params.

convert_to_fil_model
(self, output_class=False, algo='auto', fil_sparse_format='auto')¶ Create a Forest Inference (FIL) model from the trained cuML Random Forest model.
 Parameters
 output_classboolean (default = True)
This is optional and required only while performing the predict operation on the GPU. If true, return a 1 or 0 depending on whether the raw prediction exceeds the threshold. If False, just return the raw prediction.
 algostring (default = ‘auto’)
This is optional and required only while performing the predict operation on the GPU. ‘naive’  simple inference using shared memory ‘tree_reorg’  similar to naive but trees rearranged to be more coalescingfriendly ‘batch_tree_reorg’  similar to tree_reorg but predicting multiple rows per thread block auto  choose the algorithm automatically. Currently ‘batch_tree_reorg’ is used for dense storage and ‘naive’ for sparse storage
 fil_sparse_formatboolean or string (default = auto)
This variable is used to choose the type of forest that will be created in the Forest Inference Library. It is not required while using predict_model=’CPU’. ‘auto’  choose the storage type automatically (currently True is chosen by auto) False  create a dense forest True  create a sparse forest, requires algo=’naive’ or algo=’auto’
 Returns
 fil_model :
A Forest Inference model which can be used to perform inferencing on the random forest model.

convert_to_treelite_model
(self)¶ Converts the cuML RF model to a Treelite model
 Returns
 tl_to_fil_modelTreelite version of this model

fit
(self, X, y, convert_dtype=False)¶ Perform Random Forest Regression on the input data
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (int32) of shape (n_samples, 1). Acceptable formats: NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy These labels should be contiguous integers from 0 to n_classes.

get_params
(self, deep=True)¶ Returns the value of all parameters required to configure this estimator as a dictionary.
 Parameters
 deepboolean (default = True)

predict
(self, X, predict_model='GPU', algo='auto', convert_dtype=True, fil_sparse_format='auto')¶ Predicts the labels for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 predict_modelString (default = ‘GPU’)
‘GPU’ to predict using the GPU, ‘CPU’ otherwise. The GPU can only be used if the model was trained on float32 data and X is float32 or convert_dtype is set to True.
 algostring (default = ‘auto’)
This is optional and required only while performing the predict operation on the GPU. ‘naive’  simple inference using shared memory ‘tree_reorg’  similar to naive but trees rearranged to be more coalescingfriendly ‘batch_tree_reorg’  similar to tree_reorg but predicting multiple rows per thread block auto  choose the algorithm automatically. Currently ‘batch_tree_reorg’ is used for dense storage and ‘naive’ for sparse storage
 convert_dtypebool, optional (default = True)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 fil_sparse_formatboolean or string (default = auto)
This variable is used to choose the type of forest that will be created in the Forest Inference Library. It is not required while using predict_model=’CPU’. ‘auto’  choose the storage type automatically (currently True is chosen by auto) False  create a dense forest True  create a sparse forest, requires algo=’naive’ or algo=’auto’
 Returns
 yNumPy
Dense vector (int) of shape (n_samples, 1)

print_detailed
(self)¶ Prints the detailed information about the forest used to train and test the Random Forest model

print_summary
(self)¶ Prints the summary of the forest used to train and test the model

score
(self, X, y, algo='auto', convert_dtype=True, fil_sparse_format='auto', predict_model='GPU')¶ Calculates the accuracy metric score of the model for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yNumPy
Dense vector (int) of shape (n_samples, 1)
 algostring (default = ‘auto’)
This is optional and required only while performing the predict operation on the GPU. ‘naive’  simple inference using shared memory ‘tree_reorg’  similar to naive but trees rearranged to be more coalescingfriendly ‘batch_tree_reorg’  similar to tree_reorg but predicting multiple rows per thread block auto  choose the algorithm automatically. Currently ‘batch_tree_reorg’ is used for dense storage and ‘naive’ for sparse storage
 convert_dtypeboolean, default=True
whether to convert input data to correct dtype automatically
 predict_modelString (default = ‘GPU’)
‘GPU’ to predict using the GPU, ‘CPU’ otherwise. The GPU can only be used if the model was trained on float32 data and X is float32 or convert_dtype is set to True.
 fil_sparse_formatboolean or string (default = auto)
This variable is used to choose the type of forest that will be created in the Forest Inference Library. It is not required while using predict_model=’CPU’. ‘auto’  choose the storage type automatically (currently True is chosen by auto) False  create a dense forest True  create a sparse forest, requires algo=’naive’ or algo=’auto’
 Returns
 mean_square_errorfloat or
 median_abs_errorfloat or
 mean_abs_errorfloat

set_params
(self, **params)¶ Sets the value of parameters required to configure this estimator, it functions similar to the sklearn set_params.
 Parameters
 paramsdict of new params
Forest Inferencing¶

class
cuml.
ForestInference
¶ ForestInference provides GPUaccelerated inference (prediction) for random forest and boosted decision tree models.
This module does not support training models. Rather, users should train a model in another package and save it in a treelitecompatible format. (See https://github.com/dmlc/treelite) Currently, LightGBM, XGBoost and SKLearn GBDT and random forest models are supported.
Users typically create a ForestInference object by loading a saved model file with ForestInference.load. It is also possible to create it from an SKLearn model using ForestInference.load_from_sklearn. The resulting object provides a predict method for carrying out inference.
 Known limitations:
A single row of data should fit into the shared memory of a thread block, which means that more than 12288 features are not supported.
From sklearn.ensemble, only {RandomForest,GradientBoosting}{Classifier,Regressor} models are supported; other sklearn.ensemble models are currently not supported.
Importing large SKLearn models can be slow, as it is done in Python.
LightGBM categorical features are not supported.
Inference uses a dense matrix format, which is efficient for many problems but can be suboptimal for sparse datasets.
Only binary classification and regression are supported.
 Parameters
 handlecuml.Handle
If it is None, a new one is created just for this class.
Notes
For additional usage examples, see the sample notebook at https://github.com/rapidsai/cuml/blob/branch0.14/notebooks/forest_inference_demo.ipynb
Examples
In the example below, synthetic data is copied to the host before inference. ForestInference can also accept a numpy array directly at the cost of a slight performance overhead.
# Assume that the file 'xgb.model' contains a classifier model that was # previously saved by XGBoost's save_model function. import sklearn, sklearn.datasets, numpy as np from numba import cuda from cuml import ForestInference model_path = 'xgb.model' X_test, y_test = sklearn.datasets.make_classification() X_gpu = cuda.to_device(np.ascontiguousarray(X_test.astype(np.float32))) fm = ForestInference.load(model_path, output_class=True) fil_preds_gpu = fm.predict(X_gpu) accuracy_score = sklearn.metrics.accuracy_score(y_test, np.asarray(fil_preds_gpu))
Methods
load
(filename[, output_class, threshold, …])Returns a FIL instance containing the forest saved in ‘filename’ This uses Treelite to load the saved model.
load_from_sklearn
(skl_model[, output_class, …])Creates a FIL model using the scikitlearn model passed to the function.
load_from_treelite_model
(self, model[, …])Creates a FIL model using the treelite model passed to the function.
load_using_treelite_handle
(self, model_handle)Returns a FIL instance by converting a treelite model to FIL model by using the treelite ModelHandle passed.
predict
(self, X[, preds])Predicts the labels for X with the loaded forest model.
predict_proba
(self, X[, preds])Predicts the class probabilities for X with the loaded forest model.

static
load
(filename, output_class=False, threshold=0.5, algo='auto', storage_type='auto', model_type='xgboost', handle=None)¶ Returns a FIL instance containing the forest saved in ‘filename’ This uses Treelite to load the saved model.
 Parameters
 filenamestring
Path to saved model file in a treelitecompatible format (See https://treelite.readthedocs.io/en/latest/treeliteapi.html
for more information)
 output_classbool (default=False)
If True, return a 1 or 0 depending on whether the raw prediction exceeds the threshold. If False, just return the raw prediction.
 thresholdfloat (default=0.5)
Cutoff value above which a prediction is set to 1.0 Only used if the model is classification and output_class is True
 algostring (default=’auto’)
Which inference algorithm to use. See documentation in FIL.load_from_treelite_model
 storage_typestring (default=’auto’)
Inmemory storage format to be used for the FIL model. See documentation in FIL.load_from_treelite_model
 model_typestring (default=”xgboost”)
Format of the saved treelite model to be load. It can be ‘xgboost’, ‘lightgbm’, or ‘protobuf’.
 Returns
 fil_model :
A Forest Inference model which can be used to perform inferencing on the model read from the file.

static
load_from_sklearn
(skl_model, output_class=False, threshold=0.5, algo='auto', storage_type='auto', handle=None)¶ Creates a FIL model using the scikitlearn model passed to the function. This function requires Treelite 0.90 to be installed.
 Parameters
 skl_modelThe scikitlearn model from which to build the FIL version.
 output_class: boolean (default=False)
If True, return a 1 or 0 depending on whether the raw prediction exceeds the threshold. If False, just return the raw prediction.
 algostring (default=’auto’)
 name of the algo from (from algo_t enum)
 ‘AUTO’ or ‘auto’  choose the algorithm automatically;
currently ‘BATCH_TREE_REORG’ is used for dense storage, and ‘NAIVE’ for sparse storage
‘NAIVE’ or ‘naive’  simple inference using shared memory ‘TREE_REORG’ or ‘tree_reorg’  similar to naive but trees
rearranged to be more coalescingfriendly
 ‘BATCH_TREE_REORG’ or ‘batch_tree_reorg’  similar to TREE_REORG
but predicting multiple rows per thread block
 thresholdfloat (default=0.5)
Threshold is used to for classification. It is applied only if output_class == True, else it is ignored.
 storage_typestring (default=’auto’)
 Inmemory storage format to be used for the FIL model.
 ‘AUTO’ or ‘auto’  choose the storage type automatically
(currently DENSE is always used)
‘DENSE’ or ‘dense’  create a dense forest ‘SPARSE’ or ‘sparse’  create a sparse forest;
requires algo=’NAIVE’ or algo=’AUTO’.
 Returns
 fil_model :
A Forest Inference model created from the scikitlearn model passed.

load_from_treelite_model
(self, model, output_class=False, algo='auto', threshold=0.5, storage_type='auto')¶ Creates a FIL model using the treelite model passed to the function.
 Parameters
 modelthe trained model information in the treelite format
loaded from a saved model using the treelite API https://treelite.readthedocs.io/en/latest/treeliteapi.html
 output_class: boolean (default=False)
If True, return a 1 or 0 depending on whether the raw prediction exceeds the threshold. If False, just return the raw prediction.
 algostring (default=’auto’)
 name of the algo from (from algo_t enum)
 ‘AUTO’ or ‘auto’  choose the algorithm automatically;
currently ‘BATCH_TREE_REORG’ is used for dense storage, and ‘NAIVE’ for sparse storage
‘NAIVE’ or ‘naive’  simple inference using shared memory ‘TREE_REORG’ or ‘tree_reorg’  similar to naive but trees
rearranged to be more coalescingfriendly
 ‘BATCH_TREE_REORG’ or ‘batch_tree_reorg’  similar to TREE_REORG
but predicting multiple rows per thread block
 thresholdfloat (default=0.5)
Threshold is used to for classification. It is applied only if output_class == True, else it is ignored.
 storage_typestring (default=’auto’)
 Inmemory storage format to be used for the FIL model.
 ‘AUTO’ or ‘auto’  choose the storage type automatically
(currently DENSE is always used)
‘DENSE’ or ‘dense’  create a dense forest ‘SPARSE’ or ‘sparse’  create a sparse forest;
requires algo=’NAIVE’ or algo=’AUTO’
 Returns
 fil_model :
A Forest Inference model which can be used to perform inferencing on the random forest/ XGBoost model.

load_using_treelite_handle
(self, model_handle, output_class=False, algo='auto', storage_type='auto', threshold=0.5)¶ Returns a FIL instance by converting a treelite model to FIL model by using the treelite ModelHandle passed.
 Parameters
 model_handleModelhandle to the treelite forest model
(See https://treelite.readthedocs.io/en/latest/treeliteapi.html for more information)
 output_classbool (default=False)
If True, return a 1 or 0 depending on whether the raw prediction exceeds the threshold. If False, just return the raw prediction.
 thresholdfloat (default=0.5)
Cutoff value above which a prediction is set to 1.0 Only used if the model is classification and output_class is True
 algostring (default=’auto’)
Which inference algorithm to use. See documentation in FIL.load_from_treelite_model
 storage_typestring (default=’auto’)
Inmemory storage format to be used for the FIL model. See documentation in FIL.load_from_treelite_model
 Returns
 fil_model :
A Forest Inference model which can be used to perform inferencing on the random forest model.

predict
(self, X, preds=None)¶ Predicts the labels for X with the loaded forest model. By default, the result is the raw floating point output from the model, unless output_class was set to True during model loading.
See the documentation of ForestInference.load for details.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy For optimal performance, pass a device array with Cstyle layout
 preds: gpuarray or cudf.Series, shape = (n_samples,)
Optional ‘out’ location to store inference results
 Returns
 GPU array of length n_samples with inference results
 (or ‘preds’ filled with inference results if preds was specified)

predict_proba
(self, X, preds=None)¶ Predicts the class probabilities for X with the loaded forest model. The result is the raw floating point output from the model.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy For optimal performance, pass a device array with Cstyle layout
 preds: gpuarray or cudf.Series, shape = (n_samples,2)
binary probability output Optional ‘out’ location to store inference results
 Returns
 GPU array of shape (n_samples,2) with inference results
 (or ‘preds’ filled with inference results if preds was specified)
Coordinate Descent¶

class
cuml.
CD
¶ Coordinate Descent (CD) is a very common optimization algorithm that minimizes along coordinate directions to find the minimum of a function.
cuML’s CD algorithm accepts a numpy matrix or a cuDF DataFrame as the input dataset.algorithm The CD algorithm currently works with linear regression and ridge, lasso, and elasticnet penalties.
 Parameters
 loss‘squared_loss’ (Only ‘squared_loss’ is supported right now)
‘squared_loss’ uses linear regression
 alpha: float (default = 0.0001)
The constant value which decides the degree of regularization. ‘alpha = 0’ is equivalent to an ordinary least square, solved by the LinearRegression object.
 l1_ratio: float (default = 0.15)
The ElasticNet mixing parameter, with 0 <= l1_ratio <= 1. For l1_ratio = 0 the penalty is an L2 penalty. For l1_ratio = 1 it is an L1 penalty. For 0 < l1_ratio < 1, the penalty is a combination of L1 and L2.
 fit_interceptboolean (default = True)
If True, the model tries to correct for the global mean of y. If False, the model expects that you have centered the data.
 max_iterint (default = 1000)
The number of times the model should iterate through the entire dataset during training (default = 1000)
 tolfloat (default = 1e3)
The tolerance for the optimization: if the updates are smaller than tol, solver stops.
 shuffleboolean (default = True)
If set to ‘True’, a random coefficient is updated every iteration rather than looping over features sequentially by default. This (setting to ‘True’) often leads to significantly faster convergence especially when tol is higher than 1e4.
Examples
import numpy as np import cudf from cuml.solvers import CD as cumlCD cd = cumlCD(alpha=0.0) X = cudf.DataFrame() X['col1'] = np.array([1,1,2,2], dtype = np.float32) X['col2'] = np.array([1,2,2,3], dtype = np.float32) y = cudf.Series( np.array([6.0, 8.0, 9.0, 11.0], dtype = np.float32) ) reg = cd.fit(X,y) print("Coefficients:") print(reg.coef_) print("intercept:") print(reg.intercept_) X_new = cudf.DataFrame() X_new['col1'] = np.array([3,2], dtype = np.float32) X_new['col2'] = np.array([5,5], dtype = np.float32) preds = cd.predict(X_new) print(preds)
Output:
Coefficients: 0 1.0019531 1 1.9980469 Intercept: 3.0 Preds: 0 15.997 1 14.995
Methods
fit
(self, X, y[, convert_dtype])Fit the model with X and y.
predict
(self, X[, convert_dtype])Predicts the y for X.

fit
(self, X, y, convert_dtype=False)¶ Fit the model with X and y.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the fit method will, when necessary, convert y to be the same data type as X if they differ. This will increase memory used for the method.

predict
(self, X, convert_dtype=False)¶ Predicts the y for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 y: cuDF DataFrame
Dense vector (floats or doubles) of shape (n_samples, 1)
QuasiNewton¶

class
cuml.
QN
¶ QuasiNewton methods are used to either find zeroes or local maxima and minima of functions, and used by this class to optimize a cost function.
Two algorithms are implemented underneath cuML’s QN class, and which one is executed depends on the following rule:
OrthantWise Limited Memory QuasiNewton (OWLQN) if there is l1 regularization
Limited Memory BFGS (LBFGS) otherwise.
cuML’s QN class can take arraylike objects, either in host as NumPy arrays or in device (as Numba or __cuda_array_interface__ compliant).
 Parameters
 loss: ‘sigmoid’, ‘softmax’, ‘squared_loss’ (default = ‘squared_loss’)
‘sigmoid’ loss used for single class logistic regression ‘softmax’ loss used for multiclass logistic regression ‘normal’ used for normal/square loss
 fit_intercept: boolean (default = True)
If True, the model tries to correct for the global mean of y. If False, the model expects that you have centered the data.
 l1_strength: float (default = 0.0)
l1 regularization strength (if nonzero, will run OWLQN, else LBFGS). Note, that as in Scikitlearn, the bias will not be regularized.
 l2_strength: float (default = 0.0)
l2 regularization strength. Note, that as in Scikitlearn, the bias will not be regularized.
 max_iter: int (default = 1000)
Maximum number of iterations taken for the solvers to converge.
 tol: float (default = 1e3)
The training process will stop if current_loss > previous_loss  tol
 linesearch_max_iter: int (default = 50)
Max number of linesearch iterations per outer iteration of the algorithm.
 lbfgs_memory: int (default = 5)
Rank of the lbfgs inverseHessian approximation. Method will use O(lbfgs_memory * D) memory.
 verboseint or boolean (default = False)
Controls verbose level of logging.
Notes
This class contains implementations of two popular QuasiNewton methods:
Limitedmemory Broyden Fletcher Goldfarb Shanno (LBFGS) [Nocedal, Wright  Numerical Optimization (1999)]
Orthantwise limitedmemory quasinewton (OWLQN) [Andrew, Gao  ICML 2007] <https://www.microsoft.com/enus/research/publication/scalabletrainingofl1regularizedloglinearmodels/>
Examples
import cudf import numpy as np # Both import methods supported # from cuml import QN from cuml.solvers import QN X = cudf.DataFrame() X['col1'] = np.array([1,1,2,2], dtype = np.float32) X['col2'] = np.array([1,2,2,3], dtype = np.float32) y = cudf.Series( np.array([0.0, 0.0, 1.0, 1.0], dtype = np.float32) ) solver = QN() solver.fit(X,y) # Note: for now, the coefficients also include the intercept in the # last position if fit_intercept=True print("Coefficients:") print(solver.coef_.copy_to_host()) print("Intercept:") print(solver.intercept_.copy_to_host()) X_new = cudf.DataFrame() X_new['col1'] = np.array([1,5], dtype = np.float32) X_new['col2'] = np.array([2,5], dtype = np.float32) preds = solver.predict(X_new) print("Predictions:") print(preds)
Output:
Coefficients: 10.647417 0.3267412 17.158297 Intercept: 17.158297 Predictions: 0 0.0 1 1.0
 Attributes
 coef_array, shape (n_classes, n_features)
The estimated coefficients for the linear regression model. Note: shape is (n_classes, n_features + 1) if fit_intercept = True.
 intercept_array (n_classes, 1)
The independent term. If fit_intercept_ is False, will be 0.
Methods
fit
(self, X, y[, convert_dtype])Fit the model with X and y.
get_param_names
(self)predict
(self, X[, convert_dtype])Predicts the y for X.
score
(self, X, y)
fit
(self, X, y, convert_dtype=False)¶ Fit the model with X and y.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the fit method will, when necessary, convert y to be the same data type as X if they differ. This will increase memory used for the method.

get_param_names
(self)¶

predict
(self, X, convert_dtype=False)¶ Predicts the y for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 y: cuDF DataFrame
Dense vector (floats or doubles) of shape (n_samples, 1)

score
(self, X, y)¶
Support Vector Machines¶

class
cuml.svm.
SVC
(CSupport Vector Classification)¶ Construct an SVC classifier for training and predictions.
 Parameters
 handlecuml.Handle
If it is None, a new one is created for this class
 Cfloat (default = 1.0)
Penalty parameter C
 kernelstring (default=’rbf’)
Specifies the kernel function. Possible options: ‘linear’, ‘poly’, ‘rbf’, ‘sigmoid’. Currently precomputed kernels are not supported.
 degreeint (default=3)
Degree of polynomial kernel function.
 gammafloat or string (default = ‘scale’)
Coefficient for rbf, poly, and sigmoid kernels. You can specify the numeric value, or use one of the following options:  ‘auto’: gamma will be set to 1 / n_features  ‘scale’: gamma will be se to 1 / (n_features * X.var())
 coef0float (default = 0.0)
Independent term in kernel function, only signifficant for poly and sigmoid
 tolfloat (default = 1e3)
Tolerance for stopping criterion.
 cache_sizefloat (default = 200.0)
Size of the kernel cache during training in MiB. The default is a conservative value, increase it to improve the training time, at the cost of higher memory footprint. After training the kernel cache is deallocated. During prediction, we also need a temporary space to store kernel matrix elements (this can be signifficant if n_support is large). The cache_size variable sets an upper limit to the prediction buffer as well.
 max_iterint (default = 100*n_samples)
Limit the number of outer iterations in the solver
 nochange_stepsint (default = 1000)
We monitor how much our stopping criteria changes during outer iterations. If it does not change (changes less then 1e3*tol) for nochange_steps consecutive steps, then we stop training.
 verboseint or boolean (default = False)
verbosity level
Notes
The solver uses the SMO method to fit the classifier. We use the Optimized Hierarchical Decomposition [1] variant of the SMO algorithm, similar to [2]
References
[1] J. Vanek et al. A GPUArchitecture Optimized Hierarchical Decomposition Algorithm for Support VectorMachine Training, IEEE Transactions on Parallel and Distributed Systems, vol 28, no 12, 3330, (2017)
[2] Z. Wen et al. ThunderSVM: A Fast SVM Library on GPUs and CPUs, Journal of Machine Learning Research, 19, 15 (2018) https://github.com/XtraComputing/thundersvm
Examples
import numpy as np from cuml.svm import SVC X = np.array([[1,1], [2,1], [1,2], [2,2], [1,3], [2,3]], dtype=np.float32); y = np.array([1, 1, 1, 1, 1, 1], dtype=np.float32) clf = SVC(kernel='poly', degree=2, gamma='auto', C=1) clf.fit(X, y) print("Predicted labels:", clf.predict(X))
Output:
Predicted labels: [1. 1. 1. 1. 1. 1.]
 Attributes
 n_support_int
The total number of support vectors. Note: this will change in the future to represent number support vectors for each class (like in Sklearn, see https://github.com/rapidsai/cuml/issues/956 )
 support_int, shape = (n_support)
Device array of suppurt vector indices
 support_vectors_float, shape (n_support, n_cols)
Device array of support vectors
 dual_coef_float, shape = (1, n_support)
Device array of coefficients for support vectors
 intercept_int
The constant in the decision function
 fit_status_int
0 if SVM is correctly fitted
coef_
float, shape (1, n_cols)SVMBase.coef_(self)
 For additional docs, see `scikitlearn’s SVC
 <https://scikitlearn.org/stable/modules/generated/sklearn.svm.SVC.html>`_.
Methods
decision_function
(self, X)Calculates the decision function values for X.
fit
(self, X, y)Fit the model with X and y.
predict
(self, X)Predicts the class labels for X.

decision_function
(self, X)¶ Calculates the decision function values for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 Returns
 ycuDF Series
Dense vector (floats or doubles) of shape (n_samples, 1)

fit
(self, X, y)¶ Fit the model with X and y.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy

predict
(self, X)¶ Predicts the class labels for X. The returned y values are the class labels associated to sign(decision_function(X)).
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 Returns
 ycuDF Series
Dense vector (floats or doubles) of shape (n_samples, 1)

class
cuml.svm.
SVR
(Epsilon Support Vector Regression)¶ Construct an SVC classifier for training and predictions.
 Parameters
 handlecuml.Handle
If it is None, a new one is created for this class
 Cfloat (default = 1.0)
Penalty parameter C
 kernelstring (default=’rbf’)
Specifies the kernel function. Possible options: ‘linear’, ‘poly’, ‘rbf’, ‘sigmoid’. Currently precomputed kernels are not supported.
 degreeint (default=3)
Degree of polynomial kernel function.
 gammafloat or string (default = ‘scale’)
Coefficient for rbf, poly, and sigmoid kernels. You can specify the numeric value, or use one of the following options:  ‘auto’: gamma will be set to 1 / n_features  ‘scale’: gamma will be se to 1 / (n_features * X.var())
 coef0float (default = 0.0)
Independent term in kernel function, only signifficant for poly and sigmoid
 tolfloat (default = 1e3)
Tolerance for stopping criterion.
 epsilon: float (default = 0.1)
epsilon parameter of the epsironSVR model. There is no penalty associated to points that are predicted within the epsilontube around the target values.
 cache_sizefloat (default = 200 MiB)
Size of the kernel cache during training in MiB. The default is a conservative value, increase it to improve the training time, at the cost of higher memory footprint. After training the kernel cache is deallocated. During prediction, we also need a temporary space to store kernel matrix elements (this can be signifficant if n_support is large). The cache_size variable sets an upper limit to the prediction buffer as well.
 max_iterint (default = 100*n_samples)
Limit the number of outer iterations in the solver
 nochange_stepsint (default = 1000)
We monitor how much our stopping criteria changes during outer iterations. If it does not change (changes less then 1e3*tol) for nochange_steps consecutive steps, then we stop training.
 verboseint or boolean (default = False)
verbosity level
Notes
For additional docs, see Scikitlearn’s SVR.
The solver uses the SMO method to fit the regressor. We use the Optimized Hierarchical Decomposition [1] variant of the SMO algorithm, similar to [2]
References
 [1] J. Vanek et al. A GPUArchitecture Optimized Hierarchical Decomposition
Algorithm for Support VectorMachine Training, IEEE Transactions on Parallel and Distributed Systems, vol 28, no 12, 3330, (2017)
[2] Z. Wen et al. ThunderSVM: A Fast SVM Library on GPUs and CPUs, Journal * of Machine Learning Research, 19, 15 (2018)
Examples
import numpy as np from cuml.svm import SVR X = np.array([[1], [2], [3], [4], [5]], dtype=np.float32) y = np.array([1.1, 4, 5, 3.9, 1.], dtype = np.float32) reg = SVR(kernel='rbf', gamma='scale', C=10, epsilon=0.1) reg.fit(X, y) print("Predicted values:", reg.predict(X))
Output:
Predicted values: [1.200474 3.8999617 5.100488 3.7995374 1.0995375]
 Attributes
 n_support_int
The total number of support vectors. Note: this will change in the future to represent number support vectors for each class (like in Sklearn, see Issue #956)
 support_int, shape = [n_support]
Device array of suppurt vector indices
 support_vectors_float, shape [n_support, n_cols]
Device array of support vectors
 dual_coef_float, shape = [1, n_support]
Device array of coefficients for support vectors
 intercept_int
The constant in the decision function
 fit_status_int
0 if SVM is correctly fitted
coef_
float, shape [1, n_cols]SVMBase.coef_(self)
Methods
fit
(self, X, y)Fit the model with X and y.
predict
(self, X)Predicts the values for X.
score
(self, X, y)Return R^2 score of the prediction.

fit
(self, X, y)¶ Fit the model with X and y.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy

predict
(self, X)¶ Predicts the values for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 Returns
 ycuDF Series
Dense vector (floats or doubles) of shape (n_samples, 1)

score
(self, X, y)¶ Return R^2 score of the prediction.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of target values. Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 Returns
 score: float R^2 score
Nearest Neighbors Classification¶

class
cuml.neighbors.
KNeighborsClassifier
¶ KNearest Neighbors Classifier is an instancebased learning technique, that keeps training samples around for prediction, rather than trying to learn a generalizable set of model parameters.
 Parameters
 n_neighborsint (default=5)
Default number of neighbors to query
 verboseint or boolean (default = False)
Logging level
 handlecumlHandle
The cumlHandle resources to use
 algorithmstring (default=’brute’)
The query algorithm to use. Currently, only ‘brute’ is supported.
 metricstring (default=’euclidean’).
Distance metric to use.
 weightsstring (default=’uniform’)
Sample weights to use. Currently, only the uniform strategy is supported.
Notes
For additional docs, see scikitlearn’s KNeighborsClassifier.
Examples
from cuml.neighbors import KNeighborsClassifier from sklearn.datasets import make_blobs from sklearn.model_selection import train_test_split X, y = make_blobs(n_samples=100, centers=5, n_features=10) knn = KNeighborsClassifier(n_neighbors=10) X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.80) knn.fit(X_train, y_train) knn.predict(X_test)
Methods
fit
(self, X, y[, convert_dtype])Fit a GPU index for knearest neighbors classifier model.
get_param_names
(self)predict
(self, X[, convert_dtype])Use the trained knearest neighbors classifier to predict the labels for X
predict_proba
(self, X[, convert_dtype])Use the trained knearest neighbors classifier to predict the label probabilities for X
score
(self, X, y[, convert_dtype])Compute the accuracy score using the given labels and the trained knearest neighbors classifier to predict the classes for X.

fit
(self, X, y, convert_dtype=True)¶ Fit a GPU index for knearest neighbors classifier model.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, n_outputs)
Dense matrix (floats or doubles) of shape (n_samples, n_outputs). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit method will automatically convert the inputs to np.float32.

get_param_names
(self)¶

predict
(self, X, convert_dtype=True)¶ Use the trained knearest neighbors classifier to predict the labels for X
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit method will automatically convert the inputs to np.float32.

predict_proba
(self, X, convert_dtype=True)¶ Use the trained knearest neighbors classifier to predict the label probabilities for X
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit method will automatically convert the inputs to np.float32.

score
(self, X, y, convert_dtype=True)¶ Compute the accuracy score using the given labels and the trained knearest neighbors classifier to predict the classes for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit method will automatically convert the inputs to np.float32.
Nearest Neighbors Regression¶

class
cuml.neighbors.
KNeighborsRegressor
¶ KNearest Neighbors Regressor is an instancebased learning technique, that keeps training samples around for prediction, rather than trying to learn a generalizable set of model parameters.
The KNearest Neighbors Regressor will compute the average of the labels for the k closest neighbors and use it as the label.
 Parameters
 n_neighborsint (default=5)
Default number of neighbors to query
 verboseint or boolean (default = False)
Logging level
 handlecumlHandle
The cumlHandle resources to use
 algorithmstring (default=’brute’)
The query algorithm to use. Currently, only ‘brute’ is supported.
 metricstring (default=’euclidean’).
Distance metric to use.
 weightsstring (default=’uniform’)
Sample weights to use. Currently, only the uniform strategy is supported.
Notes
For additional docs, see scikitlearn’s KNeighborsClassifier.
Examples
from cuml.neighbors import KNeighborsRegressor from sklearn.datasets import make_blobs from sklearn.model_selection import train_test_split X, y = make_blobs(n_samples=100, centers=5, n_features=10) knn = KNeighborsRegressor(n_neighbors=10) X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.80) knn.fit(X_train, y_train) knn.predict(X_test)
Output:
array([3. , 1. , 1. , 3.79999995, 2. , 0. , 3.79999995, 3.79999995, 3.79999995, 0. , 3.79999995, 0. , 1. , 2. , 3. , 1. , 0. , 0. , 0. , 2. , 3. , 3. , 0. , 3. , 3.79999995, 3.79999995, 3.79999995, 3.79999995, 3. , 2. , 3.79999995, 3.79999995, 0. ])
Methods
fit
(self, X, y[, convert_dtype])Fit a GPU index for knearest neighbors regression model.
predict
(self, X[, convert_dtype])Use the trained knearest neighbors regression model to predict the labels for X
score
(self, X, y[, convert_dtype])Fit a GPU index for knearest neighbors regression model.

fit
(self, X, y, convert_dtype=True)¶ Fit a GPU index for knearest neighbors regression model.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, n_outputs)
Dense matrix (floats or doubles) of shape (n_samples, n_outputs). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit method will automatically convert the inputs to np.float32.

predict
(self, X, convert_dtype=True)¶ Use the trained knearest neighbors regression model to predict the labels for X
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit method will automatically convert the inputs to np.float32.

score
(self, X, y, convert_dtype=True)¶ Fit a GPU index for knearest neighbors regression model.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit method will automatically convert the inputs to np.float32.
Clustering¶
KMeans Clustering¶

class
cuml.
KMeans
¶ KMeans is a basic but powerful clustering method which is optimized via Expectation Maximization. It randomly selects K data points in X, and computes which samples are close to these points. For every cluster of points, a mean is computed (hence the name), and this becomes the new centroid.
cuML’s KMeans expects an arraylike object or cuDF DataFrame, and supports the scalable KMeans++ initialization method. This method is more stable than randomly selecting K points.
 Parameters
 handlecuml.Handle
If it is None, a new one is created just for this class.
 n_clustersint (default = 8)
The number of centroids or clusters you want.
 max_iterint (default = 300)
The more iterations of EM, the more accurate, but slower.
 tolfloat64 (default = 1e4)
Stopping criterion when centroid means do not change much.
 verboseint or boolean (default = False)
Logging level.
 random_stateint (default = 1)
If you want results to be the same when you restart Python, select a state.
 init‘scalablekmeans++’, ‘kmeans’ , ‘random’ or an ndarray (default = ‘scalablekmeans++’) # noqa
‘scalablekmeans++’ or ‘kmeans’: Uses fast and stable scalable kmeans++ initialization. ‘random’: Choose ‘n_cluster’ observations (rows) at random from data for the initial centroids. If an ndarray is passed, it should be of shape (n_clusters, n_features) and gives the initial centers.
 n_init: int (default = 1)
Number of instances the kmeans algorithm will be called with different seeds. The final results will be from the instance that produces lowest inertia out of n_init instances.
 oversampling_factorfloat64
scalable kmeans oversampling factor
 max_samples_per_batchint (default=1<<15)
maximum number of samples to use for each batch of the pairwise distance computation.
 oversampling_factorint (default = 2)
The amount of points to sample in scalable kmeans++ initialization for potential centroids. Increasing this value can lead to better initial centroids at the cost of memory. The total number of centroids sampled in scalable kmeans++ is oversampling_factor * n_clusters * 8.
 max_samples_per_batchint (default = 32768)
The number of data samples to use for batches of the pairwise distance computation. This computation is done throughout both fit predict. The default should suit most cases. The total number of elements in the batched pairwise distance computation is max_samples_per_batch * n_clusters. It might become necessary to lower this number when n_clusters becomes prohibitively large.
Notes
KMeans requires n_clusters to be specified. This means one needs to approximately guess or know how many clusters a dataset has. If one is not sure, one can start with a small number of clusters, and visualize the resulting clusters with PCA, UMAP or TSNE, and verify that they look appropriate.
Applications of KMeans
The biggest advantage of KMeans is its speed and simplicity. That is why KMeans is many practitioner’s first choice of a clustering algorithm. KMeans has been extensively used when the number of clusters is approximately known, such as in big data clustering tasks, image segmentation and medical clustering.
For additional docs, see scikitlearn’s Kmeans.
Examples
# Both import methods supported from cuml import KMeans from cuml.cluster import KMeans import cudf import numpy as np import pandas as pd def np2cudf(df): # convert numpy array to cuDF dataframe df = pd.DataFrame({'fea%d'%i:df[:,i] for i in range(df.shape[1])}) pdf = cudf.DataFrame() for c,column in enumerate(df): pdf[str(c)] = df[column] return pdf a = np.asarray([[1.0, 1.0], [1.0, 2.0], [3.0, 2.0], [4.0, 3.0]], dtype=np.float32) b = np2cudf(a) print("input:") print(b) print("Calling fit") kmeans_float = KMeans(n_clusters=2, n_gpu=1) kmeans_float.fit(b) print("labels:") print(kmeans_float.labels_) print("cluster_centers:") print(kmeans_float.cluster_centers_)
Output:
input: 0 1 0 1.0 1.0 1 1.0 2.0 2 3.0 2.0 3 4.0 3.0 Calling fit labels: 0 0 1 0 2 1 3 1 cluster_centers: 0 1 0 1.0 1.5 1 3.5 2.5
 Attributes
 cluster_centers_array
The coordinates of the final clusters. This represents of “mean” of each data cluster.
 labels_array
Which cluster each datapoint belongs to.
Methods
fit
(self, X[, sample_weight])Compute kmeans clustering with X.
fit_predict
(self, X[, sample_weight])Compute cluster centers and predict cluster index for each sample.
fit_transform
(self, X[, convert_dtype])Compute clustering and transform X to clusterdistance space.
get_param_names
(self)predict
(self, X[, convert_dtype, sample_weight])Predict the closest cluster each sample in X belongs to.
score
(self, X)Opposite of the value of X on the Kmeans objective.
transform
(self, X[, convert_dtype])Transform X to a clusterdistance space.

fit
(self, X, sample_weight=None)¶ Compute kmeans clustering with X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 sample_weightarraylike (device or host) shape = (n_samples,), default=None # noqa
The weights for each observation in X. If None, all observations are assigned equal weight.

fit_predict
(self, X, sample_weight=None)¶ Compute cluster centers and predict cluster index for each sample.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 sample_weightarraylike (device or host) shape = (n_samples,), default=None # noqa
The weights for each observation in X. If None, all observations are assigned equal weight.

fit_transform
(self, X, convert_dtype=False)¶ Compute clustering and transform X to clusterdistance space.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the fit_transform method will automatically convert the input to the data type which was used to train the model. This will increase memory used for the method.

get_param_names
(self)¶

predict
(self, X, convert_dtype=False, sample_weight=None)¶ Predict the closest cluster each sample in X belongs to.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 Returns
 labelsarray
 Which cluster each datapoint belongs to.

score
(self, X)¶ Opposite of the value of X on the Kmeans objective.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 Returns
 score: float
Opposite of the value of X on the Kmeans objective.

transform
(self, X, convert_dtype=False)¶ Transform X to a clusterdistance space.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the transform method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
DBSCAN¶

class
cuml.
DBSCAN
¶ DBSCAN is a very powerful yet fast clustering technique that finds clusters where data is concentrated. This allows DBSCAN to generalize to many problems if the datapoints tend to congregate in larger groups.
cuML’s DBSCAN expects an arraylike object or cuDF DataFrame, and constructs an adjacency graph to compute the distances between close neighbours.
 Parameters
 epsfloat (default = 0.5)
The maximum distance between 2 points such they reside in the same neighborhood.
 handlecuml.Handle
If it is None, a new one is created just for this class
 min_samplesint (default = 5)
The number of samples in a neighborhood such that this group can be considered as an important core point (including the point itself).
 verboseint or boolean (default = False)
Logging level
 max_mbytes_per_batch(optional) int64
Calculate batch size using no more than this number of megabytes for the pairwise distance computation. This enables the tradeoff between runtime and memory usage for making the N^2 pairwise distance computations more tractable for large numbers of samples. If you are experiencing out of memory errors when running DBSCAN, you can set this value based on the memory size of your device. Note: this option does not set the maximum total memory used in the DBSCAN computation and so this value will not be able to be set to the total memory available on the device.
 output_type(optional) {‘input’, ‘cudf’, ‘cupy’, ‘numpy’} default = None
Use it to control output type of the results and attributes. If None it’ll inherit the output type set at the module level, cuml.output_type. If that has not been changed, by default the estimator will mirror the type of the data used for each fit or predict call. If set, the estimator will override the global option for its behavior.
Notes
DBSCAN is very sensitive to the distance metric it is used with, and a large assumption is that datapoints need to be concentrated in groups for clusters to be constructed.
Applications of DBSCAN
DBSCAN’s main benefit is that the number of clusters is not a hyperparameter, and that it can find nonlinearly shaped clusters. This also allows DBSCAN to be robust to noise. DBSCAN has been applied to analyzing particle collisions in the Large Hadron Collider, customer segmentation in marketing analyses, and much more.
For additional docs, see scikitlearn’s DBSCAN.
Examples
# Both import methods supported from cuml import DBSCAN from cuml.cluster import DBSCAN import cudf import numpy as np gdf_float = cudf.DataFrame() gdf_float['0'] = np.asarray([1.0,2.0,5.0], dtype = np.float32) gdf_float['1'] = np.asarray([4.0,2.0,1.0], dtype = np.float32) gdf_float['2'] = np.asarray([4.0,2.0,1.0], dtype = np.float32) dbscan_float = DBSCAN(eps = 1.0, min_samples = 1) dbscan_float.fit(gdf_float) print(dbscan_float.labels_)
Output:
0 0 1 1 2 2
 Attributes
 labels_arraylike or cuDF series
Which cluster each datapoint belongs to. Noisy samples are labeled as 1. Format depends on cuml global output type and estimator output_type.
Methods
fit
(self, X[, out_dtype])Perform DBSCAN clustering from features.
fit_predict
(self, X[, out_dtype])Performs clustering on input_gdf and returns cluster labels.
get_param_names
(self)
fit
(self, X, out_dtype='int32')¶ Perform DBSCAN clustering from features.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 out_dtype: dtype Determines the precision of the output labels array.
default: “int32”. Valid values are { “int32”, np.int32, “int64”, np.int64}. When the number of samples exceed

fit_predict
(self, X, out_dtype='int32')¶ Performs clustering on input_gdf and returns cluster labels.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features) Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 Returns
 ycuDF Series, shape (n_samples)
cluster labels

get_param_names
(self)¶
Dimensionality Reduction and Manifold Learning¶
Principal Component Analysis¶

class
cuml.
PCA
¶ PCA (Principal Component Analysis) is a fundamental dimensionality reduction technique used to combine features in X in linear combinations such that each new component captures the most information or variance of the data. N_components is usually small, say at 3, where it can be used for data visualization, data compression and exploratory analysis.
cuML’s PCA expects an arraylike object or cuDF DataFrame, and provides 2 algorithms Full and Jacobi. Full (default) uses a full eigendecomposition then selects the top K eigenvectors. The Jacobi algorithm is much faster as it iteratively tries to correct the top K eigenvectors, but might be less accurate.
 Parameters
 copyboolean (default = True)
If True, then copies data then removes mean from data. False might cause data to be overwritten with its mean centered version.
 handlecuml.Handle
If it is None, a new one is created just for this class
 iterated_powerint (default = 15)
Used in Jacobi solver. The more iterations, the more accurate, but slower.
 n_componentsint (default = 1)
The number of top K singular vectors / values you want. Must be <= number(columns).
 random_stateint / None (default = None)
If you want results to be the same when you restart Python, select a state.
 svd_solver‘full’ or ‘jacobi’ or ‘auto’ (default = ‘full’)
Full uses a eigendecomposition of the covariance matrix then discards components. Jacobi is much faster as it iteratively corrects, but is less accurate.
 tolfloat (default = 1e7)
Used if algorithm = “jacobi”. Smaller tolerance can increase accuracy, but but will slow down the algorithm’s convergence.
 verboseint or boolean (default = False)
Logging level
 whitenboolean (default = False)
If True, decorrelates the components. This is done by dividing them by the corresponding singular values then multiplying by sqrt(n_samples). Whitening allows each component to have unit variance and removes multicollinearity. It might be beneficial for downstream tasks like LinearRegression where correlated features cause problems.
Notes
PCA considers linear combinations of features, specifically those that maximize global variance structure. This means PCA is fantastic for global structure analyses, but weak for local relationships. Consider UMAP or TSNE for a locally important embedding.
Applications of PCA
PCA is used extensively in practice for data visualization and data compression. It has been used to visualize extremely large word embeddings like Word2Vec and GloVe in 2 or 3 dimensions, large datasets of everyday objects and images, and used to distinguish between cancerous cells from healthy cells.
For additional docs, see scikitlearn’s PCA.
Examples
# Both import methods supported from cuml import PCA from cuml.decomposition import PCA import cudf import numpy as np gdf_float = cudf.DataFrame() gdf_float['0'] = np.asarray([1.0,2.0,5.0], dtype = np.float32) gdf_float['1'] = np.asarray([4.0,2.0,1.0], dtype = np.float32) gdf_float['2'] = np.asarray([4.0,2.0,1.0], dtype = np.float32) pca_float = PCA(n_components = 2) pca_float.fit(gdf_float) print(f'components: {pca_float.components_}') print(f'explained variance: {pca_float._explained_variance_}') exp_var = pca_float._explained_variance_ratio_ print(f'explained variance ratio: {exp_var}') print(f'singular values: {pca_float._singular_values_}') print(f'mean: {pca_float._mean_}') print(f'noise variance: {pca_float._noise_variance_}') trans_gdf_float = pca_float.transform(gdf_float) print(f'Inverse: {trans_gdf_float}') input_gdf_float = pca_float.inverse_transform(trans_gdf_float) print(f'Input: {input_gdf_float}')
Output:
components: 0 1 2 0 0.69225764 0.5102837 0.51028395 1 0.72165036 0.48949987 0.4895003 explained variance: 0 8.510402 1 0.48959687 explained variance ratio: 0 0.9456003 1 0.054399658 singular values: 0 4.1256275 1 0.9895422 mean: 0 2.6666667 1 2.3333333 2 2.3333333 noise variance: 0 0.0 transformed matrix: 0 1 0 2.8547091 0.42891636 1 0.121316016 0.80743366 2 2.9760244 0.37851727 Input Matrix: 0 1 2 0 1.0000001 3.9999993 4.0 1 2.0 2.0000002 1.9999999 2 4.9999995 1.0000006 1.0
 Attributes
 components_array
The top K components (VT.T[:,:n_components]) in U, S, VT = svd(X)
 explained_variance_array
How much each component explains the variance in the data given by S**2
 explained_variance_ratio_array
How much in % the variance is explained given by S**2/sum(S**2)
 singular_values_array
The top K singular values. Remember all singular values >= 0
 mean_array
The column wise mean of X. Used to mean  center the data first.
 noise_variance_float
From Bishop 1999’s Textbook. Used in later tasks like calculating the estimated covariance of X.
Methods
fit
(self, X[, y])Fit the model with X.
fit_transform
(self, X[, y])Fit the model with X and apply the dimensionality reduction on X.
get_param_names
(self)inverse_transform
(self, X[, convert_dtype])Transform data back to its original space.
transform
(self, X[, convert_dtype])Apply dimensionality reduction to X.

fit
(self, X, y=None)¶ Fit the model with X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yignored
 Returns
 cluster labels

fit_transform
(self, X, y=None)¶ Fit the model with X and apply the dimensionality reduction on X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
training data (floats or doubles), where n_samples is the number of samples, and n_features is the number of features. Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yignored
 Returns
 X_newcuDF DataFrame, shape (n_samples, n_components)

get_param_names
(self)¶

inverse_transform
(self, X, convert_dtype=False)¶ Transform data back to its original space.
In other words, return an input X_original whose transform would be X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
New data (floats or doubles), where n_samples is the number of samples and n_components is the number of components. Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the inverse_transform method will automatically convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 X_originalcuDF DataFrame, shape (n_samples, n_features)

transform
(self, X, convert_dtype=False)¶ Apply dimensionality reduction to X.
X is projected on the first principal components previously extracted from a training set.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
New data (floats or doubles), where n_samples is the number of samples and n_components is the number of components. Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the transform method will automatically convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 X_newcuDF DataFrame, shape (n_samples, n_components)
Truncated SVD¶

class
cuml.
TruncatedSVD
¶ TruncatedSVD is used to compute the top K singular values and vectors of a large matrix X. It is much faster when n_components is small, such as in the use of PCA when 3 components is used for 3D visualization.
cuML’s TruncatedSVD an arraylike object or cuDF DataFrame, and provides 2 algorithms Full and Jacobi. Full (default) uses a full eigendecomposition then selects the top K singular vectors. The Jacobi algorithm is much faster as it iteratively tries to correct the top K singular vectors, but might be less accurate.
 Parameters
 algorithm‘full’ or ‘jacobi’ or ‘auto’ (default = ‘full’)
Full uses a eigendecomposition of the covariance matrix then discards components. Jacobi is much faster as it iteratively corrects, but is less accurate.
 handlecuml.Handle
If it is None, a new one is created just for this class
 n_componentsint (default = 1)
The number of top K singular vectors / values you want. Must be <= number(columns).
 n_iterint (default = 15)
Used in Jacobi solver. The more iterations, the more accurate, but slower.
 random_stateint / None (default = None)
If you want results to be the same when you restart Python, select a state.
 tolfloat (default = 1e7)
Used if algorithm = “jacobi”. Smaller tolerance can increase accuracy, but but will slow down the algorithm’s convergence.
 verboseint or boolean (default = False)
Logging level
Notes
TruncatedSVD (the randomized version [Jacobi]) is fantastic when the number of components you want is much smaller than the number of features. The approximation to the largest singular values and vectors is very robust, however, this method loses a lot of accuracy when you want many, many components.
Applications of TruncatedSVD
TruncatedSVD is also known as Latent Semantic Indexing (LSI) which tries to find topics of a word count matrix. If X previously was centered with mean removal, TruncatedSVD is the same as TruncatedPCA. TruncatedSVD is also used in information retrieval tasks, recommendation systems and data compression.
For additional documentation, see scikitlearn’s TruncatedSVD docs.
Examples
# Both import methods supported from cuml import TruncatedSVD from cuml.decomposition import TruncatedSVD import cudf import numpy as np gdf_float = cudf.DataFrame() gdf_float['0'] = np.asarray([1.0,2.0,5.0], dtype = np.float32) gdf_float['1'] = np.asarray([4.0,2.0,1.0], dtype = np.float32) gdf_float['2'] = np.asarray([4.0,2.0,1.0], dtype = np.float32) tsvd_float = TruncatedSVD(n_components = 2, algorithm = "jacobi", n_iter = 20, tol = 1e9) tsvd_float.fit(gdf_float) print(f'components: {tsvd_float.components_}') print(f'explained variance: {tsvd_float._explained_variance_}') exp_var = tsvd_float._explained_variance_ratio_ print(f'explained variance ratio: {exp_var}') print(f'singular values: {tsvd_float._singular_values_}') trans_gdf_float = tsvd_float.transform(gdf_float) print(f'Transformed matrix: {trans_gdf_float}') input_gdf_float = tsvd_float.inverse_transform(trans_gdf_float) print(f'Input matrix: {input_gdf_float}')
Output:
components: 0 1 2 0 0.58725953 0.57233137 0.5723314 1 0.80939883 0.41525528 0.4152552 explained variance: 0 55.33908 1 16.660923 explained variance ratio: 0 0.7685983 1 0.23140171 singular values: 0 7.439024 1 4.0817795 Transformed Matrix: 0 1 2 0 5.1659107 2.512643 1 3.4638448 0.042223275 2 4.0809603 3.2164836 Input matrix: 0 1 2 0 1.0 4.000001 4.000001 1 2.0000005 2.0000005 2.0000007 2 5.000001 0.9999999 1.0000004
 Attributes
 components_array
The top K components (VT.T[:,:n_components]) in U, S, VT = svd(X)
 explained_variance_array
How much each component explains the variance in the data given by S**2
 explained_variance_ratio_array
How much in % the variance is explained given by S**2/sum(S**2)
 singular_values_array
The top K singular values. Remember all singular values >= 0
Methods
fit
(self, X[, y])Fit LSI model on training cudf DataFrame X.
fit_transform
(self, X[, y])Fit LSI model to X and perform dimensionality reduction on X.
get_param_names
(self)inverse_transform
(self, X[, convert_dtype])Transform X back to its original space.
transform
(self, X[, convert_dtype])Perform dimensionality reduction on X.

fit
(self, X, y=None)¶ Fit LSI model on training cudf DataFrame X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
y : ignored

fit_transform
(self, X, y=None)¶ Fit LSI model to X and perform dimensionality reduction on X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yignored
 Returns
 X_newcuDF DataFrame, shape (n_samples, n_components)
Reduced version of X as a dense cuDF DataFrame

get_param_names
(self)¶

inverse_transform
(self, X, convert_dtype=False)¶ Transform X back to its original space. Returns a cuDF DataFrame X_original whose transform would be X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the inverse_transform method will automatically convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 X_originalcuDF DataFrame, shape (n_samples, n_features)
Note that this is always a dense cuDF DataFrame.

transform
(self, X, convert_dtype=False)¶ Perform dimensionality reduction on X. Parameters ——— X : arraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the transform method will automatically convert the input to the data type which was used to train the model.
 Returns
 X_newcuDF DataFrame, shape (n_samples, n_components)
Reduced version of X. This will always be a dense DataFrame.
UMAP¶

class
cuml.
UMAP
¶ Uniform Manifold Approximation and Projection Finds a low dimensional embedding of the data that approximates an underlying manifold.
Adapted from https://github.com/lmcinnes/umap/blob/master/umap/umap.py
 Parameters
 n_neighbors: float (optional, default 15)
The size of local neighborhood (in terms of number of neighboring sample points) used for manifold approximation. Larger values result in more global views of the manifold, while smaller values result in more local data being preserved. In general values should be in the range 2 to 100.
 n_components: int (optional, default 2)
The dimension of the space to embed into. This defaults to 2 to provide easy visualization, but can reasonably be set to any
 n_epochs: int (optional, default None)
The number of training epochs to be used in optimizing the low dimensional embedding. Larger values result in more accurate embeddings. If None is specified a value will be selected based on the size of the input dataset (200 for large datasets, 500 for small).
 learning_rate: float (optional, default 1.0)
The initial learning rate for the embedding optimization.
 init: string (optional, default ‘spectral’)
 How to initialize the low dimensional embedding. Options are:
‘spectral’: use a spectral embedding of the fuzzy 1skeleton
‘random’: assign initial embedding positions at random.
 min_dist: float (optional, default 0.1)
The effective minimum distance between embedded points. Smaller values will result in a more clustered/clumped embedding where nearby points on the manifold are drawn closer together, while larger values will result on a more even dispersal of points. The value should be set relative to the
spread
value, which determines the scale at which embedded points will be spread out. spread: float (optional, default 1.0)
The effective scale of embedded points. In combination with
min_dist
this determines how clustered/clumped the embedded points are. set_op_mix_ratio: float (optional, default 1.0)
Interpolate between (fuzzy) union and intersection as the set operation used to combine local fuzzy simplicial sets to obtain a global fuzzy simplicial sets. Both fuzzy set operations use the product tnorm. The value of this parameter should be between 0.0 and 1.0; a value of 1.0 will use a pure fuzzy union, while 0.0 will use a pure fuzzy intersection.
 local_connectivity: int (optional, default 1)
The local connectivity required – i.e. the number of nearest neighbors that should be assumed to be connected at a local level. The higher this value the more connected the manifold becomes locally. In practice this should be not more than the local intrinsic dimension of the manifold.
 repulsion_strength: float (optional, default 1.0)
Weighting applied to negative samples in low dimensional embedding optimization. Values higher than one will result in greater weight being given to negative samples.
 negative_sample_rate: int (optional, default 5)
The number of negative samples to select per positive sample in the optimization process. Increasing this value will result in greater repulsive force being applied, greater optimization cost, but slightly more accuracy.
 transform_queue_size: float (optional, default 4.0)
For transform operations (embedding new points using a trained model_ this will control how aggressively to search for nearest neighbors. Larger values will result in slower performance but more accurate nearest neighbor evaluation.
 a: float (optional, default None)
More specific parameters controlling the embedding. If None these values are set automatically as determined by
min_dist
andspread
. b: float (optional, default None)
More specific parameters controlling the embedding. If None these values are set automatically as determined by
min_dist
andspread
. hash_input: UMAP can hash the training input so that exact embeddings
are returned when transform is called on the same data upon which the model was trained. This enables consistent behavior between calling model.fit_transform(X) and calling model.fit(X).transform(X). Not that the CPUbased UMAP reference implementation does this by default. This feature is made optional in the GPU version due to the significant overhead in copying memory to the host for computing the hash. (default = False)
 random_stateint, RandomState instance or None, optional (default=None)
random_state is the seed used by the random number generator during embedding initialization and during sampling used by the optimizer. Note: Unfortunately, achieving a high amount of parallelism during the optimization stage often comes at the expense of determinism, since many floatingpoint additions are being made in parallel without a deterministic ordering. This causes slightly different results across training sessions, even when the same seed is used for random number generation. Setting a random_state will enable consistency of trained embeddings, allowing for reproducible results to 3 digits of precision, but will do so at the expense of potentially slower training and increased memory usage.
 optim_batch_size: int (optional, default 100000 / n_components)
Used to maintain the consistency of embeddings for large datasets. The optimization step will be processed with at most optim_batch_size edges at once preventing inconsistencies. A lower batch size will yield more consistently repeatable embeddings at the cost of speed.
 callback: An instance of GraphBasedDimRedCallback class to intercept
the internal state of embeddings while they are being trained. Example of callback usage:
from cuml.internals import GraphBasedDimRedCallback class CustomCallback(GraphBasedDimRedCallback):
 def on_preprocess_end(self, embeddings):
print(embeddings.copy_to_host())
 def on_epoch_end(self, embeddings):
print(embeddings.copy_to_host())
 def on_train_end(self, embeddings):
print(embeddings.copy_to_host())
 verboseint or boolean (default = False)
Controls verbosity of logging.
Notes
This module is heavily based on Leland McInnes’ reference UMAP package. However, there are a number of differences and features that are not yet implemented in cuml.umap:
Using a nonEuclidean distance metric (support for a fixed set of nonEuclidean metrics is planned for an upcoming release).
Using a precomputed pairwise distance matrix (under consideration for future releases)
Manual initialization of initial embedding positions
In addition to these missing features, you should expect to see the final embeddings differing between cuml.umap and the reference UMAP. In particular, the reference UMAP uses an approximate kNN algorithm for large data sizes while cuml.umap always uses exact kNN.
Known issue: If a UMAP model has not yet been fit, it cannot be pickled. However, after fitting, a UMAP mode.
References
Leland McInnes, John Healy, James Melville UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction https://arxiv.org/abs/1802.03426
Methods
find_ab_params
(spread, min_dist)Function taken from UMAPlearn : https://github.com/lmcinnes/umap Fit a, b params for the differentiable curve used in lower dimensional fuzzy simplicial complex construction.
fit
(self, X[, y, convert_dtype, knn_graph])Fit X into an embedded space.
fit_transform
(self, X[, y, convert_dtype, …])Fit X into an embedded space and return that transformed output.
transform
(self, X[, convert_dtype, knn_graph])Transform X into the existing embedded space and return that transformed output.
validate_hyperparams
(self)
static
find_ab_params
(spread, min_dist)¶ Function taken from UMAPlearn : https://github.com/lmcinnes/umap Fit a, b params for the differentiable curve used in lower dimensional fuzzy simplicial complex construction. We want the smooth curve (from a predefined family with simple gradient) that best matches an offset exponential decay.

fit
(self, X, y=None, convert_dtype=True, knn_graph=None)¶ Fit X into an embedded space.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
X contains a sample per row. Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, 1)
y contains a label per row. Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 knn_graphsparse arraylike (device or host)
shape=(n_samples, n_samples) A sparse array containing the knearest neighbors of X, where the columns are the nearest neighbor indices for each row and the values are their distances. It’s important that k>=n_neighbors, so that UMAP can model the neighbors from this graph, instead of building its own internally. Users using the knn_graph parameter provide UMAP with their own run of the KNN algorithm. This allows the user to pick a custom distance function (sometimes useful on certain datasets) whereas UMAP uses euclidean by default. The custom distance function should match the metric used to train UMAP embeedings. Storing and reusing a knn_graph will also provide a speedup to the UMAP algorithm when performing a grid search. Acceptable formats: sparse SciPy ndarray, CuPy device ndarray, CSR/COO preferred other formats will go through conversion to CSR

fit_transform
(self, X, y=None, convert_dtype=True, knn_graph=None)¶ Fit X into an embedded space and return that transformed output.
There is a subtle difference between calling fit_transform(X) and calling fit().transform(). Calling fit_transform(X) will train the embeddings on X and return the embeddings. Calling fit(X).transform(X) will train the embeddings on X and then run a second optimization return the embedding after it is trained while calling
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
X contains a sample per row. Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 knn_graphsparse arraylike (device or host)
shape=(n_samples, n_samples) A sparse array containing the knearest neighbors of X, where the columns are the nearest neighbor indices for each row and the values are their distances. It’s important that k>=n_neighbors, so that UMAP can model the neighbors from this graph, instead of building its own internally. Users using the knn_graph parameter provide UMAP with their own run of the KNN algorithm. This allows the user to pick a custom distance function (sometimes useful on certain datasets) whereas UMAP uses euclidean by default. The custom distance function should match the metric used to train UMAP embeedings. Storing and reusing a knn_graph will also provide a speedup to the UMAP algorithm when performing a grid search. Acceptable formats: sparse SciPy ndarray, CuPy device ndarray, CSR/COO preferred other formats will go through conversion to CSR
 Returns
 X_newarray, shape (n_samples, n_components)
Embedding of the training data in lowdimensional space.

transform
(self, X, convert_dtype=True, knn_graph=None)¶ Transform X into the existing embedded space and return that transformed output.
Please refer to the reference UMAP implementation for information on the differences between fit_transform() and running fit() transform().
Specifically, the transform() function is stochastic: https://github.com/lmcinnes/umap/issues/158
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
New data to be transformed. Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 knn_graphsparse arraylike (device or host)
shape=(n_samples, n_samples) A sparse array containing the knearest neighbors of X, where the columns are the nearest neighbor indices for each row and the values are their distances. It’s important that k>=n_neighbors, so that UMAP can model the neighbors from this graph, instead of building its own internally. Users using the knn_graph parameter provide UMAP with their own run of the KNN algorithm. This allows the user to pick a custom distance function (sometimes useful on certain datasets) whereas UMAP uses euclidean by default. The custom distance function should match the metric used to train UMAP embeedings. Storing and reusing a knn_graph will also provide a speedup to the UMAP algorithm when performing a grid search. Acceptable formats: sparse SciPy ndarray, CuPy device ndarray, CSR/COO preferred other formats will go through conversion to CSR
 Returns
 X_newarray, shape (n_samples, n_components)
Embedding of the new data in lowdimensional space.

validate_hyperparams
(self)¶
Random Projections¶

class
cuml.random_projection.
GaussianRandomProjection
¶ Gaussian Random Projection method derivated from BaseRandomProjection class.
Random projection is a dimensionality reduction technique. Random projection methods are powerful methods known for their simplicity, computational efficiency and restricted model size. This algorithm also has the advantage to preserve distances well between any two samples and is thus suitable for methods having this requirement.
The components of the random matrix are drawn from N(0, 1 / n_components).
 Parameters
 handlecuml.Handle
If it is None, a new one is created just for this class
 n_componentsint (default = ‘auto’)
Dimensionality of the target projection space. If set to ‘auto’, the parameter is deducted thanks to Johnson–Lindenstrauss lemma. The automatic deduction make use of the number of samples and the eps parameter.
The Johnson–Lindenstrauss lemma can produce very conservative n_components parameter as it makes no assumption on dataset structure.
 epsfloat (default = 0.1)
Error tolerance during projection. Used by Johnson–Lindenstrauss automatic deduction when n_components is set to ‘auto’.
 random_stateint (default = None)
Seed used to initilize random generator
Notes
Inspired by Scikitlearn’s implementation : https://scikitlearn.org/stable/modules/random_projection.html
 Attributes
 gaussian_methodboolean
To be passed to base class in order to determine random matrix generation method

class
cuml.random_projection.
SparseRandomProjection
¶ Sparse Random Projection method derivated from BaseRandomProjection class.
Random projection is a dimensionality reduction technique. Random projection methods are powerful methods known for their simplicity, computational efficiency and restricted model size. This algorithm also has the advantage to preserve distances well between any two samples and is thus suitable for methods having this requirement.
Sparse random matrix is an alternative to dense random projection matrix (e.g. Gaussian) that guarantees similar embedding quality while being much more memory efficient and allowing faster computation of the projected data (with sparse enough matrices). If we note ‘s = 1 / density’ the components of the random matrix are drawn from:
sqrt(s) / sqrt(n_components) with probability 1 / 2s
0 with probability 1  1 / s
+sqrt(s) / sqrt(n_components) with probability 1 / 2s
 Parameters
 handlecuml.Handle
If it is None, a new one is created just for this class
 n_componentsint (default = ‘auto’)
Dimensionality of the target projection space. If set to ‘auto’, the parameter is deducted thanks to Johnson–Lindenstrauss lemma. The automatic deduction make use of the number of samples and the eps parameter.
The Johnson–Lindenstrauss lemma can produce very conservative n_components parameter as it makes no assumption on dataset structure.
 densityfloat in range (0, 1] (default = ‘auto’)
Ratio of nonzero component in the random projection matrix.
If density = ‘auto’, the value is set to the minimum density as recommended by Ping Li et al.: 1 / sqrt(n_features).
 epsfloat (default = 0.1)
Error tolerance during projection. Used by Johnson–Lindenstrauss automatic deduction when n_components is set to ‘auto’.
 dense_outputboolean (default = True)
If set to True transformed matrix will be dense otherwise sparse.
 random_stateint (default = None)
Seed used to initilize random generator
Notes
Inspired by Scikitlearn’s implementation : https://scikitlearn.org/stable/modules/random_projection.html
 Attributes
 gaussian_methodboolean
To be passed to base class in order to determine random matrix generation method

random_projection.
johnson_lindenstrauss_min_dim
(n_samples, eps=0.1)¶ In mathematics, the Johnson–Lindenstrauss lemma states that highdimensional data can be embedded into lower dimension while preserving the distances.
With p the random projection : (1  eps) u  v^2 < p(u)  p(v)^2 < (1 + eps) u  v^2
This function finds the minimum number of components to guarantee that the embedding is inside the eps error tolerance.
 Parameters
 n_samplesint
Number of samples.
 epsfloat in (0,1) (default = 0.1)
Maximum distortion rate as defined by the JohnsonLindenstrauss lemma.
 Returns
 n_componentsint
The minimal number of components to guarantee with good probability an epsembedding with n_samples.
TSNE¶

class
cuml.
TSNE
¶ TSNE (TDistributed Stochastic Neighbor Embedding) is an extremely powerful dimensionality reduction technique that aims to maintain local distances between data points. It is extremely robust to whatever dataset you give it, and is used in many areas including cancer research, music analysis and neural network weight visualizations.
Currently, cuML’s TSNE supports the fast Barnes Hut O(NlogN) TSNE approximation (derived from CannyLabs’ BH open source CUDA code). This allows TSNE to produce extremely fast embeddings when n_components = 2. cuML defaults to this algorithm. A slower but more accurate Exact algorithm is also provided.
 Parameters
 n_componentsint (default 2)
The output dimensionality size. Currently only size=2 is tested, but the ‘exact’ algorithm will support greater dimensionality in future.
 perplexityfloat (default 30.0)
Larger datasets require a larger value. Consider choosing different perplexity values from 5 to 50 and see the output differences.
 early_exaggerationfloat (default 12.0)
Controls the space between clusters. Not critical to tune this.
 learning_ratefloat (default 200.0)
The learning rate usually between (10, 1000). If this is too high, TSNE could look like a cloud / ball of points.
 n_iterint (default 1000)
The more epochs, the more stable/accurate the final embedding.
 n_iter_without_progressint (default 300)
When the KL Divergence becomes too small after some iterations, terminate TSNE early.
 min_grad_normfloat (default 1e07)
The minimum gradient norm for when TSNE will terminate early.
 metricstr ‘euclidean’ only (default ‘euclidean’)
Currently only supports euclidean distance. Will support cosine in a future release.
 initstr ‘random’ (default ‘random’)
Currently supports random intialization.
 verboseint or boolean (default = False) (default logger.level_info)
Level of verbosity. Most messages will be printed inside the Python Console.
 random_stateint (default None)
Setting this can allow future runs of TSNE to look mostly the same. It is known that TSNE tends to have vastly different outputs on many runs. Try using PCA intialization (upcoming with change #1098) to possibly counteract this problem. It is known that small perturbations can directly change the result of the embedding for parallel TSNE implementations.
 methodstr ‘barnes_hut’ or ‘exact’ (default ‘barnes_hut’)
Options are either barnes_hut or exact. It is recommended that you use the barnes hut approximation for superior O(nlogn) complexity.
 anglefloat (default 0.5)
Tradeoff between accuracy and speed. Choose between (0,2 0.8) where closer to one indicates full accuracy but slower speeds.
 learning_rate_methodstr ‘adaptive’, ‘none’ or None (default ‘adaptive’)
Either adaptive or None. Uses a special adpative method that tunes the learning rate, early exaggeration and perplexity automatically based on input size.
 n_neighborsint (default 90)
The number of datapoints you want to use in the attractive forces. Smaller values are better for preserving local structure, whilst larger values can improve global structure preservation. Default is 3 * 30 (perplexity)
 perplexity_max_iterint (default 100)
The number of epochs the best gaussian bands are found for.
 exaggeration_iterint (default 250)
To promote the growth of clusters, set this higher.
 pre_momentumfloat (default 0.5)
During the exaggeration iteration, more forcefully apply gradients.
 post_momentumfloat (default 0.8)
During the late phases, less forcefully apply gradients.
 handle(cuML Handle, default None)
You can pass in a past handle that was initialized, or we will create one for you anew!
References
van der Maaten, L.J.P. tDistributed Stochastic Neighbor Embedding https://lvdmaaten.github.io/tsne/
van der Maaten, L.J.P.; Hinton, G.E. Visualizing HighDimensional Data Using tSNE. Journal of Machine Learning Research 9:25792605, 2008.
George C. Linderman, Manas Rachh, Jeremy G. Hoskins, Stefan Steinerberger, Yuval Kluger Efficient Algorithms for tdistributed Stochastic Neighborhood Embedding
Methods
fit
(self, X[, convert_dtype])Fit X into an embedded space.
fit_transform
(self, X[, convert_dtype])Fit X into an embedded space and return that transformed output.

fit
(self, X, convert_dtype=True)¶ Fit X into an embedded space.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
X contains a sample per row. Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit method will automatically convert the inputs to np.float32.

fit_transform
(self, X, convert_dtype=True)¶ Fit X into an embedded space and return that transformed output.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
X contains a sample per row. Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit_transform method will automatically convert the inputs to np.float32.
 Returns
 X_newarray, shape (n_samples, n_components)
Embedding of the training data in lowdimensional space.
Neighbors¶
Nearest Neighbors¶

class
cuml.neighbors.
NearestNeighbors
¶ NearestNeighbors is an queries neighborhoods from a given set of datapoints. Currently, cuML supports kNN queries, which define the neighborhood as the closest k neighbors to each query point.
 Parameters
 n_neighborsint (default=5)
Default number of neighbors to query
 verboseint or boolean (default = False)
Logging level
 handlecumlHandle
The cumlHandle resources to use
 algorithmstring (default=’brute’)
The query algorithm to use. Currently, only ‘brute’ is supported.
 metricstring (default=’euclidean’).
Distance metric to use.
Notes
For an additional example see the NearestNeighbors notebook.
For additional docs, see scikitlearn’s NearestNeighbors.
Examples
import cudf from cuml.neighbors import NearestNeighbors from cuml.datasets import make_blobs X, _ = make_blobs(n_samples=25, centers=5, n_features=10, random_state=42) # build a cudf Dataframe X_cudf = cudf.DataFrame.from_gpu_matrix(X) # fit model model = NearestNeighbors(n_neighbors=3) model.fit(X) # get 3 nearest neighbors distances, indices = model.kneighbors(X_cudf) # print results print(indices) print(distances)
Output:
indices: 0 1 2 0 0 14 21 1 1 19 8 2 2 9 23 3 3 14 21 ... 22 22 18 11 23 23 16 9 24 24 17 10 distances: 0 1 2 0 0.0 4.883116 5.570006 1 0.0 3.047896 4.105496 2 0.0 3.558557 3.567704 3 0.0 3.806127 3.880100 ... 22 0.0 4.210738 4.227068 23 0.0 3.357889 3.404269 24 0.0 3.428183 3.818043
Methods
fit
(self, X[, convert_dtype])Fit GPU index for performing nearest neighbor queries.
kneighbors
(self[, X, n_neighbors, …])Query the GPU index for the k nearest neighbors of column vectors in X.

fit
(self, X, convert_dtype=True)¶ Fit GPU index for performing nearest neighbor queries.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit method will automatically convert the inputs to np.float32.

kneighbors
(self, X=None, n_neighbors=None, return_distance=True, convert_dtype=True)¶ Query the GPU index for the k nearest neighbors of column vectors in X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 n_neighborsInteger
Number of neighbors to search. If not provided, the n_neighbors from the model instance is used (default=10)
 return_distance: Boolean
If False, distances will not be returned
 convert_dtypebool, optional (default = True)
When set to True, the kneighbors method will automatically convert the inputs to np.float32.
 Returns
 distances: cuDF DataFrame or numpy ndarray
The distances of the knearest neighbors for each column vector in X
 indices: cuDF DataFrame of numpy ndarray
The indices of the knearest neighbors for each column vector in X
Nearest Neighbors Classification¶

class
cuml.neighbors.
KNeighborsClassifier
¶ KNearest Neighbors Classifier is an instancebased learning technique, that keeps training samples around for prediction, rather than trying to learn a generalizable set of model parameters.
 Parameters
 n_neighborsint (default=5)
Default number of neighbors to query
 verboseint or boolean (default = False)
Logging level
 handlecumlHandle
The cumlHandle resources to use
 algorithmstring (default=’brute’)
The query algorithm to use. Currently, only ‘brute’ is supported.
 metricstring (default=’euclidean’).
Distance metric to use.
 weightsstring (default=’uniform’)
Sample weights to use. Currently, only the uniform strategy is supported.
Notes
For additional docs, see scikitlearn’s KNeighborsClassifier.
Examples
from cuml.neighbors import KNeighborsClassifier from sklearn.datasets import make_blobs from sklearn.model_selection import train_test_split X, y = make_blobs(n_samples=100, centers=5, n_features=10) knn = KNeighborsClassifier(n_neighbors=10) X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.80) knn.fit(X_train, y_train) knn.predict(X_test)
Methods
fit
(self, X, y[, convert_dtype])Fit a GPU index for knearest neighbors classifier model.
get_param_names
(self)predict
(self, X[, convert_dtype])Use the trained knearest neighbors classifier to predict the labels for X
predict_proba
(self, X[, convert_dtype])Use the trained knearest neighbors classifier to predict the label probabilities for X
score
(self, X, y[, convert_dtype])Compute the accuracy score using the given labels and the trained knearest neighbors classifier to predict the classes for X.

fit
(self, X, y, convert_dtype=True)¶ Fit a GPU index for knearest neighbors classifier model.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, n_outputs)
Dense matrix (floats or doubles) of shape (n_samples, n_outputs). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit method will automatically convert the inputs to np.float32.

get_param_names
(self)¶

predict
(self, X, convert_dtype=True)¶ Use the trained knearest neighbors classifier to predict the labels for X
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit method will automatically convert the inputs to np.float32.

predict_proba
(self, X, convert_dtype=True)¶ Use the trained knearest neighbors classifier to predict the label probabilities for X
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit method will automatically convert the inputs to np.float32.

score
(self, X, y, convert_dtype=True)¶ Compute the accuracy score using the given labels and the trained knearest neighbors classifier to predict the classes for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit method will automatically convert the inputs to np.float32.
Nearest Neighbors Regression¶

class
cuml.neighbors.
KNeighborsRegressor
¶ KNearest Neighbors Regressor is an instancebased learning technique, that keeps training samples around for prediction, rather than trying to learn a generalizable set of model parameters.
The KNearest Neighbors Regressor will compute the average of the labels for the k closest neighbors and use it as the label.
 Parameters
 n_neighborsint (default=5)
Default number of neighbors to query
 verboseint or boolean (default = False)
Logging level
 handlecumlHandle
The cumlHandle resources to use
 algorithmstring (default=’brute’)
The query algorithm to use. Currently, only ‘brute’ is supported.
 metricstring (default=’euclidean’).
Distance metric to use.
 weightsstring (default=’uniform’)
Sample weights to use. Currently, only the uniform strategy is supported.
Notes
For additional docs, see scikitlearn’s KNeighborsClassifier.
Examples
from cuml.neighbors import KNeighborsRegressor from sklearn.datasets import make_blobs from sklearn.model_selection import train_test_split X, y = make_blobs(n_samples=100, centers=5, n_features=10) knn = KNeighborsRegressor(n_neighbors=10) X_train, X_test, y_train, y_test = train_test_split(X, y, train_size=0.80) knn.fit(X_train, y_train) knn.predict(X_test)
Output:
array([3. , 1. , 1. , 3.79999995, 2. , 0. , 3.79999995, 3.79999995, 3.79999995, 0. , 3.79999995, 0. , 1. , 2. , 3. , 1. , 0. , 0. , 0. , 2. , 3. , 3. , 0. , 3. , 3.79999995, 3.79999995, 3.79999995, 3.79999995, 3. , 2. , 3.79999995, 3.79999995, 0. ])
Methods
fit
(self, X, y[, convert_dtype])Fit a GPU index for knearest neighbors regression model.
predict
(self, X[, convert_dtype])Use the trained knearest neighbors regression model to predict the labels for X
score
(self, X, y[, convert_dtype])Fit a GPU index for knearest neighbors regression model.

fit
(self, X, y, convert_dtype=True)¶ Fit a GPU index for knearest neighbors regression model.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, n_outputs)
Dense matrix (floats or doubles) of shape (n_samples, n_outputs). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit method will automatically convert the inputs to np.float32.

predict
(self, X, convert_dtype=True)¶ Use the trained knearest neighbors regression model to predict the labels for X
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit method will automatically convert the inputs to np.float32.

score
(self, X, y, convert_dtype=True)¶ Fit a GPU index for knearest neighbors regression model.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 yarraylike (device or host) shape = (n_samples, n_features)
Dense matrix (floats or doubles) of shape (n_samples, n_features). Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = True)
When set to True, the fit method will automatically convert the inputs to np.float32.
Time Series¶
HoltWinters¶

class
cuml.
ExponentialSmoothing
¶ Implements a HoltWinters time series analysis model which is used in both forecasting future entries in a time series as well as in providing exponential smoothing, where weights are assigned against historical data with exponentially decreasing impact. This is done by analyzing three components of the data: level, trend, and seasonality.
 Parameters
 endogarraylike (device or host)
Acceptable formats: cuDF DataFrame, cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy. Note: cuDF.DataFrame types assumes data is in columns, while all other datatypes assume data is in rows. The endogenous dataset to be operated on.
 seasonal‘additive’, ‘add’, ‘multiplicative’, ‘mul’ (default = ‘additive’) # noqa
Whether the seasonal trend should be calculated additively or multiplicatively.
 seasonal_periodsint (default=2)
The seasonality of the data (how often it repeats). For monthly data this should be 12, for weekly data, this should be 7.
 start_periodsint (default=2)
Number of seasons to be used for seasonal seed values
 ts_numint (default=1)
The number of different time series that were passed in the endog param.
 epsnp.number > 0 (default=2.24e3)
The accuracy to which gradient descent should achieve. Note that changing this value may affect the forecasted results.
 handlecuml.Handle (default=None)
If it is None, a new one is created just for this class.
Examples
from cuml import ExponentialSmoothing import cudf import numpy as np data = cudf.Series([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14], dtype=np.float64) cu_hw = ExponentialSmoothing(data, seasonal_periods=12) cu_hw.fit() cu_pred = cu_hw.forecast(4) print('Forecasted points:', cu_pred)
Output
Forecasted points : 0 4.000143766093652 1 5.000000163513641 2 6.000000000174092 3 7.000000000000178
Methods
fit
(self)Perform fitting on the given endog dataset.
forecast
(self[, h, index])Forecasts future points based on the fitted model.
get_level
(self[, index])Returns the level component of the model.
get_season
(self[, index])Returns the season component of the model.
get_trend
(self[, index])Returns the trend component of the model.
score
(self[, index])Returns the score of the model.

fit
(self)¶ Perform fitting on the given endog dataset. Calculates the level, trend, season, and SSE components.

forecast
(self, h=1, index=None)¶ Forecasts future points based on the fitted model.
 Parameters
 hint (default=1)
The number of points for each series to be forecasted.
 indexint (default=None)
The index of the time series from which you want forecasted points. if None, then a cudf.DataFrame of the forecasted points from all time series is returned.
 Returns
 predscudf.DataFrame or cudf.Series
Series of forecasted points if index is provided. DataFrame of all forecasted points if index=None.

get_level
(self, index=None)¶ Returns the level component of the model.
 Parameters
 indexint (default=None)
The index of the time series from which the level will be returned. if None, then all level components are returned in a cudf.Series.
 Returns
 levelcudf.Series or cudf.DataFrame
The level component of the fitted model

get_season
(self, index=None)¶ Returns the season component of the model.
 Parameters
 indexint (default=None)
The index of the time series from which the season will be returned. if None, then all season components are returned in a cudf.Series.
 Returns
 season: cudf.Series or cudf.DataFrame
The season component of the fitted model

get_trend
(self, index=None)¶ Returns the trend component of the model.
 Parameters
 indexint (default=None)
The index of the time series from which the trend will be returned. if None, then all trend components are returned in a cudf.Series.
 Returns
 trendcudf.Series or cudf.DataFrame
The trend component of the fitted model.

score
(self, index=None)¶ Returns the score of the model.
 **Note: Currently returns the SSE, rather than the gradient of the
LogLikelihood. https://github.com/rapidsai/cuml/issues/876
 Parameters
 indexint (default=None)
The index of the time series from which the SSE will be returned. if None, then all SSEs are returned in a cudf Series.
 Returns
 scorenp.float32, np.float64, or cudf.Series
The SSE of the fitted model.
ARIMA¶

class
cuml.tsa.
ARIMA
¶ Implements a batched ARIMA model for in and outofsample timeseries prediction, with support for seasonality (SARIMA)
ARIMA stands for AutoRegressive Integrated Moving Average. See https://en.wikipedia.org/wiki/Autoregressive_integrated_moving_average
This class can fit an ARIMA(p,d,q) or ARIMA(p,d,q)(P,D,Q)_s model to a batch of time series of the same length with no missing values. The implementation is designed to give the best performance when using large batches of time series.
 Parameters
 ydataframe or arraylike (device or host)
The time series data, assumed to have each time series in columns. Acceptable formats: cuDF DataFrame, cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy.
 orderTuple[int, int, int]
The ARIMA order (p, d, q) of the model
 seasonal_order: Tuple[int, int, int, int]
The seasonal ARIMA order (P, D, Q, s) of the model
 fit_interceptbool or int
Whether to include a constant trend mu in the model (default: True)
 handlecuml.Handle
If it is None, a new one is created just for this instance
 verboseint or boolean (default = False)
Controls verbose level of logging.
 output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’}, optional
Variable to control output type of the results and attributes of the estimators. If None, it’ll inherit the output type set at the module level, cuml.output_type. If set, the estimator will override the global option for its behavior.
References
This class is heavily influenced by the Python library statsmodels, particularly statsmodels.tsa.statespace.sarimax.SARIMAX. See https://www.statsmodels.org/stable/statespace.html
Additionally the following book is a useful reference: “Time Series Analysis by State Space Methods”, J. Durbin, S.J. Koopman, 2nd Edition (2012).
Examples
import numpy as np from cuml.tsa.arima import ARIMA # Create seasonal data with a trend, a seasonal pattern and noise n_obs = 100 np.random.seed(12) x = np.linspace(0, 1, n_obs) pattern = np.array([[0.05, 0.0], [0.07, 0.03], [0.03, 0.05], [0.02, 0.025]]) noise = np.random.normal(scale=0.01, size=(n_obs, 2)) y = (np.column_stack((0.5*x, 0.25*x)) + noise + np.tile(pattern, (25, 1))) # Fit a seasonal ARIMA model model = ARIMA(y, (0,1,1), (0,1,1,4), fit_intercept=False) model.fit() # Forecast fc = model.forecast(10) print(fc)
Output:
[[ 0.55204599 0.25681163] [ 0.57430705 0.2262438 ] [ 0.48120315 0.20583011] [ 0.535594 0.24060046] [ 0.57207541 0.26695497] [ 0.59433647 0.23638713] [ 0.50123257 0.21597344] [ 0.55562342 0.25074379] [ 0.59210483 0.27709831] [ 0.61436589 0.24653047]]
 Attributes
 orderTuple[int, int, int]
The ARIMA order (p, d, q) of the model
 seasonal_order: Tuple[int, int, int, int]
The seasonal ARIMA order (P, D, Q, s) of the model
 interceptbool or int
Whether the model includes a constant trend mu
 d_y: device array
Time series data on device
 num_samples: int
Number of observations
 batch_size: int
Number of time series in the batch
 dtype: numpy.dtype
Floatingpoint type of the data and parameters
 niter: numpy.ndarray
After fitting, contains the number of iterations before convergence for each time series.
Methods
fit
(self, start_params, object], …[, maxiter])Fit the ARIMA model to each time series.
forecast
(self, nsteps)Forecast the given model nsteps into the future.
get_params
(self)Get the parameters of the model
pack
(self)Pack parameters of the model into a linearized vector x
predict
(self[, start, end])Compute insample and/or outofsample prediction for each series
set_params
(self, params, object])Set the parameters of the model
unpack
(self, x, numpy.ndarray])Unpack linearized parameter vector x into the separate parameter arrays of the model

property
aic
¶ Akaike Information Criterion

property
aicc
¶ Corrected Akaike Information Criterion

property
bic
¶ Bayesian Information Criterion

property
complexity
¶ Model complexity (number of parameters)

fit
(self, start_params: Union[Mapping[str, object], NoneType] = None, opt_disp: ‘int’ =  1, h: ‘float’ = 1e09, maxiter=1000)¶ Fit the ARIMA model to each time series.
 Parameters
 start_paramsMapping[str, object] (optional)
A mapping (e.g dictionary) of parameter names and associated arrays The key names are in {“mu”, “ar”, “ma”, “sar”, “sma”, “sigma2”} The shape of the arrays are (batch_size,) for mu parameters and (n, batch_size) for any other type, where n is the corresponding number of parameters of this type. Pass None for automatic estimation (recommended)
 opt_dispint
 Fit diagnostic level (for LBFGS solver):
1 for no output (default)
0<n<100 for output every n steps
n>100 for more detailed output
 hfloat
Finitedifferencing step size. The gradient is computed using secondorder differencing:
f(x+h)  f(x  h)
 g = —————– + O(h^2)
2 * h
 maxiterint
Maximum number of iterations of LBFGSB

forecast
(self, nsteps: ‘int’)¶ Forecast the given model nsteps into the future.
 nstepsint
The number of steps to forecast beyond end of the given series

property
llf
¶ Loglikelihood of a fit model. Shape: (batch_size,)

pack
(self) → numpy.ndarray¶ Pack parameters of the model into a linearized vector x

predict
(self, start=0, end=None)¶ Compute insample and/or outofsample prediction for each series

set_params
(self, params: Mapping[str, object])¶ Set the parameters of the model
 params: Mapping[str, np.ndarray]
A mapping (e.g dictionary) of parameter names and associated arrays The key names are in {“mu”, “ar”, “ma”, “sar”, “sma”, “sigma2”} The shape of the arrays are (batch_size,) for mu parameters and (n, batch_size) for any other type, where n is the corresponding number of parameters of this type.
MultiNode, MultiGPU Algorithms¶
KMeans Clustering¶

class
cuml.dask.cluster.
KMeans
(client=None, verbose=False, **kwargs)¶ MultiNode MultiGPU implementation of KMeans.
This version minimizes data transfer by sharing only the centroids between workers in each iteration.
Predictions are done embarrassingly parallel, using cuML’s singleGPU version.
For more information on this implementation, refer to the documentation for singleGPU KMeans.
 Parameters
 handlecuml.Handle
If it is None, a new one is created just for this class.
 n_clustersint (default = 8)
The number of centroids or clusters you want.
 max_iterint (default = 300)
The more iterations of EM, the more accurate, but slower.
 tolfloat (default = 1e4)
Stopping criterion when centroid means do not change much.
 verboseint or boolean (default = False)
Logging level for printing diagnostic information
 random_stateint (default = 1)
If you want results to be the same when you restart Python, select a state.
 init{‘scalablekmeans++’, ‘kmeans’ , ‘random’ or an ndarray}
(default = ‘scalablekmeans++’)
‘scalablekmeans++’ or ‘kmeans’: Uses fast and stable scalable kmeans++ intialization. ‘random’: Choose ‘n_cluster’ observations (rows) at random from data for the initial centroids. If an ndarray is passed, it should be of shape (n_clusters, n_features) and gives the initial centers.
 oversampling_factorint (default = 2) The amount of points to sample
in scalable kmeans++ initialization for potential centroids. Increasing this value can lead to better initial centroids at the cost of memory. The total number of centroids sampled in scalable kmeans++ is oversampling_factor * n_clusters * 8.
 max_samples_per_batchint (default = 32768) The number of data
samples to use for batches of the pairwise distance computation. This computation is done throughout both fit predict. The default should suit most cases. The total number of elements in the batched pairwise distance computation is max_samples_per_batch * n_clusters. It might become necessary to lower this number when n_clusters becomes prohibitively large.
 Attributes
 cluster_centers_cuDF DataFrame or CuPy ndarray
The coordinates of the final clusters. This represents of “mean” of each data cluster.
Methods
fit
(self, X)Fit a multinode multiGPU KMeans model
fit_predict
(self, X[, delayed])Compute cluster centers and predict cluster index for each sample.
fit_transform
(self, X[, delayed])Calls fit followed by transform using a distributed KMeans model
predict
(self, X[, delayed])Predict labels for the input
score
(self, X)Computes the inertia score for the trained KMeans centroids.
transform
(self, X[, delayed])Transforms the input into the learned centroid space
get_param_names

fit
(self, X)¶ Fit a multinode multiGPU KMeans model
 Parameters
 XDask cuDF DataFrame or CuPy backed Dask Array
 Training data to cluster.

fit_predict
(self, X, delayed=True)¶ Compute cluster centers and predict cluster index for each sample.
 Parameters
 XDask cuDF DataFrame or CuPy backed Dask Array
Data to predict
 Returns
 result: Dask cuDF DataFrame or CuPy backed Dask Array
Distributed object containing predictions

fit_transform
(self, X, delayed=True)¶ Calls fit followed by transform using a distributed KMeans model
 Parameters
 XDask cuDF DataFrame or CuPy backed Dask Array
Data to predict
 delayedbool (default = True)
Whether to execute as a delayed task or eager.
 Returns
 result: Dask cuDF DataFrame or CuPy backed Dask Array
Distributed object containing the transformed data

predict
(self, X, delayed=True)¶ Predict labels for the input
 Parameters
 XDask cuDF DataFrame or CuPy backed Dask Array
Data to predict
 delayedbool (default = True)
Whether to do a lazy prediction (and return Delayed objects) or an eagerly executed one.
 Returns
 result: Dask cuDF DataFrame or CuPy backed Dask Array
Distributed object containing predictions

score
(self, X)¶ Computes the inertia score for the trained KMeans centroids.
 Parameters
 Xdask_cudf.Dataframe
Dataframe to compute score
 Returns
 Inertial score

transform
(self, X, delayed=True)¶ Transforms the input into the learned centroid space
 Parameters
 XDask cuDF DataFrame or CuPy backed Dask Array
Data to predict
 delayedbool (default = True)
Whether to execute as a delayed task or eager.
 Returns
 result: Dask cuDF DataFrame or CuPy backed Dask Array
Distributed object containing the transformed data
Nearest Neighbors¶

class
cuml.dask.neighbors.
NearestNeighbors
(client=None, streams_per_handle=0, **kwargs)¶ Multinode MultiGPU NearestNeighbors Model.
Methods
fit
(self, X)Fit a multinode multiGPU Nearest Neighbors index
get_neighbors
(self, n_neighbors)Returns the default n_neighbors, initialized from the constructor, if n_neighbors is None.
kneighbors
(self[, X, n_neighbors, …])Query the distributed nearest neighbors index

fit
(self, X)¶ Fit a multinode multiGPU Nearest Neighbors index
 Parameters
 Xdask_cudf.Dataframe
 Returns
 self: NearestNeighbors model

get_neighbors
(self, n_neighbors)¶ Returns the default n_neighbors, initialized from the constructor, if n_neighbors is None.
 Parameters
 n_neighborsint
Number of neighbors
 Returns
 n_neighbors: int
Default n_neighbors if parameter n_neighbors is none

kneighbors
(self, X=None, n_neighbors=None, return_distance=True, _return_futures=False)¶ Query the distributed nearest neighbors index
 Parameters
 Xdask_cudf.Dataframe
Vectors to query. If not provided, neighbors of each indexed point are returned.
 n_neighborsint
Number of neighbors to query for each row in X. If not provided, the n_neighbors on the model are used.
 return_distanceboolean (default=True)
If false, only indices are returned
 Returns
 rettuple (dask_cudf.DataFrame, dask_cudf.DataFrame)
First daskcuDF DataFrame contains distances, second conains the indices.

Principal Component Analysis¶

class
cuml.dask.decomposition.
PCA
(client=None, verbose=False, **kwargs)¶ PCA (Principal Component Analysis) is a fundamental dimensionality reduction technique used to combine features in X in linear combinations such that each new component captures the most information or variance of the data. N_components is usually small, say at 3, where it can be used for data visualization, data compression and exploratory analysis.
cuML’s multinode multigpu (MNMG) PCA expects a dask cuDF input, and provides a “Full” algorithm. It uses a full eigendecomposition then selects the top K eigenvectors.
 Parameters
 handlecuml.Handle
If it is None, a new one is created just for this class
 n_componentsint (default = 1)
The number of top K singular vectors / values you want. Must be <= number(columns).
 svd_solver‘full’
Only Full algorithm is supported since it’s significantly faster on GPU then the other solvers including randomized SVD.
 verboseint or boolean (default = False)
Logging level
 whitenboolean (default = False)
If True, decorrelates the components. This is done by dividing them by the corresponding singular values then multiplying by sqrt(n_samples). Whitening allows each component to have unit variance and removes multicollinearity. It might be beneficial for downstream tasks like LinearRegression where correlated features cause problems.
Notes
PCA considers linear combinations of features, specifically those that maximise global variance structure. This means PCA is fantastic for global structure analyses, but weak for local relationships. Consider UMAP or TSNE for a locally important embedding.
Applications of PCA
PCA is used extensively in practice for data visualization and data compression. It has been used to visualize extremely large word embeddings like Word2Vec and GloVe in 2 or 3 dimensions, large datasets of everyday objects and images, and used to distinguish between cancerous cells from healthy cells.
For an additional example see the PCA notebook. For additional docs, see scikitlearn’s PCA.
Examples
from dask_cuda import LocalCUDACluster from dask.distributed import Client, wait import numpy as np from cuml.dask.decomposition import PCA from cuml.dask.datasets import make_blobs cluster = LocalCUDACluster(threads_per_worker=1) client = Client(cluster) nrows = 6 ncols = 3 n_parts = 2 X_cudf, _ = make_blobs(nrows, ncols, 1, n_parts, cluster_std=0.01, verbose=cuml.logger.level_info, random_state=10, dtype=np.float32) wait(X_cudf) print("Input Matrix") print(X_cudf.compute()) cumlModel = PCA(n_components = 1, whiten=False) XT = cumlModel.fit_transform(X_cudf) print("Transformed Input Matrix") print(XT.compute())
Output:
Input Matrix: 0 1 2 0 6.520953 0.015584 8.828546 1 6.507554 0.016524 8.836799 2 6.518214 0.010457 8.821301 0 6.520953 0.015584 8.828546 1 6.507554 0.016524 8.836799 2 6.518214 0.010457 8.821301 Transformed Input Matrix: 0 0 0.003271 1 0.011454 2 0.008182 0 0.003271 1 0.011454 2 0.008182
 Note: Everytime this code is run, the output will be different because
“make_blobs” function generates random matrices.
 Attributes
 components_array
The top K components (VT.T[:,:n_components]) in U, S, VT = svd(X)
 explained_variance_array
How much each component explains the variance in the data given by S**2
 explained_variance_ratio_array
How much in % the variance is explained given by S**2/sum(S**2)
 singular_values_array
The top K singular values. Remember all singular values >= 0
 mean_array
The column wise mean of X. Used to mean  center the data first.
 noise_variance_float
From Bishop 1999’s Textbook. Used in later tasks like calculating the estimated covariance of X.
Methods
fit
(self, X)Fit the model with X.
fit_transform
(self, X)Fit the model with X and apply the dimensionality reduction on X.
inverse_transform
(self, X[, delayed])Transform data back to its original space.
transform
(self, X[, delayed])Apply dimensionality reduction to X.
get_param_names

fit
(self, X)¶ Fit the model with X.
 Parameters
 Xdask cuDF input

fit_transform
(self, X)¶ Fit the model with X and apply the dimensionality reduction on X.
 Parameters
 Xdask cuDF
 Returns
 X_newdask cuDF

inverse_transform
(self, X, delayed=True)¶ Transform data back to its original space.
In other words, return an input X_original whose transform would be X.
 Parameters
 Xdask cuDF
 Returns
 X_originaldask cuDF

transform
(self, X, delayed=True)¶ Apply dimensionality reduction to X.
X is projected on the first principal components previously extracted from a training set.
 Parameters
 Xdask cuDF
 Returns
 X_newdask cuDF
Random Forest¶

class
cuml.dask.ensemble.
RandomForestClassifier
(workers=None, client=None, verbose=False, n_estimators=10, seed=None, **kwargs)¶ Experimental API implementing a multiGPU Random Forest classifier model which fits multiple decision tree classifiers in an ensemble. This uses Dask to partition data over multiple GPUs (possibly on different nodes).
Currently, this API makes the following assumptions: * The set of Dask workers used between instantiation, fit, and predict are all consistent * Training data comes in the form of cuDF dataframes, distributed so that each worker has at least one partition.
Future versions of the API will support more flexible data distribution and additional input types.
The distributed algorithm uses an embarrassinglyparallel approach. For a forest with N trees being built on w workers, each worker simply builds N/w trees on the data it has available locally. In many cases, partitioning the data so that each worker builds trees on a subset of the total dataset works well, but it generally requires the data to be wellshuffled in advance. Alternatively, callers can replicate all of the data across workers so that rf.fit receives w partitions, each containing the same data. This would produce results approximately identical to singleGPU fitting.
Please check the singleGPU implementation of Random Forest classifier for more information about the underlying algorithm.
 Parameters
 n_estimatorsint (default = 10)
total number of trees in the forest (not perworker)
 handlecuml.Handle
If it is None, a new one is created just for this class.
 split_criterionThe criterion used to split nodes.
0 for GINI, 1 for ENTROPY, 4 for CRITERION_END. 2 and 3 not valid for classification (default = 0)
 split_algo0 for HIST and 1 for GLOBAL_QUANTILE
(default = 1) the algorithm to determine how nodes are split in the tree.
 split_criterionThe criterion used to split nodes.
0 for GINI, 1 for ENTROPY, 4 for CRITERION_END. 2 and 3 not valid for classification (default = 0)
 bootstrapboolean (default = True)
Control bootstrapping. If set, each tree in the forest is built on a bootstrapped sample with replacement. If false, sampling without replacement is done.
 bootstrap_featuresboolean (default = False)
Control bootstrapping for features. If features are drawn with or without replacement
 rows_samplefloat (default = 1.0)
Ratio of dataset rows used while fitting each tree.
 max_depthint (default = 1)
Maximum tree depth. Unlimited (i.e, until leaves are pure), if 1.
 max_leavesint (default = 1)
Maximum leaf nodes per tree. Soft constraint. Unlimited, if 1.
 max_featuresfloat (default = ‘auto’)
Ratio of number of features (columns) to consider per node split.
 n_binsint (default = 8)
Number of bins used by the split algorithm.
 min_rows_per_nodeint (default = 2)
The minimum number of samples (rows) needed to split a node.
 quantile_per_treeboolean (default = False)
Whether quantile is computed for individual RF trees. Only relevant for GLOBAL_QUANTILE split_algo.
 n_streamsint (default = 4 )
Number of parallel streams used for forest building
 workersoptional, list of strings
Dask addresses of workers to use for computation. If None, all available Dask workers will be used.
 seedint (default = None)
Base seed for the random number generator. Unseeded by default. Does not currently fully guarantee the exact same results.
Examples
For usage examples, please see the RAPIDS notebooks repository: https://github.com/rapidsai/notebooks/blob/branch0.12/cuml/random_forest_mnmg_demo.ipynb
Methods
fit
(self, X, y[, convert_dtype])Fit the input data with a Random Forest classifier
get_params
(self[, deep])Returns the value of all parameters required to configure this estimator as a dictionary.
predict
(self, X[, output_class, algo, …])Predicts the labels for X.
predict_model_on_cpu
(self, X[, convert_dtype])Predicts the labels for X.
predict_proba
(self, X[, delayed])Predicts the probability of each class for X.
print_summary
(self)Print the summary of the forest used to train and test the model.
set_params
(self, **params)Sets the value of parameters required to configure this estimator, it functions similar to the sklearn set_params.
predict_using_fil

fit
(self, X, y, convert_dtype=False)¶ Fit the input data with a Random Forest classifier
IMPORTANT: X is expected to be partitioned with at least one partition on each Dask worker being used by the forest (self.workers).
If a worker has multiple data partitions, they will be concatenated before fitting, which will lead to additional memory usage. To minimize memory consumption, ensure that each worker has exactly one partition.
When persisting data, you can use cuml.dask.common.utils.persist_across_workers to simplify this:
X_dask_cudf = dask_cudf.from_cudf(X_cudf, npartitions=n_workers) y_dask_cudf = dask_cudf.from_cudf(y_cudf, npartitions=n_workers) X_dask_cudf, y_dask_cudf = persist_across_workers(dask_client, [X_dask_cudf, y_dask_cudf])
 (this is equivalent to calling persist with the data and workers)::
 X_dask_cudf, y_dask_cudf = dask_client.persist([X_dask_cudf,
y_dask_cudf],
workers={ X_dask_cudf=workers, y_dask_cudf=workers })
 Parameters
 XDask cuDF dataframe or CuPy backed Dask Array (n_rows, n_features)
Distributed dense matrix (floats or doubles) of shape (n_samples, n_features).
 yDask cuDF dataframe or CuPy backed Dask Array (n_rows, 1)
Labels of training examples. y must be partitioned the same way as X
 convert_dtypebool, optional (default = False)
When set to True, the fit method will, when necessary, convert y to be the same data type as X if they differ. This will increase memory used for the method.

get_params
(self, deep=True)¶ Returns the value of all parameters required to configure this estimator as a dictionary.
 Parameters
 deepboolean (default = True)

predict
(self, X, output_class=True, algo='auto', threshold=0.5, convert_dtype=True, predict_model='GPU', fil_sparse_format='auto', delayed=True)¶ Predicts the labels for X.
GPUbased prediction in a multinode, multiGPU context works by sending the subforest from each worker to the client, concatenating these into one forest with the full n_estimators set of trees, and sending this combined forest to the workers, which will each infer on their local set of data. Within the worker, this uses the cuML Forest Inference Library (cuml.fil) for highthroughput prediction.
This allows inference to scale to large datasets, but the forest transmission incurs overheads for very large trees. For inference on small datasets, this overhead may dominate prediction time.
The ‘CPU’ fallback method works with subforests inplace, broadcasting the datasets to all workers and combining predictions via a voting method at the end. This method is slower on a perrow basis but may be faster for problems with many trees and few rows.
In the 0.15 cuML release, inference will be updated with much faster tree transfer. Preliminary builds with this updated approach will be available from rapids.ai
 Parameters
 XDask cuDF dataframe or CuPy backed Dask Array (n_rows, n_features)
Distributed dense matrix (floats or doubles) of shape (n_samples, n_features).
 output_classboolean (default = True)
This is optional and required only while performing the predict operation on the GPU. If true, return a 1 or 0 depending on whether the raw prediction exceeds the threshold. If False, just return the raw prediction.
 algostring (default = ‘auto’)
This is optional and required only while performing the predict operation on the GPU. ‘naive’  simple inference using shared memory ‘tree_reorg’  similar to naive but trees rearranged to be more coalescingfriendly ‘batch_tree_reorg’  similar to tree_reorg but predicting multiple rows per thread block algo  choose the algorithm automatically. Currently ‘batch_tree_reorg’ is used for dense storage and ‘naive’ for sparse storage
 thresholdfloat (default = 0.5)
Threshold used for classification. Optional and required only while performing the predict operation on the GPU, that is for, predict_model=’GPU’. It is applied if output_class == True, else it is ignored
 convert_dtypebool, optional (default = True)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 predict_modelString (default = ‘GPU’)
‘GPU’ to predict using the GPU, ‘CPU’ otherwise. The GPU can only be used if the model was trained on float32 data and X is float32 or convert_dtype is set to True.
 fil_sparse_formatboolean or string (default = auto)
This variable is used to choose the type of forest that will be created in the Forest Inference Library. It is not required while using predict_model=’CPU’. ‘auto’  choose the storage type automatically (currently True is chosen by auto) False  create a dense forest True  create a sparse forest, requires algo=’naive’ or algo=’auto’
 delayedbool (default = True)
Whether to do a lazy prediction (and return Delayed objects) or an eagerly executed one. It is not required while using predict_model=’CPU’.
 Returns
 yDask cuDF dataframe or CuPy backed Dask Array (n_rows, 1)

predict_model_on_cpu
(self, X, convert_dtype=True)¶ Predicts the labels for X.
 Parameters
 XDask cuDF dataframe or CuPy backed Dask Array (n_rows, n_features)
Distributed dense matrix (floats or doubles) of shape (n_samples, n_features).
 convert_dtypebool, optional (default = True)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 Returns
 ———
 yDask cuDF dataframe or CuPy backed Dask Array (n_rows, 1)

predict_proba
(self, X, delayed=True, **kwargs)¶ Predicts the probability of each class for X.
See documentation of `predict’ for notes on performance.
 Parameters
 XDask cuDF dataframe or CuPy backed Dask Array (n_rows, n_features)
Distributed dense matrix (floats or doubles) of shape (n_samples, n_features).
 predict_modelString (default = ‘GPU’)
‘GPU’ to predict using the GPU, ‘CPU’ otherwise. The ‘GPU’ can only be used if the model was trained on float32 data and X is float32 or convert_dtype is set to True. Also the ‘GPU’ should only be used for binary classification problems.
 output_classboolean (default = True)
This is optional and required only while performing the predict operation on the GPU. If true, return a 1 or 0 depending on whether the raw prediction exceeds the threshold. If False, just return the raw prediction.
 algostring (default = ‘auto’)
This is optional and required only while performing the predict operation on the GPU. ‘naive’  simple inference using shared memory ‘tree_reorg’  similar to naive but trees rearranged to be more coalescingfriendly ‘batch_tree_reorg’  similar to tree_reorg but predicting multiple rows per thread block auto  choose the algorithm automatically. Currently ‘batch_tree_reorg’ is used for dense storage and ‘naive’ for sparse storage
 thresholdfloat (default = 0.5)
Threshold used for classification. Optional and required only while performing the predict operation on the GPU. It is applied if output_class == True, else it is ignored
 num_classesint (default = 2)
number of different classes present in the dataset
 convert_dtypebool, optional (default = True)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 fil_sparse_formatboolean or string (default = auto)
This variable is used to choose the type of forest that will be created in the Forest Inference Library. It is not required while using predict_model=’CPU’. ‘auto’  choose the storage type automatically (currently True is chosen by auto) False  create a dense forest True  create a sparse forest, requires algo=’naive’ or algo=’auto’
 Returns
 yNumPy
Dask cuDF dataframe or CuPy backed Dask Array (n_rows, n_classes)

print_summary
(self)¶ Print the summary of the forest used to train and test the model.

set_params
(self, **params)¶ Sets the value of parameters required to configure this estimator, it functions similar to the sklearn set_params.
 Parameters
 paramsdict of new params.

class
cuml.dask.ensemble.
RandomForestRegressor
(workers=None, client=None, verbose=False, n_estimators=10, seed=None, **kwargs)¶ Experimental API implementing a multiGPU Random Forest classifier model which fits multiple decision tree classifiers in an ensemble. This uses Dask to partition data over multiple GPUs (possibly on different nodes).
Currently, this API makes the following assumptions: * The set of Dask workers used between instantiation, fit, and predict are all consistent * Training data is comes in the form of cuDF dataframes, distributed so that each worker has at least one partition.
Future versions of the API will support more flexible data distribution and additional input types. Userfacing APIs are expected to change in upcoming versions.
The distributed algorithm uses an embarrassinglyparallel approach. For a forest with N trees being built on w workers, each worker simply builds N/w trees on the data it has available locally. In many cases, partitioning the data so that each worker builds trees on a subset of the total dataset works well, but it generally requires the data to be wellshuffled in advance. Alternatively, callers can replicate all of the data across workers so that rf.fit receives w partitions, each containing the same data. This would produce results approximately identical to singleGPU fitting.
Please check the singleGPU implementation of Random Forest classifier for more information about the underlying algorithm.
 Parameters
 n_estimatorsint (default = 10)
total number of trees in the forest (not perworker)
 handlecuml.Handle
If it is None, a new one is created just for this class.
 split_algoint (default = 1)
0 for HIST, 1 for GLOBAL_QUANTILE The type of algorithm to be used to create the trees.
 split_criterionint (default = 2)
The criterion used to split nodes. 0 for GINI, 1 for ENTROPY, 2 for MSE, 3 for MAE and 4 for CRITERION_END. 0 and 1 not valid for regression
 bootstrapboolean (default = True)
Control bootstrapping. If set, each tree in the forest is built on a bootstrapped sample with replacement. If false, sampling without replacement is done.
 bootstrap_featuresboolean (default = False)
Control bootstrapping for features. If features are drawn with or without replacement
 rows_samplefloat (default = 1.0)
Ratio of dataset rows used while fitting each tree.
 max_depthint (default = 1)
Maximum tree depth. Unlimited (i.e, until leaves are pure), if 1.
 max_leavesint (default = 1)
Maximum leaf nodes per tree. Soft constraint. Unlimited, if 1.
 max_featuresint or float or string or None (default = ‘auto’)
Ratio of number of features (columns) to consider per node split. If int then max_features/n_features. If float then max_features is a fraction. If ‘auto’ then max_features=n_features which is 1.0. If ‘sqrt’ then max_features=1/sqrt(n_features). If ‘log2’ then max_features=log2(n_features)/n_features. If None, then max_features=n_features which is 1.0.
 n_binsint (default = 8)
Number of bins used by the split algorithm.
 min_rows_per_nodeint or float (default = 2)
The minimum number of samples (rows) needed to split a node. If int then number of sample rows If float the min_rows_per_sample*n_rows
 accuracy_metricstring (default = ‘mse’)
Decides the metric used to evaluate the performance of the model. for median of abs error : ‘median_ae’ for mean of abs error : ‘mean_ae’ for mean square error’ : ‘mse’
 n_streamsint (default = 4 )
Number of parallel streams used for forest building
 workersoptional, list of strings
Dask addresses of workers to use for computation. If None, all available Dask workers will be used.
 seedint (default = None)
Base seed for the random number generator. Unseeded by default. Does not currently fully guarantee the exact same results.
Methods
fit
(self, X, y[, convert_dtype])Fit the input data with a Random Forest regression model
get_params
(self[, deep])Returns the value of all parameters required to configure this estimator as a dictionary.
predict
(self, X[, predict_model, algo, …])Predicts the regressor outputs for X.
print_summary
(self)Print the summary of the forest used to train and test the model.
set_params
(self, **params)Sets the value of parameters required to configure this estimator, it functions similar to the sklearn set_params.
predict_model_on_cpu
predict_using_fil

fit
(self, X, y, convert_dtype=False)¶ Fit the input data with a Random Forest regression model
IMPORTANT: X is expected to be partitioned with at least one partition on each Dask worker being used by the forest (self.workers).
When persisting data, you can use cuml.dask.common.utils.persist_across_workers to simplify this:
X_dask_cudf = dask_cudf.from_cudf(X_cudf, npartitions=n_workers) y_dask_cudf = dask_cudf.from_cudf(y_cudf, npartitions=n_workers) X_dask_cudf, y_dask_cudf = persist_across_workers(dask_client, [X_dask_cudf, y_dask_cudf])
 (this is equivalent to calling persist with the data and workers)::
 X_dask_cudf, y_dask_cudf = dask_client.persist([X_dask_cudf,
y_dask_cudf],
workers={ X_dask_cudf=workers, y_dask_cudf=workers })
 Parameters
 XDask cuDF dataframe or CuPy backed Dask Array (n_rows, n_features)
Distributed dense matrix (floats or doubles) of shape (n_samples, n_features).
 yDask cuDF dataframe or CuPy backed Dask Array (n_rows, 1)
Labels of training examples. y must be partitioned the same way as X
 convert_dtypebool, optional (default = False)
When set to True, the fit method will, when necessary, convert y to be the same data type as X if they differ. This will increase memory used for the method.

get_params
(self, deep=True)¶ Returns the value of all parameters required to configure this estimator as a dictionary.
 Parameters
 deepboolean (default = True)

predict
(self, X, predict_model='GPU', algo='auto', convert_dtype=True, fil_sparse_format='auto', delayed=True)¶ Predicts the regressor outputs for X.
GPUbased prediction in a multinode, multiGPU context works by sending the subforest from each worker to the client, concatenating these into one forest with the full n_estimators set of trees, and sending this combined forest to the workers, which will each infer on their local set of data. This allows inference to scale to large datasets, but the forest transmission incurs overheads for very large trees. For inference on small datasets, this overhead may dominate prediction time. Within the worker, this uses the cuML Forest Inference Library (cuml.fil) for highthroughput prediction.
The ‘CPU’ fallback method works with subforests inplace, broadcasting the datasets to all workers and combining predictions via an averaging method at the end. This method is slower on a perrow basis but may be faster for problems with many trees and few rows.
In the 0.15 cuML release, inference will be updated with much faster tree transfer.
 Parameters
 XDask cuDF dataframe or CuPy backed Dask Array (n_rows, n_features)
Distributed dense matrix (floats or doubles) of shape (n_samples, n_features).
 algostring (default = ‘auto’)
This is optional and required only while performing the predict operation on the GPU. ‘naive’  simple inference using shared memory ‘tree_reorg’  similar to naive but trees rearranged to be more coalescingfriendly ‘batch_tree_reorg’  similar to tree_reorg but predicting multiple rows per thread block algo  choose the algorithm automatically. Currently ‘batch_tree_reorg’ is used for dense storage and ‘naive’ for sparse storage
 convert_dtypebool, optional (default = True)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 predict_modelString (default = ‘GPU’)
‘GPU’ to predict using the GPU, ‘CPU’ otherwise. The GPU can only be used if the model was trained on float32 data and X is float32 or convert_dtype is set to True.
 fil_sparse_formatboolean or string (default = auto)
This variable is used to choose the type of forest that will be created in the Forest Inference Library. It is not required while using predict_model=’CPU’. ‘auto’  choose the storage type automatically (currently True is chosen by auto) False  create a dense forest True  create a sparse forest, requires algo=’naive’ or algo=’auto’
 delayedbool (default = True)
Whether to do a lazy prediction (and return Delayed objects) or an eagerly executed one.
 Returns
 yDask cuDF dataframe or CuPy backed Dask Array (n_rows, 1)

print_summary
(self)¶ Print the summary of the forest used to train and test the model.

set_params
(self, **params)¶ Sets the value of parameters required to configure this estimator, it functions similar to the sklearn set_params.
 Parameters
 paramsdict of new params
Truncated SVD¶

class
cuml.dask.decomposition.
TruncatedSVD
(client=None, **kwargs)¶  Parameters
 handlecuml.Handle
If it is None, a new one is created just for this class
 n_componentsint (default = 1)
The number of top K singular vectors / values you want. Must be <= number(columns).
 svd_solver‘full’
Only Full algorithm is supported since it’s significantly faster on GPU then the other solvers including randomized SVD.
 verboseint or boolean (default = False)
Logging level
Examples
from dask_cuda import LocalCUDACluster from dask.distributed import Client, wait import numpy as np from cuml.dask.decomposition import TruncatedSVD from cuml.dask.datasets import make_blobs cluster = LocalCUDACluster(threads_per_worker=1) client = Client(cluster) nrows = 6 ncols = 3 n_parts = 2 X_cudf, _ = make_blobs(nrows, ncols, 1, n_parts, cluster_std=1.8, verbose=cuml.logger.level_info, random_state=10, dtype=np.float32) wait(X_cudf) print("Input Matrix") print(X_cudf.compute()) cumlModel = TruncatedSVD(n_components = 1) XT = cumlModel.fit_transform(X_cudf) print("Transformed Input Matrix") print(XT.compute())
Output:
Input Matrix: 0 1 2 0 8.519647 8.519222 8.865648 1 6.107700 8.350124 10.351215 2 8.026635 9.442240 7.561770 0 8.519647 8.519222 8.865648 1 6.107700 8.350124 10.351215 2 8.026635 9.442240 7.561770 Transformed Input Matrix: 0 0 14.928891 1 14.487295 2 14.431235 0 14.928891 1 14.487295 2 14.431235
 Note: Everytime this code is run, the output will be different because
“make_blobs” function generates random matrices.
 Attributes
 components_array
The top K components (VT.T[:,:n_components]) in U, S, VT = svd(X)
 explained_variance_array
How much each component explains the variance in the data given by S**2
 explained_variance_ratio_array
How much in % the variance is explained given by S**2/sum(S**2)
 singular_values_array
The top K singular values. Remember all singular values >= 0
Methods
fit
(self, X[, _transform])Fit the model with X.
fit_transform
(self, X)Fit the model with X and apply the dimensionality reduction on X.
inverse_transform
(self, X[, delayed])Transform data back to its original space.
transform
(self, X[, delayed])Apply dimensionality reduction to X.
get_param_names

fit
(self, X, _transform=False)¶ Fit the model with X.
 Parameters
 Xdask cuDF input

fit_transform
(self, X)¶ Fit the model with X and apply the dimensionality reduction on X.
 Parameters
 Xdask cuDF
 Returns
 X_newdask cuDF

inverse_transform
(self, X, delayed=True)¶ Transform data back to its original space.
In other words, return an input X_original whose transform would be X.
 Parameters
 Xdask cuDF
 Returns
 X_originaldask cuDF

transform
(self, X, delayed=True)¶ Apply dimensionality reduction to X.
X is projected on the first principal components previously extracted from a training set.
 Parameters
 Xdask cuDF
 Returns
 X_newdask cuDF
Manifold¶

class
cuml.dask.manifold.
UMAP
(model, client=None, **kwargs)¶ Uniform Manifold Approximation and Projection Finds a low dimensional embedding of the data that approximates an underlying manifold.
Adapted from https://github.com/lmcinnes/umap/blob/master/umap/umap.py
Notes
This module is heavily based on Leland McInnes’ reference UMAP package. However, there are a number of differences and features that are not yet implemented in cuml.umap: * Using a nonEuclidean distance metric (support for a fixed set
of nonEuclidean metrics is planned for an upcoming release).
 Using a precomputed pairwise distance matrix (under consideration
for future releases)
Manual initialization of initial embedding positions
In addition to these missing features, you should expect to see the final embeddings differing between cuml.umap and the reference UMAP. In particular, the reference UMAP uses an approximate kNN algorithm for large data sizes while cuml.umap always uses exact kNN.
Known issue: If a UMAP model has not yet been fit, it cannot be pickled
References
Leland McInnes, John Healy, James Melville
UMAP: Uniform Manifold Approximation and Projection for Dimension Reduction https://arxiv.org/abs/1802.03426
Examples
from dask_cuda import LocalCUDACluster from dask.distributed import Client from cuml.dask.datasets import make_blobs from cuml.manifold import UMAP from cuml.dask.manifold import UMAP as MNMG_UMAP import numpy as np cluster = LocalCUDACluster(threads_per_worker=1) client = Client(cluster) X, y = make_blobs(1000, 10, centers=42, cluster_std=0.1, dtype=np.float32, n_parts=2, output='array') local_model = UMAP() selection = np.random.choice(1000, 100) X_train = X[selection].compute() y_train = y[selection].compute() local_model.fit(X_train, y=y_train) distributed_model = MNMG_UMAP(local_model) embedding = distributed_model.transform(X)
 Note: Everytime this code is run, the output will be different because
“make_blobs” function generates random matrices.
Methods
transform
(self, X[, convert_dtype])Transform X into the existing embedded space and return that transformed output.

transform
(self, X, convert_dtype=True)¶ Transform X into the existing embedded space and return that transformed output.
Please refer to the reference UMAP implementation for information on the differences between fit_transform() and running fit() transform().
Specifically, the transform() function is stochastic: https://github.com/lmcinnes/umap/issues/158
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
New data to be transformed. Acceptable formats: dask cuDF, dask CuPy/NumPy/Numba Array
 Returns
 X_newarray, shape (n_samples, n_components)
Embedding of the new data in lowdimensional space.
Linear Models¶

class
cuml.dask.linear_model.
LinearRegression
(client=None, verbose=False, **kwargs)¶ LinearRegression is a simple machine learning model where the response y is modelled by a linear combination of the predictors in X.
cuML’s dask Linear Regression (multinode multigpu) expects dask cuDF DataFrame and provides an algorithms, Eig, to fit a linear model. And provides an eigendecompositionbased algorithm to fit a linear model. (SVD, which is more stable than eig, will be added in an upcoming version.) Eig algorithm is usually preferred when the X is a tall and skinny matrix. As the number of features in X increases, the accuracy of Eig algorithm drops.
This is an experimental implementation of dask Linear Regresion. It supports input X that has more than one column. Single column input X will be supported after SVD algorithm is added in an upcoming version.
 Parameters
 algorithm‘eig’
Eig uses a eigendecomposition of the covariance matrix, and is much faster. SVD is slower, but guaranteed to be stable.
 fit_interceptboolean (default = True)
LinearRegression adds an additional term c to correct for the global mean of y, modeling the reponse as “x * beta + c”. If False, the model expects that you have centered the data.
 normalizeboolean (default = False)
If True, the predictors in X will be normalized by dividing by its L2 norm. If False, no scaling will be done.
 Attributes
 coef_cuDF series, shape (n_features)
The estimated coefficients for the linear regression model.
 intercept_array
The independent term. If fit_intercept_ is False, will be 0.
Methods
fit
(self, X, y)Fit the model with X and y.
predict
(self, X[, delayed])Make predictions for X and returns a dask collection.
get_param_names

fit
(self, X, y)¶ Fit the model with X and y.
 Parameters
 XDask cuDF dataframe or CuPy backed Dask Array (n_rows, n_features)
Features for regression
 yDask cuDF dataframe or CuPy backed Dask Array (n_rows, 1)
Labels (outcome values)

predict
(self, X, delayed=True)¶ Make predictions for X and returns a dask collection.
 Parameters
 XDask cuDF dataframe or CuPy backed Dask Array (n_rows, n_features)
Distributed dense matrix (floats or doubles) of shape (n_samples, n_features).
 delayedbool (default = True)
Whether to do a lazy prediction (and return Delayed objects) or an eagerly executed one.
 Returns
 yDask cuDF dataframe or CuPy backed Dask Array (n_rows, 1)

class
cuml.dask.linear_model.
Ridge
(client=None, verbose=False, **kwargs)¶ Ridge extends LinearRegression by providing L2 regularization on the coefficients when predicting response y with a linear combination of the predictors in X. It can reduce the variance of the predictors, and improves the conditioning of the problem.
cuML’s dask Ridge (multinode multigpu) expects dask cuDF DataFrame and provides an algorithms, Eig, to fit a linear model. And provides an eigendecompositionbased algorithm to fit a linear model. (SVD, which is more stable than eig, will be added in an upcoming version) Eig algorithm is usually preferred when the X is a tall and skinny matrix. As the number of features in X increases, the accuracy of Eig algorithm drops.
This is an experimental implementation of dask Ridge Regresion. It supports input X that has more than one column. Single column input X will be supported after SVD algorithm is added in an upcoming version.
 Parameters
 alphafloat (default = 1.0)
Regularization strength  must be a positive float. Larger values specify stronger regularization. Array input will be supported later.
 solver{‘eig’}
Eig uses a eigendecomposition of the covariance matrix, and is much faster. Other solvers will be supported in the future.
 fit_interceptboolean (default = True)
If True, Ridge adds an additional term c to correct for the global mean of y, modeling the reponse as “x * beta + c”. If False, the model expects that you have centered the data.
 normalizeboolean (default = False)
If True, the predictors in X will be normalized by dividing by it’s L2 norm. If False, no scaling will be done.
 Attributes
 coef_array, shape (n_features)
The estimated coefficients for the linear regression model.
 intercept_array
The independent term. If fit_intercept_ is False, will be 0.
Methods
fit
(self, X, y)Fit the model with X and y.
predict
(self, X[, delayed])Make predictions for X and returns a dask collection.
get_param_names

fit
(self, X, y)¶ Fit the model with X and y.
 Parameters
 XDask cuDF dataframe or CuPy backed Dask Array (n_rows, n_features)
Features for regression
 yDask cuDF dataframe or CuPy backed Dask Array (n_rows, 1)
Labels (outcome values)

predict
(self, X, delayed=True)¶ Make predictions for X and returns a dask collection.
 Parameters
 XDask cuDF dataframe or CuPy backed Dask Array (n_rows, n_features)
Distributed dense matrix (floats or doubles) of shape (n_samples, n_features).
 delayedbool (default = True)
Whether to do a lazy prediction (and return Delayed objects) or an eagerly executed one.
 Returns
 yDask cuDF dataframe or CuPy backed Dask Array (n_rows, 1)

class
cuml.dask.linear_model.
Lasso
(client=None, **kwargs)¶ Lasso extends LinearRegression by providing L1 regularization on the coefficients when predicting response y with a linear combination of the predictors in X. It can zero some of the coefficients for feature selection and improves the conditioning of the problem.
cuML’s Lasso an arraylike object or cuDF DataFrame and uses coordinate descent to fit a linear model.
 Parameters
 alphafloat (default = 1.0)
Constant that multiplies the L1 term. alpha = 0 is equivalent to an ordinary least square, solved by the LinearRegression class. For numerical reasons, using alpha = 0 with the Lasso class is not advised. Given this, you should use the LinearRegression class.
 fit_interceptboolean (default = True)
If True, Lasso tries to correct for the global mean of y. If False, the model expects that you have centered the data.
 normalizeboolean (default = False)
If True, the predictors in X will be normalized by dividing by it’s L2 norm. If False, no scaling will be done.
 max_iterint (default = 1000)
The maximum number of iterations
 tolfloat (default = 1e3)
The tolerance for the optimization: if the updates are smaller than tol, the optimization code checks the dual gap for optimality and continues until it is smaller than tol.
 selection{‘cyclic’, ‘random’} (default=’cyclic’)
If set to ‘random’, a random coefficient is updated every iteration rather than looping over features sequentially by default. This (setting to ‘random’) often leads to significantly faster convergence especially when tol is higher than 1e4.
 Attributes
 coef_array, shape (n_features)
The estimated coefficients for the linear regression model.
 intercept_array
The independent term. If fit_intercept_ is False, will be 0.
 For additional docs, see `scikitlearn’s Lasso
 <https://scikitlearn.org/stable/modules/generated/sklearn.linear_model.Lasso.html>`_.
Methods
fit
(self, X, y)Fit the model with X and y.
predict
(self, X[, delayed])Predicts the y for X.

fit
(self, X, y)¶ Fit the model with X and y.
 Parameters
 XDask cuDF DataFrame or CuPy backed Dask Array
Dense matrix (floats or doubles) of shape (n_samples, n_features).
 yDask cuDF DataFrame or CuPy backed Dask Array
Dense matrix (floats or doubles) of shape (n_samples, n_features).

predict
(self, X, delayed=True)¶ Predicts the y for X.
 Parameters
 XDask cuDF DataFrame or CuPy backed Dask Array
Dense matrix (floats or doubles) of shape (n_samples, n_features).
 delayedbool (default = True)
Whether to do a lazy prediction (and return Delayed objects) or an eagerly executed one.
 Returns
 yDask cuDF DataFrame or CuPy backed Dask Array
Dense matrix (floats or doubles) of shape (n_samples, n_features).

class
cuml.dask.linear_model.
ElasticNet
(client=None, **kwargs)¶ ElasticNet extends LinearRegression with combined L1 and L2 regularizations on the coefficients when predicting response y with a linear combination of the predictors in X. It can reduce the variance of the predictors, force some coefficients to be small, and improves the conditioning of the problem.
cuML’s ElasticNet an arraylike object or cuDF DataFrame, uses coordinate descent to fit a linear model.
 Parameters
 alphafloat (default = 1.0)
Constant that multiplies the L1 term. alpha = 0 is equivalent to an ordinary least square, solved by the LinearRegression object. For numerical reasons, using alpha = 0 with the Lasso object is not advised. Given this, you should use the LinearRegression object.
 l1_ratio: float (default = 0.5)
The ElasticNet mixing parameter, with 0 <= l1_ratio <= 1. For l1_ratio = 0 the penalty is an L2 penalty. For l1_ratio = 1 it is an L1 penalty. For 0 < l1_ratio < 1, the penalty is a combination of L1 and L2.
 fit_interceptboolean (default = True)
If True, Lasso tries to correct for the global mean of y. If False, the model expects that you have centered the data.
 normalizeboolean (default = False)
If True, the predictors in X will be normalized by dividing by it’s L2 norm. If False, no scaling will be done.
 max_iterint (default = 1000)
The maximum number of iterations
 tolfloat (default = 1e3)
The tolerance for the optimization: if the updates are smaller than tol, the optimization code checks the dual gap for optimality and continues until it is smaller than tol.
 selection{‘cyclic’, ‘random’} (default=’cyclic’)
If set to ‘random’, a random coefficient is updated every iteration rather than looping over features sequentially by default. This (setting to ‘random’) often leads to significantly faster convergence especially when tol is higher than 1e4.
 handlecuml.Handle
If it is None, a new one is created just for this class.
 output_type(optional) {‘input’, ‘cudf’, ‘cupy’, ‘numpy’} default = None
Use it to control output type of the results and attributes. If None it’ll inherit the output type set at the module level, cuml.output_type. If that has not been changed, by default the estimator will mirror the type of the data used for each fit or predict call. If set, the estimator will override the global option for its behavior.
 Attributes
 coef_array, shape (n_features)
The estimated coefficients for the linear regression model.
 intercept_array
The independent term. If fit_intercept_ is False, will be 0.
 For additional docs, see `scikitlearn’s ElasticNet
 <https://scikitlearn.org/stable/modules/generated/sklearn.linear_model.ElasticNet.html>`_.
Methods
fit
(self, X, y)Fit the model with X and y.
predict
(self, X[, delayed])Predicts the y for X.

fit
(self, X, y)¶ Fit the model with X and y.
 Parameters
 XDask cuDF DataFrame or CuPy backed Dask Array
Dense matrix (floats or doubles) of shape (n_samples, n_features).
 yDask cuDF DataFrame or CuPy backed Dask Array
Dense matrix (floats or doubles) of shape (n_samples, n_features).

predict
(self, X, delayed=True)¶ Predicts the y for X.
 Parameters
 XDask cuDF DataFrame or CuPy backed Dask Array
Dense matrix (floats or doubles) of shape (n_samples, n_features).
 delayedbool (default = True)
Whether to do a lazy prediction (and return Delayed objects) or an eagerly executed one.
 Returns
 yDask cuDF DataFrame or CuPy backed Dask Array
Dense matrix (floats or doubles) of shape (n_samples, n_features).
Solvers¶

class
cuml.dask.solvers.
CD
(client=None, **kwargs)¶ ModelParallel MultiGPU Linear Regression Model. Single Process Multi GPU supported currently
Methods
fit
(self, X, y)Fit the model with X and y.
predict
(self, X[, delayed])Make predictions for X and returns a dask collection.

fit
(self, X, y)¶ Fit the model with X and y.
 Parameters
 XDask cuDF dataframe or CuPy backed Dask Array (n_rows, n_features)
Features for regression
 yDask cuDF dataframe or CuPy backed Dask Array (n_rows, 1)
Labels (outcome values)

predict
(self, X, delayed=True)¶ Make predictions for X and returns a dask collection.
 Parameters
 XDask cuDF dataframe or CuPy backed Dask Array (n_rows, n_features)
Distributed dense matrix (floats or doubles) of shape (n_samples, n_features).
 delayedbool (default = True)
Whether to do a lazy prediction (and return Delayed objects) or an eagerly executed one.
 Returns
 yDask cuDF dataframe or CuPy backed Dask Array (n_rows, 1)

Dask Base Classes and Mixins¶

class
cuml.dask.common.base.
BaseEstimator
(client=None, verbose=False, **kwargs)¶

class
cuml.dask.common.base.
DelayedParallelFunc
¶

class
cuml.dask.common.base.
DelayedPredictionMixin
¶

class
cuml.dask.common.base.
DelayedTransformMixin
¶

class
cuml.dask.common.base.
DelayedInverseTransformMixin
¶