cuML API Reference

Module Configuration

Output Data Type Configuration

cuml.common.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__ objs	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'>

cuml.common.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__ objs	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.ndarray'>

Verbosity Levels

cuML follows a verbosity model similar to Scikit-learn’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:

Verbosity Levels

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

Feature and Label Encoding (Single-GPU)

class cuml.preprocessing.LabelEncoder.LabelEncoder(*, handle_unknown='error', handle=None, verbose=False, output_type=None)

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.

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

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: object

Methods

fit(y[, _classes])	Fit a LabelEncoder (nvcategory) instance to a set of categories
fit_transform(y[, z])	Simultaneously fit and transform an input
get_param_names(self)	Returns a list of hyperparameter names owned by this class.
inverse_transform(y)	Revert ordinal label to original label
transform(y)	Transform an input into its categorical keys.

fit(y, _classes=None)

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

_classes: int or None.

Passed by the dask client when dask LabelEncoder is used.

Returns

selfLabelEncoder

A fitted instance of itself to allow method chaining

fit_transform(y: cudf.core.series.Series, z=None) → cudf.core.series.Series

Simultaneously fit and transform an input

This is functionally equivalent to (but faster than) LabelEncoder().fit(y).transform(y)

get_param_names(self)

Returns a list of hyperparameter names owned by this class. It is expected that every child class overrides this method and appends its extra set of parameters that it in-turn owns. This is to simplify the implementation of get_params and set_params methods.

inverse_transform(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(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, handle=None, verbose=False, output_type=None)

A multi-class dummy encoder for labels.

Parameters

neg_labelinteger

label to be used as the negative binary label

pos_labelinteger

label to be used as the positive binary label

sparse_outputbool

whether to return sparse arrays for transformed output

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

Examples

Create an array with labels and dummy encode them

import cupy as cp
import cupyx
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]

Attributes

classes_

Python descriptor object to control getting/setting CumlArray attributes on Base objects. See the Estimator Guide for an in depth guide.

Methods

fit(y)	Fit label binarizer
fit_transform(y)	Fit label binarizer and transform multi-class labels to their dummy-encoded representation.
get_param_names(self)	Returns a list of hyperparameter names owned by this class.
inverse_transform(y[, threshold])	Transform binary labels back to original multi-class labels
transform(y)	Transform multi-class labels to their dummy-encoded representation labels.

fit(y) → cuml.preprocessing.label.LabelBinarizer

Fit label binarizer

Parameters

yarray of shape [n_samples,] or [n_samples, n_classes]

Target values. The 2-d matrix should only contain 0 and 1, represents multilabel classification.

Returns

selfreturns an instance of self.

fit_transform(y) → cuml.common.array_sparse.SparseCumlArray

Fit label binarizer and transform multi-class labels to their dummy-encoded representation.

Parameters

yarray of shape [n_samples,] or [n_samples, n_classes]

Returns

arrarray with encoded labels

get_param_names(self)

Returns a list of hyperparameter names owned by this class. It is expected that every child class overrides this method and appends its extra set of parameters that it in-turn owns. This is to simplify the implementation of get_params and set_params methods.

inverse_transform(y, threshold=None) → cuml.common.array.CumlArray

Transform binary labels back to original multi-class labels

Parameters

yarray of shape [n_samples, n_classes]

thresholdfloat this value is currently ignored

Returns

arrarray with original labels

transform(y) → cuml.common.array_sparse.SparseCumlArray

Transform multi-class labels to their dummy-encoded representation labels.

Parameters

yarray of shape [n_samples,] or [n_samples, n_classes]

Returns

arrarray with encoded labels

cuml.preprocessing.label_binarize(y, classes, neg_label=0, pos_label=1, sparse_output=False) → cuml.common.array_sparse.SparseCumlArray

A stateless helper function to dummy encode multi-class labels.

Parameters

yarray-like 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', handle=None, verbose=False, output_type=None)

Encode categorical features as a one-hot 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 one-hot (aka ‘one-of-K’ 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 one-hot 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=True

This feature is 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 one-hot encoded columns for this feature will be all zeros. In the inverse transform, an unknown category will be denoted as None.

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

Attributes

drop_idx_array of shape (n_features,)

drop_idx_[i] is the index in categories_[i] of the category to be dropped for each feature. None if all the transformed features will be retained.

Methods

fit(X)	Fit OneHotEncoder to X.
fit_transform(X[, y])	Fit OneHotEncoder to X, then transform X.
get_param_names(self)	Returns a list of hyperparameter names owned by this class.
inverse_transform(X)	Convert the data back to the original representation.
transform(X)	Transform X using one-hot encoding.

property categories_

Returns categories used for the one hot encoding in the correct order.

fit(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(X, y=None)

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 2-d array

Transformed input.

get_param_names(self)

Returns a list of hyperparameter names owned by this class. It is expected that every child class overrides this method and appends its extra set of parameters that it in-turn owns. This is to simplify the implementation of get_params and set_params methods.

inverse_transform(X)

Convert the data back to the original representation. In case unknown categories are encountered (all zeros in the one-hot 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

Xarray-like or sparse matrix, shape [n_samples, n_encoded_features]

The transformed data.

Returns

X_trcudf.DataFrame or cupy.ndarray

Inverse transformed array.

transform(X)

Transform X using one-hot encoding.

Parameters

Xcudf.DataFrame or cupy.ndarray

The data to encode.

Returns

X_outsparse matrix if sparse=True else a 2-d array

Transformed input.

class cuml.preprocessing.TargetEncoder.TargetEncoder(n_folds=4, smooth=0, seed=42, split_method='interleaved', output_type='auto')

A cudf based implementation of target encoding [1], which converts one or mulitple categorical variables, ‘Xs’, with the average of corresponding values of the target variable, ‘Y’. The input data is grouped by the columns Xs and the aggregated mean value of Y of each group is calculated to replace each value of Xs. Several optimizations are applied to prevent label leakage and parallelize the execution.

Parameters

n_foldsint (default=4)

Default number of folds for fitting training data. To prevent label leakage in fit, we split data into n_folds and encode one fold using the target variables of the remaining folds.

smoothint or float (default=0)

Count of samples to smooth the encoding. 0 means no smoothing.

seedint (default=42)

Random seed

split_method{‘random’, ‘continuous’, ‘interleaved’},

default=’interleaved’ Method to split train data into n_folds. ‘random’: random split. ‘continuous’: consecutive samples are grouped into one folds. ‘interleaved’: samples are assign to each fold in a round robin way.

output_type: {‘cupy’, ‘numpy’, ‘auto’}, default = ‘auto’

The data type of output. If ‘auto’, it matches input data.

References

1

https://maxhalford.github.io/blog/target-encoding/

Examples

Converting a categorical implementation to a numerical one

from cudf import DataFrame, Series

train = DataFrame({'category': ['a', 'b', 'b', 'a'],
                   'label': [1, 0, 1, 1]})
test = DataFrame({'category': ['a', 'c', 'b', 'a']})

encoder = TargetEncoder()
train_encoded = encoder.fit_transform(train.category, train.label)
test_encoded = encoder.transform(test.category)
print(train_encoded)
print(test_encoded)

Output:

[1. 1. 0. 1.]
[1.   0.75 0.5  1.  ]

Methods

fit(x, y)	Fit a TargetEncoder instance to a set of categories
fit_transform(x, y)	Simultaneously fit and transform an input
transform(x)	Transform an input into its categorical keys.

fit(x, y)

Fit a TargetEncoder instance to a set of categories

Parameters

x: cudf.Series or cudf.DataFrame or cupy.ndarray

categories to be encoded. It’s elements may or may not be unique

ycudf.Series or cupy.ndarray

Series containing the target variable.

Returns

selfTargetEncoder

A fitted instance of itself to allow method chaining

fit_transform(x, y)

Simultaneously fit and transform an input

This is functionally equivalent to (but faster than) TargetEncoder().fit(y).transform(y)

transform(x)

Transform an input into its categorical keys.

This is intended for test data. For fitting and transforming the training data, prefer fit_transform.

Parameters

xcudf.Series

Input keys to be transformed. Its values doesn’t have to match the categories given to fit

Returns

encodedcupy.ndarray

The ordinally encoded input series

Text Preprocessing (Single-GPU)

class cuml.preprocessing.text.stem.PorterStemmer(mode='NLTK_EXTENSIONS')

A word stemmer based on the Porter stemming algorithm.

Porter, M. “An algorithm for suffix stripping.” Program 14.3 (1980): 130-137.

See http://www.tartarus.org/~martin/PorterStemmer/ for the homepage of the algorithm.

Martin Porter has endorsed several modifications to the Porter algorithm since writing his original paper, and those extensions are included in the implementations on his website. Additionally, others have proposed further improvements to the algorithm, including NLTK contributors. Only below mode is supported currently PorterStemmer.NLTK_EXTENSIONS

• Implementation that includes further improvements devised by NLTK contributors or taken from other modified implementations found on the web.

Parameters

mode: Modes of stemming (Only supports (NLTK_EXTENSIONS) currently)

default(“NLTK_EXTENSIONS”)

Examples

import cudf
from cuml.preprocessing.text.stem import PorterStemmer
stemmer = PorterStemmer()
word_str_ser =  cudf.Series(['revival','singing','adjustable'])
print(stemmer.stem(word_str_ser))

Output:

0     reviv
1      sing
2    adjust
dtype: object

Methods

stem(word_str_ser)

Stem Words using Porter stemmer

stem(word_str_ser)

Stem Words using Porter stemmer

Parameters

word_str_sercudf.Series

A string series of words to stem

Returns

stemmed_sercudf.Series

Stemmed words strings series

Feature and Label Encoding (Dask-based Multi-GPU)

class cuml.dask.preprocessing.LabelBinarizer(*, client=None, **kwargs)

A distributed version of LabelBinarizer for one-hot encoding a collection of labels.

Examples

Create an array with labels and dummy encode them

import cupy as cp
import cupyx
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(y)	Fit label binarizer
fit_transform(y)	Fit the label encoder and return transformed labels
inverse_transform(y[, threshold])	Invert a set of encoded labels back to original labels
transform(y)	Transform and return encoded labels

fit(y)

Fit label binarizer

Parameters

yDask.Array of shape [n_samples,] or [n_samples, n_classes]

chunked by row. Target values. The 2-d matrix should only contain 0 and 1, represents multilabel classification.

Returns

selfreturns an instance of self.

fit_transform(y)

Fit the label encoder and return transformed labels

Parameters

yDask.Array of shape [n_samples,] or [n_samples, n_classes]

target values. The 2-d matrix should only contain 0 and 1, represents multilabel classification.

Returns

arrDask.Array backed by CuPy arrays containing encoded labels

inverse_transform(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(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 one-hot 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 one-hot (aka ‘one-of-K’ 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 one-hot encoded columns for this feature will be all zeros. In the inverse transform, an unknown category will be denoted as None.

Methods

fit(X)	Fit a multi-node multi-gpu OneHotEncoder to X.
fit_transform(X[, delayed])	Fit OneHotEncoder to X, then transform X.
inverse_transform(X[, delayed])	Convert the data back to the original representation.
transform(X[, delayed])	Transform X using one-hot encoding.

fit(X)

Fit a multi-node multi-gpu 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(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(X, delayed=True)

Convert the data back to the original representation. In case unknown categories are encountered (all zeros in the one-hot 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(X, delayed=True)

Transform X using one-hot 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.

Feature Extraction (Single-GPU)

class cuml.feature_extraction.text.CountVectorizer(input=None, encoding=None, decode_error=None, strip_accents=None, lowercase=True, preprocessor=None, tokenizer=None, stop_words=None, token_pattern=None, ngram_range=(1, 1), analyzer='word', max_df=1.0, min_df=1, max_features=None, vocabulary=None, binary=False, dtype=<class 'numpy.float32'>, delimiter=' ')

Convert a collection of text documents to a matrix of token counts

If you do not provide an a-priori dictionary then the number of features will be equal to the vocabulary size found by analyzing the data.

Parameters

lowercaseboolean, True by default

Convert all characters to lowercase before tokenizing.

preprocessorcallable or None (default)

Override the preprocessing (string transformation) stage while preserving the tokenizing and n-grams generation steps.

stop_wordsstring {‘english’}, list, or None (default)

If ‘english’, a built-in stop word list for English is used. If a list, that list is assumed to contain stop words, all of which will be removed from the input documents. If None, no stop words will be used. max_df can be set to a value to automatically detect and filter stop words based on intra corpus document frequency of terms.

ngram_rangetuple (min_n, max_n), default=(1, 1)

The lower and upper boundary of the range of n-values for different word n-grams or char n-grams to be extracted. All values of n such such that min_n <= n <= max_n will be used. For example an ngram_range of (1, 1) means only unigrams, (1, 2) means unigrams and bigrams, and (2, 2) means only bigrams.

analyzerstring, {‘word’, ‘char’, ‘char_wb’}

Whether the feature should be made of word n-gram or character n-grams. Option ‘char_wb’ creates character n-grams only from text inside word boundaries; n-grams at the edges of words are padded with space.

max_dffloat in range [0.0, 1.0] or int, default=1.0

When building the vocabulary ignore terms that have a document frequency strictly higher than the given threshold (corpus-specific stop words). If float, the parameter represents a proportion of documents, integer absolute counts. This parameter is ignored if vocabulary is not None.

min_dffloat in range [0.0, 1.0] or int, default=1

When building the vocabulary ignore terms that have a document frequency strictly lower than the given threshold. This value is also called cut-off in the literature. If float, the parameter represents a proportion of documents, integer absolute counts. This parameter is ignored if vocabulary is not None.

max_featuresint or None, default=None

If not None, build a vocabulary that only consider the top max_features ordered by term frequency across the corpus. This parameter is ignored if vocabulary is not None.

vocabularycudf.Series, optional

If not given, a vocabulary is determined from the input documents.

binaryboolean, default=False

If True, all non zero counts are set to 1. This is useful for discrete probabilistic models that model binary events rather than integer counts.

dtypetype, optional

Type of the matrix returned by fit_transform() or transform().

delimiterstr, whitespace by default

String used as a replacement for stop words if stop_words is not None. Typically the delimiting character between words is a good choice.

Attributes

vocabulary_cudf.Series[str]

Array mapping from feature integer indices to feature name.

stop_words_cudf.Series[str]

Terms that were ignored because they either:

• occurred in too many documents (max_df)

• occurred in too few documents (min_df)

• were cut off by feature selection (max_features).

This is only available if no vocabulary was given.

Methods

fit(raw_documents)	Build a vocabulary of all tokens in the raw documents.
fit_transform(raw_documents)	Build the vocabulary and return document-term matrix.
get_feature_names()	Array mapping from feature integer indices to feature name.
inverse_transform(X)	Return terms per document with nonzero entries in X.
transform(raw_documents)	Transform documents to document-term matrix.

fit(raw_documents)

Build a vocabulary of all tokens in the raw documents.

Parameters

raw_documentscudf.Series

A Series of string documents

Returns

self

fit_transform(raw_documents)

Build the vocabulary and return document-term matrix.

Equivalent to self.fit(X).transform(X) but preprocess X only once.

Parameters

raw_documentscudf.Series

A Series of string documents

Returns

Xcupy csr array of shape (n_samples, n_features)

Document-term matrix.

get_feature_names()

Array mapping from feature integer indices to feature name.

Returns

feature_namesSeries

A list of feature names.

inverse_transform(X)

Return terms per document with nonzero entries in X.

Parameters

Xarray-like of shape (n_samples, n_features)

Document-term matrix.

Returns

X_invlist of cudf.Series of shape (n_samples,)

List of Series of terms.

transform(raw_documents)

Transform documents to document-term matrix.

Extract token counts out of raw text documents using the vocabulary fitted with fit or the one provided to the constructor.

Parameters

raw_documentscudf.Series

A Series of string documents

Returns

Xcupy csr array of shape (n_samples, n_features)

Document-term matrix.

class cuml.feature_extraction.text.HashingVectorizer(input=None, encoding=None, decode_error=None, strip_accents=None, lowercase=True, preprocessor=None, tokenizer=None, stop_words=None, token_pattern=None, ngram_range=(1, 1), analyzer='word', n_features=1048576, binary=False, norm='l2', alternate_sign=True, dtype=<class 'numpy.float32'>, delimiter=' ')

Convert a collection of text documents to a matrix of token occurrences

It turns a collection of text documents into a cupyx.scipy.sparse matrix holding token occurrence counts (or binary occurrence information), possibly normalized as token frequencies if norm=’l1’ or projected on the euclidean unit sphere if norm=’l2’.

This text vectorizer implementation uses the hashing trick to find the token string name to feature integer index mapping.

This strategy has several advantages:

• it is very low memory scalable to large datasets as there is no need to store a vocabulary dictionary in memory which is even more important as GPU’s that are often memory constrained

• it is fast to pickle and un-pickle as it holds no state besides the constructor parameters

• it can be used in a streaming (partial fit) or parallel pipeline as there is no state computed during fit.

There are also a couple of cons (vs using a CountVectorizer with an in-memory vocabulary):

• there is no way to compute the inverse transform (from feature indices to string feature names) which can be a problem when trying to introspect which features are most important to a model.

• there can be collisions: distinct tokens can be mapped to the same feature index. However in practice this is rarely an issue if n_features is large enough (e.g. 2 ** 18 for text classification problems).

• no IDF weighting as this would render the transformer stateful.

The hash function employed is the signed 32-bit version of Murmurhash3.

Parameters

lowercasebool, default=True

Convert all characters to lowercase before tokenizing.

preprocessorcallable or None (default)

Override the preprocessing (string transformation) stage while preserving the tokenizing and n-grams generation steps.

stop_wordsstring {‘english’}, list, default=None

If ‘english’, a built-in stop word list for English is used. There are several known issues with ‘english’ and you should consider an alternative. If a list, that list is assumed to contain stop words, all of which will be removed from the resulting tokens. Only applies if analyzer == 'word'.

ngram_rangetuple (min_n, max_n), default=(1, 1)

The lower and upper boundary of the range of n-values for different word n-grams or char n-grams to be extracted. All values of n such such that min_n <= n <= max_n will be used. For example an ngram_range of (1, 1) means only unigrams, (1, 2) means unigrams and bigrams, and (2, 2) means only bigrams.

analyzerstring, {‘word’, ‘char’, ‘char_wb’}

Whether the feature should be made of word n-gram or character n-grams. Option ‘char_wb’ creates character n-grams only from text inside word boundaries; n-grams at the edges of words are padded with space.

n_featuresint, default=(2 ** 20)

The number of features (columns) in the output matrices. Small numbers of features are likely to cause hash collisions, but large numbers will cause larger coefficient dimensions in linear learners.

binarybool, default=False.

If True, all non zero counts are set to 1. This is useful for discrete probabilistic models that model binary events rather than integer counts.

norm{‘l1’, ‘l2’}, default=’l2’

Norm used to normalize term vectors. None for no normalization.

alternate_signbool, default=True

When True, an alternating sign is added to the features as to approximately conserve the inner product in the hashed space even for small n_features. This approach is similar to sparse random projection.

dtypetype, optional

Type of the matrix returned by fit_transform() or transform().

delimiterstr, whitespace by default

String used as a replacement for stop words if stop_words is not None. Typically the delimiting character between words is a good choice.

See also

CountVectorizer, TfidfVectorizer

Examples

from cuml.feature_extraction.text import HashingVectorizer
corpus = [
    'This is the first document.',
    'This document is the second document.',
    'And this is the third one.',
    'Is this the first document?',
]
vectorizer = HashingVectorizer(n_features=2**4)
X = vectorizer.fit_transform(corpus)
print(X.shape)

Output:

(4, 16)

Methods

fit(X[, y])	This method only checks the input type and the model parameter.
fit_transform(X[, y])	Transform a sequence of documents to a document-term matrix.
partial_fit(X[, y])	Does nothing: This transformer is stateless This method is just there to mark the fact that this transformer can work in a streaming setup.
transform(raw_documents)	Transform documents to document-term matrix.

fit(X, y=None)

This method only checks the input type and the model parameter. It does not do anything meaningful as this transformer is stateless

Parameters

Xcudf.Series

A Series of string documents

fit_transform(X, y=None)

Transform a sequence of documents to a document-term matrix.

Parameters

Xiterable over raw text documents, length = n_samples

Samples. Each sample must be a text document (either bytes or unicode strings, file name or file object depending on the constructor argument) which will be tokenized and hashed.

yany

Ignored. This parameter exists only for compatibility with sklearn.pipeline.Pipeline.

Returns

Xsparse CuPy CSR matrix of shape (n_samples, n_features)

Document-term matrix.

partial_fit(X, y=None)

Does nothing: This transformer is stateless This method is just there to mark the fact that this transformer can work in a streaming setup.

Parameters

Xcudf.Series(A Series of string documents).

transform(raw_documents)

Transform documents to document-term matrix.

Extract token counts out of raw text documents using the vocabulary fitted with fit or the one provided to the constructor.

Parameters

raw_documentscudf.Series

A Series of string documents

Returns

Xsparse CuPy CSR matrix of shape (n_samples, n_features)

Document-term matrix.

class cuml.feature_extraction.text.TfidfVectorizer(input=None, encoding=None, decode_error=None, strip_accents=None, lowercase=True, preprocessor=None, tokenizer=None, stop_words=None, token_pattern=None, ngram_range=(1, 1), analyzer='word', max_df=1.0, min_df=1, max_features=None, vocabulary=None, binary=False, dtype=<class 'numpy.float32'>, delimiter=' ', norm='l2', use_idf=True, smooth_idf=True, sublinear_tf=False)

Convert a collection of raw documents to a matrix of TF-IDF features.

Equivalent to CountVectorizer followed by TfidfTransformer.

Parameters

lowercaseboolean, True by default

Convert all characters to lowercase before tokenizing.

preprocessorcallable or None (default)

Override the preprocessing (string transformation) stage while preserving the tokenizing and n-grams generation steps.

stop_wordsstring {‘english’}, list, or None (default)

If ‘english’, a built-in stop word list for English is used. If a list, that list is assumed to contain stop words, all of which will be removed from the input documents. If None, no stop words will be used. max_df can be set to a value to automatically detect and filter stop words based on intra corpus document frequency of terms.

ngram_rangetuple (min_n, max_n), default=(1, 1)

The lower and upper boundary of the range of n-values for different word n-grams or char n-grams to be extracted. All values of n such such that min_n <= n <= max_n will be used. For example an ngram_range of (1, 1) means only unigrams, (1, 2) means unigrams and bigrams, and (2, 2) means only bigrams.

analyzerstring, {‘word’, ‘char’, ‘char_wb’}

Whether the feature should be made of word n-gram or character n-grams. Option ‘char_wb’ creates character n-grams only from text inside word boundaries; n-grams at the edges of words are padded with space.

max_dffloat in range [0.0, 1.0] or int, default=1.0

When building the vocabulary ignore terms that have a document frequency strictly higher than the given threshold (corpus-specific stop words). If float, the parameter represents a proportion of documents, integer absolute counts. This parameter is ignored if vocabulary is not None.

min_dffloat in range [0.0, 1.0] or int, default=1

When building the vocabulary ignore terms that have a document frequency strictly lower than the given threshold. This value is also called cut-off in the literature. If float, the parameter represents a proportion of documents, integer absolute counts. This parameter is ignored if vocabulary is not None.

max_featuresint or None, default=None

If not None, build a vocabulary that only consider the top max_features ordered by term frequency across the corpus. This parameter is ignored if vocabulary is not None.

vocabularycudf.Series, optional

If not given, a vocabulary is determined from the input documents.

binaryboolean, default=False

If True, all non zero counts are set to 1. This is useful for discrete probabilistic models that model binary events rather than integer counts.

dtypetype, optional

Type of the matrix returned by fit_transform() or transform().

delimiterstr, whitespace by default

String used as a replacement for stop words if stop_words is not None. Typically the delimiting character between words is a good choice.

norm{‘l1’, ‘l2’}, default=’l2’

Each output row will have unit norm, either:

• ‘l2’: Sum of squares of vector elements is 1. The cosine similarity between two vectors is their dot product when l2 norm has been applied.

• ‘l1’: Sum of absolute values of vector elements is 1.

use_idfbool, default=True

Enable inverse-document-frequency reweighting.

smooth_idfbool, default=True

Smooth idf weights by adding one to document frequencies, as if an extra document was seen containing every term in the collection exactly once. Prevents zero divisions.

sublinear_tfbool, default=False

Apply sublinear tf scaling, i.e. replace tf with 1 + log(tf).

Notes

The stop_words_ attribute can get large and increase the model size when pickling. This attribute is provided only for introspection and can be safely removed using delattr or set to None before pickling.

This class is largely based on scikit-learn 0.23.1’s TfIdfVectorizer code, which is provided under the BSD-3 license.

Attributes

idf_array of shape (n_features)

The inverse document frequency (IDF) vector; only defined if use_idf is True.

vocabulary_cudf.Series[str]

Array mapping from feature integer indices to feature name.

stop_words_cudf.Series[str]

Terms that were ignored because they either:

• occurred in too many documents (max_df)

• occurred in too few documents (min_df)

• were cut off by feature selection (max_features).

This is only available if no vocabulary was given.

Methods

fit(raw_documents)	Learn vocabulary and idf from training set.
fit_transform(raw_documents)	Learn vocabulary and idf, return document-term matrix.
transform(raw_documents)	Transform documents to document-term matrix.

fit(raw_documents)

Learn vocabulary and idf from training set.

Parameters

raw_documentscudf.Series

A Series of string documents

Returns

selfobject

Fitted vectorizer.

fit_transform(raw_documents)

Learn vocabulary and idf, return document-term matrix. This is equivalent to fit followed by transform, but more efficiently implemented.

Parameters

raw_documentscudf.Series

A Series of string documents

Returns

Xcupy csr array of shape (n_samples, n_features)

Tf-idf-weighted document-term matrix.

transform(raw_documents)

Transform documents to document-term matrix. Uses the vocabulary and document frequencies (df) learned by fit (or fit_transform).

Parameters

raw_documentscudf.Series

A Series of string documents

Returns

Xcupy csr array of shape (n_samples, n_features)

Tf-idf-weighted document-term matrix.

Feature Extraction (Dask-based Multi-GPU)

class cuml.dask.feature_extraction.text.TfidfTransformer(*, client=None, verbose=False, **kwargs)

Distributed TF-IDF transformer

Examples

import cupy as cp
from sklearn.datasets import fetch_20newsgroups
from sklearn.feature_extraction.text import CountVectorizer
from dask_cuda import LocalCUDACluster
from dask.distributed import Client
from cuml.dask.common import to_sparse_dask_array
from cuml.dask.naive_bayes import MultinomialNB
import dask
from cuml.dask.feature_extraction.text import TfidfTransformer

# Create a local CUDA cluster
cluster = LocalCUDACluster()
client = Client(cluster)

# Load corpus
twenty_train = fetch_20newsgroups(subset='train',
                        shuffle=True, random_state=42)
cv = CountVectorizer()
xformed = cv.fit_transform(twenty_train.data).astype(cp.float32)
X = to_sparse_dask_array(xformed, client)

y = dask.array.from_array(twenty_train.target, asarray=False,
                    fancy=False).astype(cp.int32)

mutli_gpu_transformer = TfidfTransformer()
X_transormed = mutli_gpu_transformer.fit_transform(X)
X_transormed.compute_chunk_sizes()

model = MultinomialNB()
model.fit(X_transormed, y)
model.score(X_transormed, y)

Output:

array(0.93264981)

Methods

fit(X)	Fit distributed TFIDF Transformer
fit_transform(X)	Fit distributed TFIDFTransformer and then transform the given set of data samples.
transform(X)	Use distributed TFIDFTransformer to transform the given set of data samples.

fit(X)

Fit distributed TFIDF Transformer

Parameters

Xdask.Array with blocks containing dense or sparse cupy arrays

Returns

cuml.dask.feature_extraction.text.TfidfTransformer instance

fit_transform(X)

Fit distributed TFIDFTransformer and then transform the given set of data samples.

Parameters

Xdask.Array with blocks containing dense or sparse cupy arrays

Returns

dask.Array with blocks containing transformed sparse cupy arrays

transform(X)

Use distributed TFIDFTransformer to transform the given set of data samples.

Parameters

Xdask.Array with blocks containing dense or sparse cupy arrays

Returns

dask.Array with blocks containing transformed sparse cupy arrays

Dataset Generation (Single-GPU)

random_state

Determines random number generation for dataset creation. Pass an int for reproducible output across multiple function calls.

cuml.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 array-like, optional (default=100)

If int, it is the total number of points equally divided among clusters. If array-like, each element of the sequence indicates the number of samples per cluster.

n_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 array-like, 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])

cuml.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 n-class classification problem. This initially creates clusters of points normally distributed (std=1) about vertices of an n_informative-dimensional hypercube with sides of length 2*class_sep and assigns an equal number of clusters to each class. It introduces interdependence between these features and adds various types of further noise to the data. Without shuffling, X horizontally stacks features in the following order: the primary n_informative features, followed by n_redundant linear combinations of the informative features, followed by n_repeated duplicates, drawn randomly with replacement from the informative and redundant features. The remaining features are filled with random noise. Thus, without shuffling, all useful features are contained in the columns X[:, :n_informative + n_redundant + n_repeated].

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 and n_features-n_informative-n_redundant-n_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.

weightsarray-like of shape (n_classes,) or (n_classes - 1,), (default=None)

The proportions of samples assigned to each class. If None, then classes are balanced. Note that if len(weights) == n_classes - 1, then the last class weight is automatically inferred. More than n_samples samples may be returned if the sum of weights exceeds 1.

flip_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:

1. Firstly, we generate X from a standard univariate instead of zeros. This saves memory as we don’t need to generate univariates each time for each feature class (informative, repeated, etc.) while also providing the added speedup of generating a big matrix on GPU

2. We generate order=F construction. We exploit the fact that X is a generated from a univariate normal, and covariance is introduced with matrix multiplications. Which means, we can generate X as a 1D array and just reshape it to the desired order, which only updates the metadata and eliminates copies

3. Lastly, we also shuffle by construction. Centroid indices are permuted for each sample, and then we construct the data for each centroid. This shuffle works for both order=C and order=F and eliminates any need for secondary copies

References

1

I. Guyon, “Design of experiments for the NIPS 2003 variable selection benchmark”, 2003.

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]

cuml.datasets.make_regression(n_samples=100, n_features=2, n_informative=2, n_targets=1, bias=0.0, effective_rank=None, tail_strength=0.5, noise=0.0, shuffle=True, coef=False, random_state=None, dtype='single', handle=None) → typing.Union[typing.Tuple[CumlArray, CumlArray], typing.Tuple[CumlArray, CumlArray, CumlArray]]

Generate a random regression problem.

See https://scikit-learn.org/stable/modules/generated/sklearn.datasets.make_regression.html # noqa: E501

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.

Examples

from cuml.datasets.regression import make_regression
from cuml.linear_model import LinearRegression

# Create regression problem
data, values = make_regression(n_samples=200, n_features=12,
                               n_informative=7, bias=-4.2, noise=0.3)

# Perform a linear regression on this problem
lr = LinearRegression(fit_intercept = True, normalize = False,
                      algorithm = "eig")
reg = lr.fit(data, values)
print(reg.coef_)

cuml.datasets.make_arima(batch_size=1000, n_obs=100, order=(1, 1, 1), seasonal_order=(0, 0, 0, 0), intercept=False, random_state=None, dtype='double', handle=None)

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’

handle: cuml.Handle

If it is None, a new one is created just for this function call

Returns

out: array-like, shape (n_obs, batch_size)

Array of the requested type containing the generated dataset

Examples

from cuml.datasets import make_arima
y = make_arima(1000, 100, (2,1,2), (0,1,2,12), 0)

Dataset Generation (Dask-based Multi-GPU)

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 Dask-Cupy arrays containing blobs for a randomly generated set of centroids.

This function calls make_blobs from cuml.datasets on each Dask worker and aggregates them into a single Dask Dataframe.

For more information on Scikit-learn’s make_blobs:.

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 array-like, 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 n-class classification problem.

This initially creates clusters of points normally distributed (std=1) about vertices of an n_informative-dimensional hypercube with sides of length 2 * class_sep and assigns an equal number of clusters to each class. It introduces interdependence between these features and adds various types of further noise to the data.

Without shuffling, X horizontally stacks features in the following order: the primary n_informative features, followed by n_redundant linear combinations of the informative features, followed by n_repeated duplicates, drawn randomly with replacement from the informative and redundant features. The remaining features are filled with random noise. Thus, without shuffling, all useful features are contained in the columns X[:, :n_informative + n_redundant + n_repeated].

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 and n_features-n_informative-n_redundant-n_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.

weightsarray-like of shape (n_classes,) or (n_classes - 1,) , (default=None)

The proportions of samples assigned to each class. If None, then classes are balanced. Note that if len(weights) == n_classes - 1, then the last class weight is automatically inferred. More than n_samples samples may be returned if the sum of weights exceeds 1.

flip_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:

1. We generate centroids of shape (n_centroids, n_informative)

2. We generate an informative covariance of shape (n_centroids, n_informative, n_informative)

3. We generate a redundant covariance of shape (n_informative, n_redundant)

4. We generate the indices for the repeated features We pass along the references to the futures of the above arrays with each part to the single GPU cuml.datasets.classification.make_classification so that each part (and worker) has access to the correct values to generate data from the same covariances

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 bell-shaped 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

XDask-CuPy 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 rank-fat tail singular profile.

The output is generated by applying a (potentially biased) random linear regression model with “n_informative” nonzero regressors to the previously generated input and some gaussian centered noise with some adjustable scale.

Parameters

n_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’)

Row-major or Col-major

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

XDask-CuPy array of shape [n_samples, n_features]

The input samples.

yDask-CuPy array of shape [n_samples] or [n_samples, n_targets]

The output values.

coefDask-CuPy array of shape [n_features] or [n_features, n_targets], optional

The coefficient of the underlying linear model. It is returned only if coef is True.

Notes

Known Performance Limitations:

1. When effective_rank is set and use_full_low_rank is True, we cannot generate order F by construction, and an explicit transpose is performed on each part. This may cause memory to spike (other parameters make order F by construction)

2. When n_targets > 1 and order = 'F' as above, we have to explicity transpose the y array. If coef = True, then we also explicity transpose the ground_truth array

3. When shuffle = True and order = F, there are memory spikes to shuffle the F order arrays

Note

If out-of-memory errors are encountered in any of the above configurations, try increasing the n_parts parameter.

Array Wrappers (Internal API)

class cuml.common.CumlArray(data=None, owner=None, dtype=None, shape=None, order=None)

Array represents an abstracted array allocation. It can be instantiated by itself, creating an rmm.DeviceBuffer underneath, or can be instantiated by __cuda_array_interface__ or __array_interface__ compliant arrays, in which case it’ll keep a reference to that data underneath. Also can be created from a pointer, specifying the characteristics of the array, in that case the owner of the data referred to by the pointer should be specified explicitly.

Parameters

datarmm.DeviceBuffer, cudf.Buffer, array_like, int, bytes, bytearray or memoryview

An array-like object or integer representing a device or host pointer to pre-allocated memory.

ownerobject, optional

Python object to which the lifetime of the memory allocation is tied. If provided, a reference to this object is kept in this Buffer.

dtypedata-type, optional

Any object that can be interpreted as a numpy or cupy data type.

shapeint or tuple of ints, optional

Shape of created array.

order: string, optional

Whether to create a F-major or C-major array.

Notes

cuml Array is not meant as an end-user array library. It is meant for cuML/RAPIDS developer consumption. Therefore it contains the minimum functionality. Its functionality is hidden by base.pyx to provide automatic output format conversion so that the users see the important attributes in whatever format they prefer.

Todo: support cuda streams in the constructor. See: https://github.com/rapidsai/cuml/issues/1712 https://github.com/rapidsai/cuml/pull/1396

Attributes

ptrint

Pointer to the data

sizeint

Size of the array data in bytes

_ownerPython Object

Object that owns the data of the array

shapetuple of ints

Shape of the array

order{‘F’, ‘C’}

‘F’ or ‘C’ to indicate Fortran-major or C-major order of the array

stridestuple of ints

Strides of the data

__cuda_array_interface__dictionary

__cuda_array_interface__ to interop with other libraries.

Methods

empty(shape, dtype[, order])	Create an empty Array with an allocated but uninitialized DeviceBuffer
full(shape, value, dtype[, order])	Create an Array with an allocated DeviceBuffer initialized to value.
ones(shape[, dtype, order])	Create an Array with an allocated DeviceBuffer initialized to zeros.
serialize()	Generate an equivalent serializable representation of an object.
to_output([output_type, output_dtype])	Convert array to output format
zeros(shape[, dtype, order])	Create an Array with an allocated DeviceBuffer initialized to zeros.

item

classmethod empty(shape, dtype, order='F')

Create an empty Array with an allocated but uninitialized DeviceBuffer

Parameters

dtypedata-type, optional

Any object that can be interpreted as a numpy or cupy data type.

shapeint or tuple of ints, optional

Shape of created array.

order: string, optional

Whether to create a F-major or C-major array.

classmethod full(shape, value, dtype, order='F')

Create an Array with an allocated DeviceBuffer initialized to value.

Parameters

dtypedata-type, optional

Any object that can be interpreted as a numpy or cupy data type.

shapeint or tuple of ints, optional

Shape of created array.

order: string, optional

Whether to create a F-major or C-major array.

classmethod ones(shape, dtype='float32', order='F')

Create an Array with an allocated DeviceBuffer initialized to zeros.

Parameters

dtypedata-type, optional

Any object that can be interpreted as a numpy or cupy data type.

shapeint or tuple of ints, optional

Shape of created array.

order: string, optional

Whether to create a F-major or C-major array.

to_output(output_type='cupy', output_dtype=None)

Convert array to output format

Parameters

output_typestring

Format to convert the array to. Acceptable formats are:

• ‘cupy’ - to cupy array

• ‘numpy’ - to numpy (host) array

• ‘numba’ - to numba device array

• ‘dataframe’ - to cuDF DataFrame

• ‘series’ - to cuDF Series

• ‘cudf’ - to cuDF Series if array is single dimensional, to

  DataFrame otherwise

output_dtypestring, optional

Optionally cast the array to a specified dtype, creating a copy if necessary.

classmethod zeros(shape, dtype='float32', order='F')

Create an Array with an allocated DeviceBuffer initialized to zeros.

Parameters

dtypedata-type, optional

Any object that can be interpreted as a numpy or cupy data type.

shapeint or tuple of ints, optional

Shape of created array.

order: string, optional

Whether to create a F-major or C-major array.

Metrics (regression, classification, and distance)

cuml.metrics.regression.mean_absolute_error(y_true, y_pred, sample_weight=None, multioutput='uniform_average')

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_truearray-like (device or host) shape = (n_samples,)

or (n_samples, n_outputs) Ground truth (correct) target values.

y_predarray-like (device or host) shape = (n_samples,)

or (n_samples, n_outputs) Estimated target values.

sample_weightarray-like (device or host) shape = (n_samples,), optional

Sample weights.

multioutputstring in [‘raw_values’, ‘uniform_average’]

or array-like of shape (n_outputs) Defines aggregating of multiple output values. Array-like 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 non-negative floating point. The best value is 0.0.

cuml.metrics.regression.mean_squared_error(y_true, y_pred, sample_weight=None, multioutput='uniform_average', squared=True)

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_truearray-like (device or host) shape = (n_samples,)

or (n_samples, n_outputs) Ground truth (correct) target values.

y_predarray-like (device or host) shape = (n_samples,)

or (n_samples, n_outputs) Estimated target values.

sample_weightarray-like (device or host) shape = (n_samples,), optional

Sample weights.

multioutputstring in [‘raw_values’, ‘uniform_average’]

or array-like of shape (n_outputs) Defines aggregating of multiple output values. Array-like 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 non-negative 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(y_true, y_pred, sample_weight=None, multioutput='uniform_average', squared=True)

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_truearray-like (device or host) shape = (n_samples,)

or (n_samples, n_outputs) Ground truth (correct) target values.

y_predarray-like (device or host) shape = (n_samples,)

or (n_samples, n_outputs) Estimated target values.

sample_weightarray-like (device or host) shape = (n_samples,), optional

Sample weights.

multioutputstring in [‘raw_values’, ‘uniform_average’]

or array-like of shape (n_outputs) Defines aggregating of multiple output values. Array-like 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 non-negative 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=True, handle=None) → double

Calculates r2 score between y and y_hat

Parameters

yarray-like (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_hatarray-like (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 low-dimensional 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.confusion_matrix(y_true, y_pred, labels=None, sample_weight=None, normalize=None, convert_dtype=False) → cuml.common.array.CumlArray

Compute confusion matrix to evaluate the accuracy of a classification.

Parameters

y_truearray-like (device or host) shape = (n_samples,)

or (n_samples, n_outputs) Ground truth (correct) target values.

y_predarray-like (device or host) shape = (n_samples,)

or (n_samples, n_outputs) Estimated target values.

labelsarray-like (device or host) shape = (n_classes,), optional

List of labels to index the matrix. This may be used to reorder or select a subset of labels. If None is given, those that appear at least once in y_true or y_pred are used in sorted order.

sample_weightarray-like (device or host) shape = (n_samples,), optional

Sample weights.

normalizestring in [‘true’, ‘pred’, ‘all’]

Normalizes confusion matrix over the true (rows), predicted (columns) conditions or all the population. If None, confusion matrix will not be normalized.

convert_dtypebool, optional (default = False)

When set to True, the confusion matrix method will automatically convert the predictions, ground truth, and labels arrays to np.int32.

Returns

Carray-like (device or host) shape = (n_classes, n_classes)

Confusion matrix.

cuml.metrics.kl_divergence(P, Q, handle=None, convert_dtype=True)

Calculates the “Kullback-Leibler” Divergence The KL divergence tells us how well the probability distribution Q approximates the probability distribution P It is often also used as a ‘distance metric’ between two probablity ditributions (not symmetric)

Parameters

PDense array of probabilities corresponding to distribution P

shape = (n_samples, 1) Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy.

QDense array of probabilities corresponding to distribution Q

shape = (n_samples, 1) Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy.

handle : cuml.Handle

convert_dtypebool, optional (default = True)

When set to True, the method will, convert P and Q to be the same data type: float32. This will increase memory used for the method.

float

The KL Divergence value

cuml.metrics.log_loss(y_true, y_pred, eps=1e-15, normalize=True, sample_weight=None) → float

Log loss, aka logistic loss or cross-entropy loss. This is the loss function used in (multinomial) logistic regression and extensions of it such as neural networks, defined as the negative log-likelihood of a logistic model that returns y_pred probabilities for its training data y_true. The log loss is only defined for two or more labels.

Parameters

y_truearray-like, shape = (n_samples,)

y_predarray-like of float,

shape = (n_samples, n_classes) or (n_samples,)

epsfloat

Log loss is undefined for p=0 or p=1, so probabilities are clipped to max(eps, min(1 - eps, p)).

normalizebool, optional (default=True)

If true, return the mean loss per sample. Otherwise, return the sum of the per-sample losses.

sample_weightarray-like of shape (n_samples,), default=None

Sample weights.

Returns

lossfloat

Notes

The logarithm used is the natural logarithm (base-e).

References

C.M. Bishop (2006). Pattern Recognition and Machine Learning. Springer, p. 209.

Examples

>>> from cuml.metrics import log_loss
>>> import numpy as np
>>> log_loss(np.array([1, 0, 0, 1]),
...          np.array([[.1, .9], [.9, .1], [.8, .2], [.35, .65]]))
0.21616...

cuml.metrics.roc_auc_score(y_true, y_score)

Compute Area Under the Receiver Operating Characteristic Curve (ROC AUC) from prediction scores.

Note

this implementation can only be used with binary classification.

Parameters

y_truearray-like of shape (n_samples,)

True labels. The binary cases expect labels with shape (n_samples,)

y_scorearray-like of shape (n_samples,)

Target scores. In the binary cases, these can be either probability estimates or non-thresholded decision values (as returned by decision_function on some classifiers). The binary case expects a shape (n_samples,), and the scores must be the scores of the class with the greater label.

Returns

aucfloat

Examples

>>> import numpy as np
>>> from cuml.metrics import roc_auc_score
>>> y_true = np.array([0, 0, 1, 1])
>>> y_scores = np.array([0.1, 0.4, 0.35, 0.8])
>>> print(roc_auc_score(y_true, y_scores))
0.75

cuml.metrics.precision_recall_curve(y_true, probs_pred) → Tuple[cuml.common.array.CumlArray, cuml.common.array.CumlArray, cuml.common.array.CumlArray]

Compute precision-recall pairs for different probability thresholds

Note

this implementation is restricted to the binary classification task. The precision is the ratio tp / (tp + fp) where tp is the number of true positives and fp the number of false positives. The precision is intuitively the ability of the classifier not to label as positive a sample that is negative.

The recall is the ratio tp / (tp + fn) where tp is the number of true positives and fn the number of false negatives. The recall is intuitively the ability of the classifier to find all the positive samples. The last precision and recall values are 1. and 0. respectively and do not have a corresponding threshold. This ensures that the graph starts on the y axis.

Read more in the scikit-learn’s User Guide.

Parameters

y_truearray, shape = [n_samples]

True binary labels, {0, 1}.

probas_predarray, shape = [n_samples]

Estimated probabilities or decision function.

Returns

precisionarray, shape = [n_thresholds + 1]

Precision values such that element i is the precision of predictions with score >= thresholds[i] and the last element is 1.

recallarray, shape = [n_thresholds + 1]

Decreasing recall values such that element i is the recall of predictions with score >= thresholds[i] and the last element is 0.

thresholdsarray, shape = [n_thresholds <= len(np.unique(probas_pred))]

Increasing thresholds on the decision function used to compute precision and recall.

Examples

import numpy as np
from cuml.metrics import precision_recall_curve
y_true = np.array([0, 0, 1, 1])
y_scores = np.array([0.1, 0.4, 0.35, 0.8])
precision, recall, thresholds = precision_recall_curve(
    y_true, y_scores)
print(precision)
print(recall)
print(thresholds)

Output:

array([0.66666667, 0.5       , 1.        , 1.        ])
array([1. , 0.5, 0.5, 0. ])
array([0.35, 0.4 , 0.8 ])

cuml.metrics.pairwise_distances.pairwise_distances(X, Y=None, metric='euclidean', handle=None, convert_dtype=True, metric_arg=2, **kwds)

Compute the distance matrix from a vector array X and optional Y.

This method takes either one or two vector arrays, and returns a distance matrix.

If Y is given (default is None), then the returned matrix is the pairwise distance between the arrays from both X and Y.

Valid values for metric are:

• From scikit-learn: [‘cityblock’, ‘cosine’, ‘euclidean’, ‘l1’, ‘l2’, ‘manhattan’].

  Sparse matrices are supported, see ‘sparse_pairwise_distances’.

• From scipy.spatial.distance: [‘sqeuclidean’]

  See the documentation for scipy.spatial.distance for details on this metric. Sparse matrices are supported.

Parameters

XDense or sparse matrix (device or host) of shape

(n_samples_x, n_features) Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy, or cupyx.scipy.sparse for sparse input

Yarray-like (device or host) of shape (n_samples_y, n_features), optional

Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy

metric{“cityblock”, “cosine”, “euclidean”, “l1”, “l2”, “manhattan”, “sqeuclidean”}

The metric to use when calculating distance between instances in a feature array.

convert_dtypebool, optional (default = True)

When set to True, the 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.

Returns

Darray [n_samples_x, n_samples_x] or [n_samples_x, n_samples_y]

A distance matrix D such that D_{i, j} is the distance between the ith and jth vectors of the given matrix X, if Y is None. If Y is not None, then D_{i, j} is the distance between the ith array from X and the jth array from Y.

Examples

>>> import cupy as cp
>>> from cuml.metrics import pairwise_distances
>>>
>>> X = cp.array([[2.0, 3.0], [3.0, 5.0], [5.0, 8.0]])
>>> Y = cp.array([[1.0, 0.0], [2.0, 1.0]])
>>>
>>> # Euclidean Pairwise Distance, Single Input:
>>> pairwise_distances(X, metric='euclidean')
array([[0.        , 2.23606798, 5.83095189],
    [2.23606798, 0.        , 3.60555128],
    [5.83095189, 3.60555128, 0.        ]])
>>>
>>> # Cosine Pairwise Distance, Multi-Input:
>>> pairwise_distances(X, Y, metric='cosine')
array([[0.4452998 , 0.13175686],
    [0.48550424, 0.15633851],
    [0.47000106, 0.14671817]])
>>>
>>> # Manhattan Pairwise Distance, Multi-Input:
>>> pairwise_distances(X, Y, metric='manhattan')
array([[ 4.,  2.],
    [ 7.,  5.],
    [12., 10.]])

cuml.metrics.pairwise_distances.sparse_pairwise_distances(X, Y=None, metric='euclidean', handle=None, convert_dtype=True, metric_arg=2, **kwds)

Compute the distance matrix from a vector array X and optional Y.

This method takes either one or two sparse vector arrays, and returns a dense distance matrix.

If Y is given (default is None), then the returned matrix is the pairwise distance between the arrays from both X and Y.

Valid values for metric are:

• From scikit-learn: [‘cityblock’, ‘cosine’, ‘euclidean’, ‘l1’, ‘l2’, ‘manhattan’].

• From scipy.spatial.distance: [‘sqeuclidean’, ‘canberra’, ‘minkowski’, ‘jaccard’, ‘chebyshev’, ‘dice’]

  See the documentation for scipy.spatial.distance for details on these metrics.

• [‘inner_product’, ‘hellinger’]

Xarray-like (device or host) of shape (n_samples_x, n_features)

Acceptable formats: SciPy or Cupy sparse array

Yarray-like (device or host) of shape (n_samples_y, n_features), optional

Acceptable formats: SciPy or Cupy sparse array

metric{“cityblock”, “cosine”, “euclidean”, “l1”, “l2”, “manhattan”, “sqeuclidean”, “canberra”, “lp”, “inner_product”, “minkowski”, “jaccard”, “hellinger”, “chebyshev”, “linf”, “dice”}

The metric to use when calculating distance between instances in a feature array.

convert_dtypebool, optional (default = True)

When set to True, the 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.

metric_argfloat, optional (default = 2)

Additionnal metric-specific argument. For Minkowski it’s the p-norm to apply.

Returns

Darray [n_samples_x, n_samples_x] or [n_samples_x, n_samples_y]

A dense distance matrix D such that D_{i, j} is the distance between the ith and jth vectors of the given matrix X, if Y is None. If Y is not None, then D_{i, j} is the distance between the ith array from X and the jth array from Y.

Examples

>>> import cupyx
>>> from cuml.metrics import sparse_pairwise_distances
>>>
>>> X = cupyx.scipy.sparse.random(2, 3, density=0.5)
>>> Y = cupyx.scipy.sparse.random(1, 3, density=0.5)
>>> X.todense()
array([[0.02797998, 0.        , 0.66309184],
       [0.        , 0.        , 0.92316052]])
>>> Y.todense()
array([[0.        , 0.        , 0.32750517]])
>>> # Cosine Pairwise Distance, Single Input:
>>> sparse_pairwise_distances(X, metric='cosine')
array([[0.        , 0.00088907],
       [0.00088907, 0.        ]])
>>>
>>> # Squared euclidean Pairwise Distance, Multi-Input:
>>> sparse_pairwise_distances(X, Y, metric='sqeuclidean')
array([[0.11340129],
       [0.3548053]])
>>>
>>> # Canberra Pairwise Distance, Multi-Input:
>>> sparse_pairwise_distances(X, Y, metric='canberra')
array([[1.33877214],
       [0.47627064]])

Metrics (clustering and manifold learning)

cuml.metrics.trustworthiness.trustworthiness(X, X_embedded, handle=None, n_neighbors=5, metric='euclidean', should_downcast=True, convert_dtype=False, batch_size=512) → double

Expresses to what extent the local structure is retained in embedding. The score is defined in the range [0, 1].

Parameters

Xarray-like (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_embeddedarray-like (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 low-dimensional embedding

cuml.metrics.cluster.adjusted_rand_index.adjusted_rand_score(labels_true, labels_pred, handle=None, convert_dtype=True) → float

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) → float

Computes the entropy of a distribution for given probability values.

Parameters

clusteringarray-like (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.

cuml.metrics.cluster.homogeneity_score.cython_homogeneity_score(labels_true, labels_pred, handle=None) → float

Computes the homogeneity metric of a cluster labeling given a ground truth.

A clustering result satisfies homogeneity if all of its clusters contain only data points which are members of a single class.

This metric is independent of the absolute values of the labels: a permutation of the class or cluster label values won’t change the score value in any way.

This metric is not symmetric: switching label_true with label_pred will return the completeness_score which will be different in general.

The labels in labels_pred and labels_true are assumed to be drawn from a contiguous set (Ex: drawn from {2, 3, 4}, but not from {2, 4}). If your set of labels looks like {2, 4}, convert them to something like {0, 1}.

Parameters

labels_predarray-like (device or host) shape = (n_samples,)

The labels predicted by the model for the test dataset. Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy

labels_truearray-like (device or host) shape = (n_samples,)

The ground truth labels (ints) of the test dataset. Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy

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

float

The homogeneity of the predicted labeling given the ground truth. Score between 0.0 and 1.0. 1.0 stands for perfectly homogeneous labeling.

cuml.metrics.cluster.silhouette_score.cython_silhouette_samples(X, labels, metric='euclidean', chunksize=None, handle=None)

Calculate the silhouette coefficient for each sample in the provided data.

Given a set of cluster labels for every sample in the provided data, compute the mean intra-cluster distance (a) and the mean nearest-cluster distance (b) for each sample. The silhouette coefficient for a sample is then (b - a) / max(a, b).

Parameters

Xarray-like, shape = (n_samples, n_features)

The feature vectors for all samples.

labelsarray-like, shape = (n_samples,)

The assigned cluster labels for each sample.

metricstring

A string representation of the distance metric to use for evaluating the silhouette score. Available options are “cityblock”, “cosine”, “euclidean”, “l1”, “l2”, “manhattan”, and “sqeuclidean”.

chunksizeinteger (default = None)

An integer, 1 <= chunksize <= n_samples to tile the pairwise distance matrix computations, so as to reduce the quadratic memory usage of having the entire pairwise distance matrix in GPU memory. If None, chunksize will automatically be set to 40000, which through experiments has proved to be a safe number for the computation to run on a GPU with 16 GB VRAM.

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.

cuml.metrics.cluster.silhouette_score.cython_silhouette_score(X, labels, metric='euclidean', chunksize=None, handle=None)

Calculate the mean silhouette coefficient for the provided data.

Given a set of cluster labels for every sample in the provided data, compute the mean intra-cluster distance (a) and the mean nearest-cluster distance (b) for each sample. The silhouette coefficient for a sample is then (b - a) / max(a, b).

Parameters

Xarray-like, shape = (n_samples, n_features)

The feature vectors for all samples.

labelsarray-like, shape = (n_samples,)

The assigned cluster labels for each sample.

metricstring

A string representation of the distance metric to use for evaluating the silhouette score. Available options are “cityblock”, “cosine”, “euclidean”, “l1”, “l2”, “manhattan”, and “sqeuclidean”.

chunksizeinteger (default = None)

An integer, 1 <= chunksize <= n_samples to tile the pairwise distance matrix computations, so as to reduce the quadratic memory usage of having the entire pairwise distance matrix in GPU memory. If None, chunksize will automatically be set to 40000, which through experiments has proved to be a safe number for the computation to run on a GPU with 16 GB VRAM.

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.

cuml.metrics.cluster.completeness_score.cython_completeness_score(labels_true, labels_pred, handle=None) → float

Completeness metric of a cluster labeling given a ground truth.

A clustering result satisfies completeness if all the data points that are members of a given class are elements of the same cluster.

This metric is independent of the absolute values of the labels: a permutation of the class or cluster label values won’t change the score value in any way.

This metric is not symmetric: switching label_true with label_pred will return the homogeneity_score which will be different in general.

The labels in labels_pred and labels_true are assumed to be drawn from a contiguous set (Ex: drawn from {2, 3, 4}, but not from {2, 4}). If your set of labels looks like {2, 4}, convert them to something like {0, 1}.

Parameters

labels_predarray-like (device or host) shape = (n_samples,)

The labels predicted by the model for the test dataset. Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy

labels_truearray-like (device or host) shape = (n_samples,)

The ground truth labels (ints) of the test dataset. Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy

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

float

The completeness of the predicted labeling given the ground truth. Score between 0.0 and 1.0. 1.0 stands for perfectly complete labeling.

cuml.metrics.cluster.mutual_info_score.cython_mutual_info_score(labels_true, labels_pred, handle=None) → float

Computes the Mutual Information between two clusterings.

The Mutual Information is a measure of the similarity between two labels of the same data.

This metric is independent of the absolute values of the labels: a permutation of the class or cluster label values won’t change the score value in any way.

This metric is furthermore symmetric: switching label_true with label_pred will return the same score value. This can be useful to measure the agreement of two independent label assignments strategies on the same dataset when the real ground truth is not known.

The labels in labels_pred and labels_true are assumed to be drawn from a contiguous set (Ex: drawn from {2, 3, 4}, but not from {2, 4}). If your set of labels looks like {2, 4}, convert them to something like {0, 1}.

Parameters

handlecuml.Handle

labels_predarray-like (device or host) shape = (n_samples,)

A clustering of the data (ints) into disjoint subsets. Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy

labels_truearray-like (device or host) shape = (n_samples,)

A clustering of the data (ints) into disjoint subsets. Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy

Returns

float

Mutual information, a non-negative value

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 cpu-based algorithm (typically scikit-learn, 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 CPU-based 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(data, **override_args)	Runs the cpu-based algorithm’s fit method on specified data
run_cuml(data, **override_args)	Runs the cuml-based algorithm’s fit method on specified data

setup_cpu
setup_cuml

run_cpu(data, **override_args)

Runs the cpu-based algorithm’s fit method on specified data

run_cuml(data, **override_args)

Runs the cuml-based algorithm’s fit method on specified data

cuml.benchmark.algorithms.algorithm_by_name(name)

Returns the algorithm pair with the name ‘name’ (case-insensitive)

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 cpu-based 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 generator-specific parameters)

The function should return a 2-tuple (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(*, algorithm='eig', fit_intercept=True, normalize=False, handle=None, verbose=False, output_type=None)

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.

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

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)

fit(self, X, y, convert_dtype=True) → u’LinearRegression’

Fit the model with X and y.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the train 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)

Logistic Regression

class cuml.LogisticRegression(*, penalty='l2', tol=0.0001, C=1.0, fit_intercept=True, class_weight=None, max_iter=1000, linesearch_max_iter=50, verbose=False, l1_ratio=None, solver='qn', handle=None, output_type=None)

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 array-like 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 single-class (using sigmoid loss) and multiple-class (using softmax loss) variants, depending on the input variables

Only one solver option is currently available: Quasi-Newton (QN) algorithms. Even though it is presented as a single option, this solver resolves to two different algorithms underneath:

• Orthant-Wise Limited Memory Quasi-Newton (OWL-QN) if there is l1 regularization

• Limited Memory BFGS (L-BFGS) otherwise.

Note that, just like in Scikit-learn, 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 L-BFGS solver will be used. If ‘l1’ is selected, solver OWL-QN will be used. If ‘elasticnet’ is selected, OWL-QN will be used if l1_ratio > 0, otherwise L-BFGS will be used.

tol: float (default = 1e-4)

Tolerance for stopping criteria. The exact stopping conditions depend on the chosen solver. Check the solver’s documentation for more details:

• Quasi-Newton (L-BFGS/OWL-QN)

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.

class_weight: dict or ‘balanced’, default=None

By default all classes have a weight one. However, a dictionary can be provided with weights associated with classes in the form {class_label: weight}. The “balanced” mode uses the values of y to automatically adjust weights inversely proportional to class frequencies in the input data as n_samples / (n_classes * np.bincount(y)). Note that these weights will be multiplied with sample_weight (passed through the fit method) if sample_weight is specified.

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

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

l1_ratio: float or None, optional (default=None)

The Elastic-Net 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 L-BFGS or OWL-QN 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.

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.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

Notes

cuML’s LogisticRegression uses a different solver that the equivalent Scikit-learn, except when there is no penalty and solver=lbfgs is used in Scikit-learn. This can cause (smaller) differences in the coefficients and predictions of the model, similar to using different solvers in Scikit-learn.

For additional information, see Scikit-learn’s LogisticRegression.

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_)
print("Intercept:")
print(reg.intercept_)

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.

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[, sample_weight, 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
set_params(self, **params)

decision_function(self, X, convert_dtype=False) → CumlArray

Gives confidence score for X

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense or sparse matrix containing floats or doubles. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = False)

When set to True, the decision_function 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

scorecuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, n_classes)

Confidence score

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

fit(self, X, y, sample_weight=None, convert_dtype=True) → u’LogisticRegression’

Fit the model with X and y.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense or sparse matrix containing floats or doubles. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

sample_weightarray-like (device or host) shape = (n_samples,), default=None

The weights for each observation in X. If None, all observations are assigned equal weight. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the train 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=True) → CumlArray

Predicts the y for X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense or sparse matrix containing floats or doubles. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Predicted values

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

predict_log_proba(self, X, convert_dtype=True) → CumlArray

Predicts the log class probabilities for each class in X

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense or sparse matrix containing floats or doubles. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the predict_log_proba 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

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, n_classes)

Logaright of predicted class probabilities

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

predict_proba(self, X, convert_dtype=True) → CumlArray

Predicts the class probabilities for each class in X

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense or sparse matrix containing floats or doubles. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the predict_proba 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

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, n_classes)

Predicted class probabilities

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

set_params(self, **params)

Ridge Regression

class cuml.Ridge(*, alpha=1.0, solver='eig', fit_intercept=True, normalize=False, handle=None, output_type=None, verbose=False)

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 array-like 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.

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.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

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 Scikit-learn’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([1e-5])
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)
set_params(self, **params)

fit(self, X, y, convert_dtype=True) → u’Ridge’

Fit the model with X and y.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the train 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)

set_params(self, **params)

Lasso Regression

class cuml.Lasso(*, alpha=1.0, fit_intercept=True, normalize=False, max_iter=1000, tol=0.001, selection='cyclic', handle=None, output_type=None, verbose=False)

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 array-like 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 = 1e-3)

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 1e-4.

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.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

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_param_names(self)
set_params(self, **params)

fit(self, X, y, convert_dtype=True) → u’Lasso’

Fit the model with X and y.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the train 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)

set_params(self, **params)

ElasticNet Regression

class cuml.ElasticNet(*, alpha=1.0, l1_ratio=0.5, fit_intercept=True, normalize=False, max_iter=1000, tol=0.001, selection='cyclic', handle=None, output_type=None, verbose=False)

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 array-like 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 = 1e-3)

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 1e-4.

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.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

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_param_names(self)
set_params(self, **params)

fit(self, X, y, convert_dtype=True) → u’ElasticNet’

Fit the model with X and y.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the train 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)

set_params(self, **params)

Mini Batch SGD Classifier

class cuml.MBSGDClassifier(*, loss='hinge', penalty='l2', alpha=0.0001, l1_ratio=0.15, fit_intercept=True, epochs=1000, tol=0.001, shuffle=True, learning_rate='constant', eta0=0.001, power_t=0.5, batch_size=32, n_iter_no_change=5, handle=None, verbose=False, output_type=None)

Linear models (linear SVM, logistic regression, or linear regression) fitted by minimizing a regularized empirical loss with mini-batch SGD. The MBSGD Classifier implementation is experimental and and it uses a different algorithm than sklearn’s SGDClassifier. In order to improve the results obtained from cuML’s MBSGDClassifier: * Reduce the batch size * Increase the eta0 * Increase the number of iterations Since cuML is analyzing the data in batches using a small eta0 might not let the model learn as much as scikit learn does. Furthermore, decreasing the batch size might seen an increase in the time required to fit the model.

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

l1_ratio: float (default=0.15)

The l1_ratio is used only when penalty = elasticnet. The value for l1_ratio should be 0 <= l1_ratio <= 1. When l1_ratio = 0 then the penalty = 'l2' and if l1_ratio = 1 then penalty = 'l1'

batch_size: int (default = 32)

It sets the number of samples that will be included in each batch.

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 = 1e-3)

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

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

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_param_names(self)
predict(self, X[, convert_dtype])	Predicts the y for X.
set_params(self, **params)

fit(self, X, y, convert_dtype=True) → u’MBSGDClassifier’

Fit the model with X and y.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the train 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) → CumlArray

Predicts the y for X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Predicted values

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

set_params(self, **params)

Mini Batch SGD Regressor

class cuml.MBSGDRegressor(*, loss='squared_loss', penalty='l2', alpha=0.0001, l1_ratio=0.15, fit_intercept=True, epochs=1000, tol=0.001, shuffle=True, learning_rate='constant', eta0=0.001, power_t=0.5, batch_size=32, n_iter_no_change=5, handle=None, verbose=False, output_type=None)

Linear regression model fitted by minimizing a regularized empirical loss with mini-batch SGD. The MBSGD Regressor implementation is experimental and and it uses a different algorithm than sklearn’s SGDClassifier. In order to improve the results obtained from cuML’s MBSGD Regressor: * Reduce the batch size * Increase the eta0 * Increase the number of iterations Since cuML is analyzing the data in batches using a small eta0 might not let the model learn as much as scikit learn does. Furthermore, decreasing the batch size might seen an increase in the time required to fit the model.

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.

l1_ratio: float (default=0.15)

The l1_ratio is used only when penalty = elasticnet. The value for l1_ratio should be 0 <= l1_ratio <= 1. When l1_ratio = 0 then the penalty = 'l2' and if l1_ratio = 1 then penalty = 'l1'

batch_size: int (default = 32)

It sets the number of samples that will be included in each batch.

epochsint (default = 1000)

The number of times the model should iterate through the entire dataset during training (default = 1000)

tolfloat (default = 1e-3)

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

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

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_param_names(self)
predict(self, X[, convert_dtype])	Predicts the y for X.
set_params(self, **params)

fit(self, X, y, convert_dtype=True) → u’MBSGDRegressor’

Fit the model with X and y.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the train 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) → CumlArray

Predicts the y for X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Predicted values

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

set_params(self, **params)

Multiclass Classification

class cuml.multiclass.MulticlassClassifier(estimator, *, handle=None, verbose=False, output_type=None, strategy='ovr')

Wrapper around scikit-learn multiclass classifiers that allows to choose different multiclass strategies.

The input can be any kind of cuML compatible array, and the output type follows cuML’s output type configuration rules.

Berofe passing the data to scikit-learn, it is converted to host (numpy) array. Under the hood the data is partitioned for binary classification, and it is transformed back to the device by the cuML estimator. These copies back and forth the device and the host have some overhead. For more details see issue https://github.com/rapidsai/cuml/issues/2876.

Parameters

estimatorcuML estimator

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

strategy: string {‘ovr’, ‘ovo’}, default=’ovr’

Multiclass classification strategy: ‘ovr’: one vs. rest or ‘ovo’: one vs. one

Examples

>>> from cuml.linear_model import LogisticRegression
>>> from cuml.multiclass import MulticlassClassifier
>>> from cuml.datasets.classification import make_classification
>>>
>>> X, y = make_classification(n_samples=10, n_features=6, n_informative=4,
...                            n_classes=3, random_state=137)
>>>
>>> cls = MulticlassClassifier(LogisticRegression(), strategy='ovo')
>>> cls.fit(X,y)
>>> cls.predict(X)
array([1, 1, 1, 0, 0, 2, 2, 2, 0, 1])

Attributes

classes_float, shape (n_classes_)

Array of class labels.

n_classes_int

Number of classes.

Methods

decision_function(X)	Calculate the decision function.
fit(X, y)	Fit a multiclass classifier.
get_param_names(self)	Returns a list of hyperparameter names owned by this class.
predict(X)	Predict using multi class classifier.

decision_function(X) → cuml.common.array.CumlArray

Calculate the decision function.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

Returns

resultscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Decision function values

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

fit(X, y) → cuml.multiclass.multiclass.MulticlassClassifier

Fit a multiclass classifier.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix of any dtype. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

get_param_names(self)

Returns a list of hyperparameter names owned by this class. It is expected that every child class overrides this method and appends its extra set of parameters that it in-turn owns. This is to simplify the implementation of get_params and set_params methods.

predict(X) → cuml.common.array.CumlArray

Predict using multi class classifier.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

Returns

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Predicted values

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

class cuml.multiclass.OneVsOneClassifier(estimator, *args, handle=None, verbose=False, output_type=None)

Wrapper around Sckit-learn’s class with the same name. The input can be any kind of cuML compatible array, and the output type follows cuML’s output type configuration rules.

Berofe passing the data to scikit-learn, it is converted to host (numpy) array. Under the hood the data is partitioned for binary classification, and it is transformed back to the device by the cuML estimator. These copies back and forth the device and the host have some overhead. For more details see issue https://github.com/rapidsai/cuml/issues/2876.

For documentation see scikit-learn’s OneVsOneClassifier.

Parameters

estimatorcuML estimator

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

Examples

>>> from cuml.linear_model import LogisticRegression
>>> from cuml.multiclass import OneVsOneClassifier
>>> from cuml.datasets.classification import make_classification
>>>
>>> X, y = make_classification(n_samples=10, n_features=6, n_informative=4,
...                            n_classes=3, random_state=137)
>>>
>>> cls = OneVsOneClassifier(LogisticRegression())
>>> cls.fit(X,y)
>>> cls.predict(X)
array([1, 1, 1, 0, 0, 2, 2, 2, 0, 1])

Methods

get_param_names(self)

Returns a list of hyperparameter names owned by this class.

get_param_names(self)

Returns a list of hyperparameter names owned by this class. It is expected that every child class overrides this method and appends its extra set of parameters that it in-turn owns. This is to simplify the implementation of get_params and set_params methods.

class cuml.multiclass.OneVsRestClassifier(estimator, *args, handle=None, verbose=False, output_type=None)

Wrapper around Sckit-learn’s class with the same name. The input can be any kind of cuML compatible array, and the output type follows cuML’s output type configuration rules.

Berofe passing the data to scikit-learn, it is converted to host (numpy) array. Under the hood the data is partitioned for binary classification, and it is transformed back to the device by the cuML estimator. These copies back and forth the device and the host have some overhead. For more details see issue https://github.com/rapidsai/cuml/issues/2876.

For documentation see scikit-learn’s OneVsRestClassifier.

Parameters

estimatorcuML estimator

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

Examples

>>> from cuml.linear_model import LogisticRegression
>>> from cuml.multiclass import OneVsRestClassifier
>>> from cuml.datasets.classification import make_classification
>>>
>>> X, y = make_classification(n_samples=10, n_features=6, n_informative=4,
...                            n_classes=3, random_state=137)
>>>
>>> cls = OneVsRestClassifier(LogisticRegression())
>>> cls.fit(X,y)
>>> cls.predict(X)
array([1, 1, 1, 0, 1, 2, 2, 2, 0, 1])

Methods

get_param_names(self)

Returns a list of hyperparameter names owned by this class.

get_param_names(self)

Returns a list of hyperparameter names owned by this class. It is expected that every child class overrides this method and appends its extra set of parameters that it in-turn owns. This is to simplify the implementation of get_params and set_params methods.

Naive Bayes

class cuml.naive_bayes.MultinomialNB(*, alpha=1.0, fit_prior=True, class_prior=None, output_type=None, handle=None, verbose=False)

Naive Bayes classifier for multinomial models

The multinomial Naive Bayes classifier is suitable for classification with discrete features (e.g., word counts for text classification).

The multinomial distribution normally requires integer feature counts. However, in practice, fractional counts such as tf-idf may also work.

Parameters

alphafloat

Additive (Laplace/Lidstone) smoothing parameter (0 for no smoothing).

fit_priorboolean

Whether to learn class prior probabilities or no. If false, a uniform prior will be used.

class_priorarray-like, size (n_classes)

Prior probabilities of the classes. If specified, the priors are not adjusted according to the data.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

Examples

Load the 20 newsgroups dataset from Scikit-learn and train a Naive Bayes classifier.

import cupy as cp

from sklearn.datasets import fetch_20newsgroups
from sklearn.feature_extraction.text import CountVectorizer

from cuml.naive_bayes import MultinomialNB

# Load corpus

twenty_train = fetch_20newsgroups(subset='train',
                        shuffle=True, random_state=42)

# Turn documents into term frequency vectors

count_vect = CountVectorizer()
features = count_vect.fit_transform(twenty_train.data)

# Put feature vectors and labels on the GPU

X = cp.sparse.csr_matrix(features.tocsr(), dtype=cp.float32)
y = cp.asarray(twenty_train.target, dtype=cp.int32)

# Train model

model = MultinomialNB()
model.fit(X, y)

# Compute accuracy on training set

model.score(X, y)

Output:

0.9244298934936523

Methods

get_param_names(self)

Returns a list of hyperparameter names owned by this class.

get_param_names(self)

Returns a list of hyperparameter names owned by this class. It is expected that every child class overrides this method and appends its extra set of parameters that it in-turn owns. This is to simplify the implementation of get_params and set_params methods.

class cuml.naive_bayes.BernoulliNB(*, alpha=1.0, binarize=0.0, fit_prior=True, class_prior=None, output_type=None, handle=None, verbose=False)

Naive Bayes classifier for multivariate Bernoulli models. Like MultinomialNB, this classifier is suitable for discrete data. The difference is that while MultinomialNB works with occurrence counts, BernoulliNB is designed for binary/boolean features.

Parameters

alphafloat, default=1.0

Additive (Laplace/Lidstone) smoothing parameter (0 for no smoothing).

binarizefloat or None, default=0.0

Threshold for binarizing (mapping to booleans) of sample features. If None, input is presumed to already consist of binary vectors.

fit_priorbool, default=True

Whether to learn class prior probabilities or not. If false, a uniform prior will be used.

class_priorarray-like of shape (n_classes,), default=None

Prior probabilities of the classes. If specified the priors are not adjusted according to the data.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

References

C.D. Manning, P. Raghavan and H. Schuetze (2008). Introduction to Information Retrieval. Cambridge University Press, pp. 234-265. https://nlp.stanford.edu/IR-book/html/htmledition/the-bernoulli-model-1.html A. McCallum and K. Nigam (1998). A comparison of event models for naive Bayes text classification. Proc. AAAI/ICML-98 Workshop on Learning for Text Categorization, pp. 41-48. V. Metsis, I. Androutsopoulos and G. Paliouras (2006). Spam filtering with naive Bayes – Which naive Bayes? 3rd Conf. on Email and Anti-Spam (CEAS).

Examples

>>> import cupy as cp
>>> rng = cp.random.RandomState(1)
>>> X = rng.randint(5, size=(6, 100))
>>> Y = cp.array([1, 2, 3, 4, 4, 5])
>>> from cuml.naive_bayes import BernoulliNB
>>> clf = BernoulliNB()
>>> clf.fit(X, Y)
BernoulliNB()
>>> print(clf.predict(X[2:3]))
[3]

Attributes

class_count_ndarray of shape (n_classes)

Number of samples encountered for each class during fitting. This value is weighted by the sample weight when provided.

class_log_prior_ndarray of shape (n_classes)

Log probability of each class (smoothed).

classes_ndarray of shape (n_classes,)

Class labels known to the classifier

feature_count_ndarray of shape (n_classes, n_features)

Number of samples encountered for each (class, feature) during fitting. This value is weighted by the sample weight when provided.

feature_log_prob_ndarray of shape (n_classes, n_features)

Empirical log probability of features given a class, P(x_i|y).

n_features_int

Number of features of each sample.

Methods

get_param_names(self)

Returns a list of hyperparameter names owned by this class.

get_param_names(self)

Returns a list of hyperparameter names owned by this class. It is expected that every child class overrides this method and appends its extra set of parameters that it in-turn owns. This is to simplify the implementation of get_params and set_params methods.

Stochastic Gradient Descent

class cuml.SGD(*, loss='squared_loss', penalty='none', alpha=0.0001, l1_ratio=0.15, fit_intercept=True, epochs=1000, tol=0.001, shuffle=True, learning_rate='constant', eta0=0.001, power_t=0.5, batch_size=32, n_iter_no_change=5, handle=None, output_type=None, verbose=False)

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 = 1e-3)

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

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.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

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=np.float32)
pred_data['col2'] = np.asarray([5, 5], dtype=np.float32)
cu_sgd = cumlSGD(learning_rate='constant', eta0=0.005, epochs=2000,
                fit_intercept=True, batch_size=2,
                tol=0.0, penalty='none', loss='squared_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 :  0.0041877031326293945
cuML coef :  0      0.984174
             1      0.009776
            dtype: float32
cuML predictions :  [3.005588  2.0214138]

Attributes

classes_

Python descriptor object to control getting/setting CumlArray attributes on Base objects. See the Estimator Guide for an in depth guide.

coef_

Python descriptor object to control getting/setting CumlArray attributes on Base objects. See the Estimator Guide for an in depth guide.

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.
predictClass(self, X[, convert_dtype])	Predicts the y for X.

fit(self, X, y, convert_dtype=False) → u’SGD’

Fit the model with X and y.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = False)

When set to True, the train 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) → CumlArray

Predicts the y for X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Predicted values

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

predictClass(self, X, convert_dtype=False) → CumlArray

Predicts the y for X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = False)

When set to True, the predictClass 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

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Predicted values

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

Random Forest

class cuml.ensemble.RandomForestClassifier(*, split_criterion=0, handle=None, verbose=False, output_type=None, **kwargs)

Implements a Random Forest classifier model which fits multiple decision tree classifiers in an ensemble.

Note

Note that the underlying algorithm for tree node splits differs from that used in scikit-learn. By default, the cuML Random Forest uses a histogram-based algorithm to determine splits, rather than an exact count. You can tune the size of the histograms with the n_bins parameter.

Note

You can export cuML Random Forest models and run predictions with them on machines without an NVIDIA GPUs. See https://docs.rapids.ai/api/cuml/nightly/pickling_cuml_models.html for more details.

Known Limitations: This is an early release of the cuML Random Forest code. It contains a few known limitations:

• GPU-based inference is only supported if the model was trained with 32-bit (float32) datatypes. CPU-based 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.

• While training the model for multi class classification problems, using deep trees or max_features=1.0 provides better performance.

• Prediction of classes is currently different from how scikit-learn predicts:

  • scikit-learn predicts random forest classifiers by obtaining class probabilities from each component tree, then averaging these class probabilities over all the ensemble members, and finally resolving to the label with highest probability as prediction.

  • cuml random forest classifier prediction differs in that, each component tree generates labels instead of class probabilities; with the most frequent label over all the trees (the statistical mode) resolved as prediction.

  The above differences might cause marginal variations in accuracy in tradeoff to better performance. See: https://github.com/rapidsai/cuml/issues/3764

Parameters

n_estimatorsint (default = 100)

Number of trees in the forest. (Default changed to 100 in cuML 0.11)

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)

Deprecated and currrently has no effect. .. deprecated:: 21.06

bootstrapboolean (default = True)

Control bootstrapping. If True, each tree in the forest is built on a bootstrapped sample with replacement. If False, the whole dataset is used to build each tree.

max_samplesfloat (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 scikit-learn’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 = 128)

Number of bins used by the split algorithm. For large problems, particularly those with highly-skewed input data, increasing the number of bins may improve accuracy.

n_streamsint (default = 4)

Number of parallel streams used for forest building.

min_samples_leafint or float (default = 1)

The minimum number of samples (rows) in each leaf node. If int, then min_samples_leaf represents the minimum number. If float, then min_samples_leaf represents a fraction and ceil(min_samples_leaf * n_rows) is the minimum number of samples for each leaf node.

min_samples_splitint or float (default = 2)

The minimum number of samples required to split an internal node. If int, then min_samples_split represents the minimum number. If float, then min_samples_split represents a fraction and ceil(min_samples_split * n_rows) is the minimum number of samples for each split.

min_impurity_decreasefloat (default = 0.0)

Minimum decrease in impurity requried for node to be spilt.

use_experimental_backendboolean (default = True)

Deprecated and currrently has no effect. .. deprecated:: 21.08

max_batch_size: int (default = 4096)

Maximum number of nodes that can be processed in a given batch.

random_stateint (default = None)

Seed for the random number generator. Unseeded by default. Does not currently fully guarantee the exact same results. Note: Parameter `seed` is removed since release 0.19.

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

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.
convert_to_treelite_model(self)	Converts the cuML RF model to a Treelite model
fit(self, X, y[, convert_dtype])	Perform Random Forest Classification on the input data
get_detailed_text(self)	Obtain the detailed information for the random forest model, as text
get_json(self)	Export the Random Forest model as a JSON string
get_summary_text(self)	Obtain the text summary of the random forest model
predict(self, X[, predict_model, threshold, …])	Predicts the labels for X.
predict_proba(self, X[, algo, num_classes, …])	Predicts class probabilites for X.
score(self, X, y[, threshold, algo, …])	Calculates the accuracy metric score of the model for X.

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 coalescing-friendly

• '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=True)

Perform Random Forest Classification on the input data

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix of type np.int32. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the method will automatically convert the inputs to np.float32.

convert_dtypebool, optional (default = True)

When set to True, the fit method will, when necessary, convert y to be of dtype int32. This will increase memory used for the method.

get_detailed_text(self)

Obtain the detailed information for the random forest model, as text

get_json(self)

Export the Random Forest model as a JSON string

get_summary_text(self)

Obtain the text summary of the random forest model

predict(self, X, predict_model='GPU', threshold=0.5, algo='auto', num_classes=None, convert_dtype=True, fil_sparse_format='auto') → CumlArray

Predicts the labels for X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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 classification problems.

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 coalescing-friendly

• '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.

num_classesint (default = None)

number of different classes present in the dataset.

Deprecated since version 0.16: Parameter ‘num_classes’ is deprecated and will be removed in an upcoming version. The number of classes passed must match the number of classes the model was trained on.

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

ycuDF, CuPy or NumPy object depending on cuML’s output typeconfiguration, shape =(n_samples, 1)

predict_proba(self, X, algo='auto', num_classes=None, convert_dtype=True, fil_sparse_format='auto') → CumlArray

Predicts class probabilites for X. This function uses the GPU implementation of predict. Therefore, data with ‘dtype = np.float32’ should be used with this function.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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 coalescing-friendly

• '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

num_classesint (default = None)

number of different classes present in the dataset.

Deprecated since version 0.16: Parameter ‘num_classes’ is deprecated and will be removed in an upcoming version. The number of classes passed must match the number of classes the model was trained on.

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

ycuDF, CuPy or NumPy object depending on cuML’s output typeconfiguration, shape =(n_samples, 1)

score(self, X, y, threshold=0.5, algo='auto', num_classes=None, predict_model='GPU', convert_dtype=True, fil_sparse_format='auto')

Calculates the accuracy metric score of the model for X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix of type np.int32. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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 coalescing-friendly

• '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_classesint (default = None)

number of different classes present in the dataset.

Deprecated since version 0.16: Parameter ‘num_classes’ is deprecated and will be removed in an upcoming version. The number of classes passed must match the number of classes the model was trained on.

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 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]

class cuml.ensemble.RandomForestRegressor(*, split_criterion=2, accuracy_metric='r2', handle=None, verbose=False, output_type=None, **kwargs)

Implements a Random Forest regressor model which fits multiple decision trees in an ensemble.

Note

Note that the underlying algorithm for tree node splits differs from that used in scikit-learn. By default, the cuML Random Forest uses a histogram-based algorithm to determine splits, rather than an exact count. You can tune the size of the histograms with the n_bins parameter.

Note

You can export cuML Random Forest models and run predictions with them on machines without an NVIDIA GPUs. See https://docs.rapids.ai/api/cuml/nightly/pickling_cuml_models.html for more details.

Known Limitations: This is an early release of the cuML Random Forest code. It contains a few known limitations:

• GPU-based inference is only supported if the model was trained with 32-bit (float32) datatypes. CPU-based 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)

split_algoint (default = 1)

The algorithm to determine how nodes are split in the tree. Can be changed only for the old backend [deprecated]. 0 for HIST and 1 for GLOBAL_QUANTILE. Default is GLOBAL_QUANTILE. The default backend does not support HIST. HIST currently uses a slower tree-building algorithm so GLOBAL_QUANTILE is recommended for most cases.

Deprecated since version 21.06: Parameter ‘split_algo’ is deprecated and will be removed in subsequent release.

split_criterionint (default = 2)

The criterion used to split nodes. 0 for GINI, 1 for ENTROPY, 2 for MSE 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, the whole dataset is used to build each tree.

max_samplesfloat (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. Note that this default differs from scikit-learn’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 = 128)

Number of bins used by the split algorithm. For large problems, particularly those with highly-skewed input data, increasing the number of bins may improve accuracy.

n_streamsint (default = 4 )

Number of parallel streams used for forest building

min_samples_leafint or float (default = 1)

The minimum number of samples (rows) in each leaf node. If int, then min_samples_leaf represents the minimum number. If float, then min_samples_leaf represents a fraction and ceil(min_samples_leaf * n_rows) is the minimum number of samples for each leaf node.

min_samples_splitint or float (default = 2)

The minimum number of samples required to split an internal node. If int, then min_samples_split represents the minimum number. If float, then min_samples_split represents a fraction and ceil(min_samples_split * n_rows) is the minimum number of samples for each split.

min_impurity_decreasefloat (default = 0.0)

The minimum decrease in impurity required for node to be split

accuracy_metricstring (default = ‘r2’)

Decides the metric used to evaluate the performance of the model. In the 0.16 release, the default scoring metric was changed from mean squared error to r-squared. for r-squared : ‘r2’ for median of abs error : ‘median_ae’ for mean of abs error : ‘mean_ae’ for mean square error’ : ‘mse’

use_experimental_backendboolean (default = True)

Deprecated and currrently has no effect. .. deprecated:: 21.08

max_batch_size: int (default = 4096)

Maximum number of nodes that can be processed in a given batch.

random_stateint (default = None)

Seed for the random number generator. Unseeded by default. Does not currently fully guarantee the exact same results. Note: Parameter `seed` is removed since release 0.19.

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

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=128,
                    min_samples_leaf=1,
                    min_samples_split=2,
                    n_estimators=40, accuracy_metric='r2')
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.
convert_to_treelite_model(self)	Converts the cuML RF model to a Treelite model
fit(self, X, y[, convert_dtype])	Perform Random Forest Regression on the input data
get_detailed_text(self)	Obtain the detailed information for the random forest model, as text
get_json(self)	Export the Random Forest model as a JSON string
get_summary_text(self)	Obtain the text summary of the random forest model
predict(self, X[, predict_model, algo, …])	Predicts the labels for X.
score(self, X, y[, algo, convert_dtype, …])	Calculates the accuracy metric score of the model for X.

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 = False)

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 coalescing-friendly

• '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=True)

Perform Random Forest Regression on the input data

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the train 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_detailed_text(self)

Obtain the detailed information for the random forest model, as text

get_json(self)

Export the Random Forest model as a JSON string

get_summary_text(self)

Obtain the text summary of the random forest model

predict(self, X, predict_model='GPU', algo='auto', convert_dtype=True, fil_sparse_format='auto') → CumlArray

Predicts the labels for X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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 coalescing-friendly

• '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

ycuDF, CuPy or NumPy object depending on cuML’s output typeconfiguration, shape =(n_samples, 1)

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. In the 0.16 release, the default scoring metric was changed from mean squared error to r-squared.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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 coalescing-friendly

• '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

Forest Inferencing

class cuml.ForestInference(*, handle=None, output_type=None, verbose=False)

ForestInference provides GPU-accelerated 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 treelite-compatible 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, otherwise (starting from 5000-12288 features) FIL might infer slower

• From sklearn.ensemble, only {RandomForest,GradientBoosting,ExtraTrees}{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 classification and regression are supported.

• Many other random forest implementations including LightGBM, and SKLearn GBDTs make use of 64-bit floating point parameters, but the underlying library for ForestInference uses only 32-bit parameters. Because of the truncation that will occur when loading such models into ForestInference, you may observe a slight degradation in accuracy.

Parameters

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

Notes

For additional usage examples, see the sample notebook at https://github.com/rapidsai/cuml/blob/branch-0.15/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

common_load_params_docstring(func)
common_predict_params_docstring(func)
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 scikit-learn model passed to the function.
load_from_treelite_model(self, model[, …])	Creates a FIL model using the treelite model
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, safe_dtype_conversion])	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.

common_load_params_docstring(func)

common_predict_params_docstring(func)

static load(filename, output_class=False, threshold=0.5, algo='auto', storage_type='auto', blocks_per_sm=0, threads_per_tree=1, n_items=0, compute_shape_str=False, 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 treelite-compatible format (See https://treelite.readthedocs.io/en/latest/treelite-api.html for more information)

output_class: boolean (default=False)

For a Classification model output_class must be True. For a Regression model output_class must be False.

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 coalescing-friendly

• '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 or boolean (default=’auto’)

In-memory storage format to be used for the FIL model:

• 'auto': Choose the storage type automatically (currently DENSE is always used)

• False: Create a dense forest

• True: Create a sparse forest. Requires algo=’NAIVE’ or algo=’AUTO’

blocks_per_sminteger (default=0)

(experimental) Indicates how the number of thread blocks to lauch for the inference kernel is determined.

• 0 (default): Launches the number of blocks proportional to the number of data rows

• >= 1: Attempts to lauch blocks_per_sm blocks per SM. This will fail if blocks_per_sm blocks result in more threads than the maximum supported number of threads per GPU. Even if successful, it is not guaranteed that blocks_per_sm blocks will run on an SM concurrently.

compute_shape_strboolean (default=False)

if True or equivalent, creates a ForestInference.shape_str (writes a human-readable forest shape description as a multiline ascii string)

model_typestring (default=”xgboost”)

Format of the saved treelite model to be load. It can be ‘xgboost’, ‘xgboost_json’, ‘lightgbm’.

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', blocks_per_sm=0, threads_per_tree=1, n_items=0, compute_shape_str=False, handle=None)

Creates a FIL model using the scikit-learn model passed to the function. This function requires Treelite 1.0.0+ to be installed.

Parameters

skl_model

The scikit-learn model from which to build the FIL version.

output_class: boolean (default=False)

For a Classification model output_class must be True. For a Regression model output_class must be False.

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 coalescing-friendly

• '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 or boolean (default=’auto’)

In-memory storage format to be used for the FIL model:

• 'auto': Choose the storage type automatically (currently DENSE is always used)

• False: Create a dense forest

• True: Create a sparse forest. Requires algo=’NAIVE’ or algo=’AUTO’

blocks_per_sminteger (default=0)

(experimental) Indicates how the number of thread blocks to lauch for the inference kernel is determined.

• 0 (default): Launches the number of blocks proportional to the number of data rows

• >= 1: Attempts to lauch blocks_per_sm blocks per SM. This will fail if blocks_per_sm blocks result in more threads than the maximum supported number of threads per GPU. Even if successful, it is not guaranteed that blocks_per_sm blocks will run on an SM concurrently.

compute_shape_strboolean (default=False)

if True or equivalent, creates a ForestInference.shape_str (writes a human-readable forest shape description as a multiline ascii string)

Returns

fil_model

A Forest Inference model created from the scikit-learn model passed.

load_from_treelite_model(self, model, output_class=False, algo='auto', threshold=0.5, storage_type='auto', blocks_per_sm=0, threads_per_tree=1, n_items=0, compute_shape_str=False)

Creates a FIL model using the treelite model

passed to the function.

Parameters

model

the trained model information in the treelite format loaded from a saved model using the treelite API https://treelite.readthedocs.io/en/latest/treelite-api.html

output_class: boolean (default=False)

For a Classification model output_class must be True. For a Regression model output_class must be False.

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 coalescing-friendly

• '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 or boolean (default=’auto’)

In-memory storage format to be used for the FIL model:

• 'auto': Choose the storage type automatically (currently DENSE is always used)

• False: Create a dense forest

• True: Create a sparse forest. Requires algo=’NAIVE’ or algo=’AUTO’

blocks_per_sminteger (default=0)

(experimental) Indicates how the number of thread blocks to lauch for the inference kernel is determined.

• 0 (default): Launches the number of blocks proportional to the number of data rows

• >= 1: Attempts to lauch blocks_per_sm blocks per SM. This will fail if blocks_per_sm blocks result in more threads than the maximum supported number of threads per GPU. Even if successful, it is not guaranteed that blocks_per_sm blocks will run on an SM concurrently.

compute_shape_strboolean (default=False)

if True or equivalent, creates a ForestInference.shape_str (writes a human-readable forest shape description as a multiline ascii string)

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, blocks_per_sm=0, threads_per_tree=1, n_items=0, compute_shape_str=False)

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/treelite-api.html for more information)

output_class: boolean (default=False)

For a Classification model output_class must be True. For a Regression model output_class must be False.

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 coalescing-friendly

• '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 or boolean (default=’auto’)

In-memory storage format to be used for the FIL model:

• 'auto': Choose the storage type automatically (currently DENSE is always used)

• False: Create a dense forest

• True: Create a sparse forest. Requires algo=’NAIVE’ or algo=’AUTO’

blocks_per_sminteger (default=0)

(experimental) Indicates how the number of thread blocks to lauch for the inference kernel is determined.

• 0 (default): Launches the number of blocks proportional to the number of data rows

• >= 1: Attempts to lauch blocks_per_sm blocks per SM. This will fail if blocks_per_sm blocks result in more threads than the maximum supported number of threads per GPU. Even if successful, it is not guaranteed that blocks_per_sm blocks will run on an SM concurrently.

compute_shape_strboolean (default=False)

if True or equivalent, creates a ForestInference.shape_str (writes a human-readable forest shape description as a multiline ascii string)

Returns

fil_model

A Forest Inference model which can be used to perform inferencing on the random forest model.

predict(self, X, preds=None, safe_dtype_conversion=False) → CumlArray

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

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series. For optimal performance, pass a float device array with C-style layout preds : gpuarray or cudf.Series, shape = (n_samples,)

Optional ‘out’ location to store inference results

safe_dtype_conversionbool (default = False)

FIL converts data to np.float32 when needed. Set this parameter to True to enable checking for information loss during that conversion, but note that this check can have a significant performance penalty. Parameter will be dropped in a future version.

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, safe_dtype_conversion=False) → CumlArray

Predicts the class probabilities for X with the loaded forest model. The result is the raw floating point output from the model.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series. For optimal performance, pass a float device array with C-style layout preds : gpuarray or cudf.Series, shape = (n_samples,2)

Binary probability output Optional ‘out’ location to store inference results

safe_dtype_conversionbool (default = False)

FIL converts data to np.float32 when needed. Set this parameter to True to enable checking for information loss during that conversion, but note that this check can have a significant performance penalty. Parameter will be dropped in a future version.

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(*, loss='squared_loss', alpha=0.0001, l1_ratio=0.15, fit_intercept=True, normalize=False, max_iter=1000, tol=0.001, shuffle=True, handle=None, output_type=None, verbose=False)

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 elastic-net 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 = 1e-3)

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 1e-4.

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

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

Attributes

coef_

Python descriptor object to control getting/setting CumlArray attributes on Base objects. See the Estimator Guide for an in depth guide.

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) → u’CD’

Fit the model with X and y.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = False)

When set to True, the train 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) → CumlArray

Predicts the y for X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Predicted values

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

Quasi-Newton

class cuml.QN(*, loss='sigmoid', fit_intercept=True, l1_strength=0.0, l2_strength=0.0, max_iter=1000, tol=0.0001, delta=None, linesearch_max_iter=50, lbfgs_memory=5, verbose=False, handle=None, output_type=None, warm_start=False)

Quasi-Newton 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:

• Orthant-Wise Limited Memory Quasi-Newton (OWL-QN) if there is l1 regularization

• Limited Memory BFGS (L-BFGS) otherwise.

cuML’s QN class can take array-like 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 non-zero, will run OWL-QN, else L-BFGS). Note, that as in Scikit-learn, the bias will not be regularized.

l2_strength: float (default = 0.0)

l2 regularization strength. Note, that as in Scikit-learn, the bias will not be regularized.

max_iter: int (default = 1000)

Maximum number of iterations taken for the solvers to converge.

tol: float (default = 1e-4)

The training process will stop if

norm(current_loss_grad, inf) <= tol * max(current_loss, tol).

This differs slightly from the gtol-controlled stopping condition in scipy.optimize.minimize(method=’L-BFGS-B’):

norm(current_loss_projected_grad, inf) <= gtol.

Note, sklearn.LogisticRegression() uses the sum of softmax/logistic loss over the input data, whereas cuML uses the average. As a result, Scikit-learn’s loss is usually sample_size times larger than cuML’s. To account for the differences you may divide the tol by the sample size; this would ensure that the cuML solver does not stop earlier than the Scikit-learn solver.

delta: Optional[float] (default = None)

The training process will stop if

abs(current_loss - previous_loss) <= delta * max(current_loss, tol).

When None, it’s set to tol * 0.01; when 0, the check is disabled. Given the current step k, parameter previous_loss here is the loss at the step k - p, where p is a small positive integer set internally.

Note, this parameter corresponds to ftol in scipy.optimize.minimize(method=’L-BFGS-B’), which is set by default to a miniscule 2.2e-9 and is not exposed in sklearn.LogisticRegression(). This condition is meant to protect the solver against doing vanishingly small linesearch steps or zigzagging. You may choose to set delta = 0 to make sure the cuML solver does not stop earlier than the Scikit-learn solver.

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 inverse-Hessian approximation. Method will use O(lbfgs_memory * D) memory.

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

warm_startbool, default=False

When set to True, reuse the solution of the previous call to fit as initialization, otherwise, just erase the previous solution.

Notes

This class contains implementations of two popular Quasi-Newton methods:

• Limited-memory Broyden Fletcher Goldfarb Shanno (L-BFGS) [Nocedal, Wright - Numerical Optimization (1999)]

• Orthant-wise limited-memory quasi-newton (OWL-QN) [Andrew, Gao - ICML 2007]

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_)
print("Intercept:")
print(solver.intercept_)

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)

QN.coef_(self)

intercept_array (n_classes, 1)

The independent term. If fit_intercept is False, will be 0.

Methods

fit(self, X, y[, sample_weight, 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)

property coef_

fit(self, X, y, sample_weight=None, convert_dtype=False) → u’QN’

Fit the model with X and y.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense or sparse matrix containing floats or doubles. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

sample_weightarray-like (device or host) shape = (n_samples,), default=None

The weights for each observation in X. If None, all observations are assigned equal weight. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = False)

When set to True, the train 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) → CumlArray

Predicts the y for X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense or sparse matrix containing floats or doubles. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Predicted values

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

score(self, X, y)

Support Vector Machines

class cuml.svm.SVC(C-Support Vector Classification)

Construct an SVC classifier for training and predictions.

Parameters

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.

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 = 1e-3)

Tolerance for stopping criterion.

cache_sizefloat (default = 1024.0)

Size of the kernel cache during training in MiB. 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.

class_weightdict or string (default=None)

Weights to modify the parameter C for class i to class_weight[i]*C. The string ‘balanced’ is also accepted, in which case class_weight[i] = n_samples / (n_classes * n_samples_of_class[i])

max_iterint (default = 100*n_samples)

Limit the number of outer iterations in the solver

multiclass_strategystr (‘ovo’ or ‘ovr’, default ‘ovo’)

Multiclass classification strategy. 'ovo' uses OneVsOneClassifier while 'ovr' selects OneVsRestClassifier

nochange_stepsint (default = 1000)

We monitor how much our stopping criteria changes during outer iterations. If it does not change (changes less then 1e-3*tol) for nochange_steps consecutive steps, then we stop training.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

probability: bool (default = False)

Enable or disable probability estimates.

random_state: int (default = None)

Seed for random number generator (used only when probability = True). Currently this argument is not used and a waring will be printed if the user provides it.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

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].

For additional docs, see scikitlearn’s SVC.

References

1

J. Vanek et al. A GPU-Architecture 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, 1-5 (2018)

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 support 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_float

SVC.intercept_(self)

fit_status_int

0 if SVM is correctly fitted

coef_float, shape (1, n_cols)

SVMBase.coef_(self)

classes_: shape (n_classes_,)

Array of class labels

n_classes_int

Number of classes

Methods

decision_function(self, X)	Calculates the decision function values for X.
fit(self, X, y[, sample_weight, convert_dtype])	Fit the model with X and y.
get_param_names(self)
predict(self, X[, convert_dtype])	Predicts the class labels for X. The returned y values are the class
predict_log_proba(self, X)	Predicts the log probabilities for X (returns log(predict_proba(x)).
predict_proba(self, X[, log])	Predicts the class probabilities for X.

property classes_

decision_function(self, X) → CumlArray

Calculates the decision function values for X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

Returns

resultscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Decision function values

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

fit(self, X, y, sample_weight=None, convert_dtype=True) → u’SVC’

Fit the model with X and y.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix of any dtype. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

sample_weightarray-like (device or host) shape = (n_samples,), default=None

The weights for each observation in X. If None, all observations are assigned equal weight. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the train 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)

property intercept_

predict(self, X, convert_dtype=True) → CumlArray

Predicts the class labels for X. The returned y values are the class labels associated to sign(decision_function(X)).

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Predicted values

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

predict_log_proba(self, X) → CumlArray

Predicts the log probabilities for X (returns log(predict_proba(x)).

The model has to be trained with probability=True to use this method.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

Returns

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, n_classes)

Log of predicted probabilities

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

predict_proba(self, X, log=False) → CumlArray

Predicts the class probabilities for X.

The model has to be trained with probability=True to use this method.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

log: boolean (default = False)

Whether to return log probabilities.

Returns

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, n_classes)

Predicted probabilities

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

class cuml.svm.SVR(Epsilon Support Vector Regression)

Construct an SVC classifier for training and predictions.

Parameters

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.

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 = 1e-3)

Tolerance for stopping criterion.

epsilon: float (default = 0.1)

epsilon parameter of the epsiron-SVR model. There is no penalty associated to points that are predicted within the epsilon-tube around the target values.

cache_sizefloat (default = 1024.0)

Size of the kernel cache during training in MiB. 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 1e-3*tol) for nochange_steps consecutive steps, then we stop training.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

Notes

For additional docs, see Scikit-learn’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 GPU-Architecture 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, 1-5 (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

SVMBase.intercept_(self)

fit_status_int

0 if SVM is correctly fitted

coef_float, shape [1, n_cols]

SVMBase.coef_(self)

Methods

fit(self, X, y[, sample_weight, convert_dtype])	Fit the model with X and y.
predict(self, X[, convert_dtype])	Predicts the values for X.

fit(self, X, y, sample_weight=None, convert_dtype=True) → u’SVR’

Fit the model with X and y.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

sample_weightarray-like (device or host) shape = (n_samples,), default=None

The weights for each observation in X. If None, all observations are assigned equal weight. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the train 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=True) → CumlArray

Predicts the values for X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Predicted values

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

Nearest Neighbors Classification

class cuml.neighbors.KNeighborsClassifier(*, weights='uniform', handle=None, verbose=False, output_type=None, **kwargs)

K-Nearest Neighbors Classifier is an instance-based 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

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.

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

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)

Output:

array([3, 1, 1, 0, 2, 0, 0, 0, 0, 0, 0, 0, 1, 2, 3, 1, 0, 0, 0, 2, 3, 3,
       0, 3, 0, 0, 0, 0, 3, 2, 0, 0, 0], dtype=int32)

Attributes

classes_

Python descriptor object to control getting/setting CumlArray attributes on Base objects. See the Estimator Guide for an in depth guide.

y

Python descriptor object to control getting/setting CumlArray attributes on Base objects. See the Estimator Guide for an in depth guide.

Methods

fit(self, X, y[, convert_dtype])	Fit a GPU index for k-nearest neighbors classifier model.
get_param_names(self)
predict(self, X[, convert_dtype])	Use the trained k-nearest neighbors classifier to
predict_proba(self, X[, convert_dtype])	Use the trained k-nearest neighbors classifier to

fit(self, X, y, convert_dtype=True) → u’KNeighborsClassifier’

Fit a GPU index for k-nearest neighbors classifier model.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the method will automatically convert the inputs to np.float32.

get_param_names(self)

predict(self, X, convert_dtype=True) → CumlArray

Use the trained k-nearest neighbors classifier to predict the labels for X

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the method will automatically convert the inputs to np.float32.

Returns

X_newcuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Labels predicted

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

predict_proba(self, X, convert_dtype=True) → typing.Union[CumlArray, typing.Tuple]

Use the trained k-nearest neighbors classifier to predict the label probabilities for X

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the method will automatically convert the inputs to np.float32.

Returns

X_newcuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Labels probabilities

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

Nearest Neighbors Regression

class cuml.neighbors.KNeighborsRegressor(*, weights='uniform', handle=None, verbose=False, output_type=None, **kwargs)

K-Nearest Neighbors Regressor is an instance-based learning technique, that keeps training samples around for prediction, rather than trying to learn a generalizable set of model parameters.

The K-Nearest 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

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.

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

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.        ])

Attributes

y

Python descriptor object to control getting/setting CumlArray attributes on Base objects. See the Estimator Guide for an in depth guide.

Methods

fit(self, X, y[, convert_dtype])	Fit a GPU index for k-nearest neighbors regression model.
get_param_names(self)
predict(self, X[, convert_dtype])	Use the trained k-nearest neighbors regression model to

fit(self, X, y, convert_dtype=True) → u’KNeighborsRegressor’

Fit a GPU index for k-nearest neighbors regression model.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the method will automatically convert the inputs to np.float32.

get_param_names(self)

predict(self, X, convert_dtype=True) → CumlArray

Use the trained k-nearest neighbors regression model to predict the labels for X

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the method will automatically convert the inputs to np.float32.

Returns

X_newcuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, n_features)

Predicted values

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

Clustering

K-Means Clustering

class cuml.KMeans(*, handle=None, n_clusters=8, max_iter=300, tol=0.0001, verbose=False, random_state=1, init='scalable-k-means++', n_init=1, oversampling_factor=2.0, max_samples_per_batch=32768, output_type=None)

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 array-like object or cuDF DataFrame, and supports the scalable KMeans++ initialization method. This method is more stable than randomly selecting K points.

Parameters

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.

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 = 1e-4)

Stopping criterion when centroid means do not change much.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

random_stateint (default = 1)

If you want results to be the same when you restart Python, select a state.

init‘scalable-kmeans++’, ‘k-means||’ , ‘random’ or an ndarray (default = ‘scalable-k-means++’) # noqa

‘scalable-k-means++’ or ‘k-means||’: 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 k-means 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 (default = 2.0)

The amount of points to sample in scalable k-means++ 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 k-means++ 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.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

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 T-SNE, 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)
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 k-means 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 cluster-distance space.
get_param_names(self)
predict(self, X[, convert_dtype, …])	Predict the closest cluster each sample in X belongs to.
score(self, X[, y, sample_weight, convert_dtype])	Opposite of the value of X on the K-means objective.
transform(self, X[, convert_dtype])	Transform X to a cluster-distance space.

fit(self, X, sample_weight=None) → u’KMeans’

Compute k-means clustering with X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

sample_weightarray-like (device or host) shape = (n_samples,), default=None

The weights for each observation in X. If None, all observations are assigned equal weight. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

fit_predict(self, X, sample_weight=None) → CumlArray

Compute cluster centers and predict cluster index for each sample.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

sample_weightarray-like (device or host) shape = (n_samples,), default=None

The weights for each observation in X. If None, all observations are assigned equal weight. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

Returns

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Cluster indexes

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

fit_transform(self, X, convert_dtype=False, sample_weight=None) → CumlArray

Compute clustering and transform X to cluster-distance space.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = False)

When set to True, the fit_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.

sample_weightarray-like (device or host) shape = (n_samples,), default=None

The weights for each observation in X. If None, all observations are assigned equal weight. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

Returns

X_newcuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, n_clusters)

Transformed data

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

get_param_names(self)

predict(self, X, convert_dtype=False, sample_weight=None, normalize_weights=True) → CumlArray

Predict the closest cluster each sample in X belongs to.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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.

sample_weightarray-like (device or host) shape = (n_samples,), default=None

The weights for each observation in X. If None, all observations are assigned equal weight. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

Returns

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Cluster indexes

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

score(self, X, y=None, sample_weight=None, convert_dtype=True)

Opposite of the value of X on the K-means objective.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

sample_weightarray-like (device or host) shape = (n_samples,), default=None

The weights for each observation in X. If None, all observations are assigned equal weight. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the score 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

scorefloat

Opposite of the value of X on the K-means objective.

transform(self, X, convert_dtype=False) → CumlArray

Transform X to a cluster-distance space.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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.

Returns

X_newcuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, n_clusters)

Transformed data

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

DBSCAN

class cuml.DBSCAN(*, eps=0.5, handle=None, min_samples=5, metric='euclidean', verbose=False, max_mbytes_per_batch=None, output_type=None, calc_core_sample_indices=True)

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 array-like 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

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.

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).

metric: {‘euclidean’, ‘precomputed’}, default = ‘euclidean’

The metric to use when calculating distances between points. If metric is ‘precomputed’, X is assumed to be a distance matrix and must be square.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

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 trade-off 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{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

calc_core_sample_indices(optional) boolean (default = True)

Indicates whether the indices of the core samples should be calculated. The the attribute core_sample_indices_ will not be used, setting this to False will avoid unnecessary kernel launches

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 non-linearly 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_array-like 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.

core_sample_indices_array-like or cuDF series

The indices of the core samples. Only calculated if calc_core_sample_indices==True

Methods

fit(self, X[, out_dtype])	Perform DBSCAN clustering from features.
fit_predict(self, X[, out_dtype])	Performs clustering on X and returns cluster labels.
get_param_names(self)

fit(self, X, out_dtype='int32') → u’DBSCAN’

Perform DBSCAN clustering from features.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

out_dtype: dtype Determines the precision of the output labels array.

default: “int32”. Valid values are { “int32”, np.int32, “int64”, np.int64}.

fit_predict(self, X, out_dtype='int32') → CumlArray

Performs clustering on X and returns cluster labels.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

out_dtype: dtype Determines the precision of the output labels array.

default: “int32”. Valid values are { “int32”, np.int32, “int64”, np.int64}.

Returns

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Cluster labels

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

get_param_names(self)

Agglomerative Clustering

class cuml.AgglomerativeClustering(*, n_clusters=2, affinity='euclidean', linkage='single', handle=None, verbose=False, connectivity='knn', n_neighbors=10, output_type=None)

Agglomerative Clustering

Recursively merges the pair of clusters that minimally increases a given linkage distance.

Parameters

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.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

n_clustersint (default = 2)

The number of clusters to find.

affinitystr, default=’euclidean’

Metric used to compute the linkage. Can be “euclidean”, “l1”, “l2”, “manhattan”, or “cosine”. If connectivity is “knn” only “euclidean” is accepted.

linkage{“single”}, default=”single”

Which linkage criterion to use. The linkage criterion determines which distance to use between sets of observations. The algorithm will merge the pairs of clusters that minimize this criterion. - ‘single’ uses the minimum of the distances between all

observations of the two sets.

n_neighborsint (default = 15)

The number of neighbors to compute when connectivity = “knn”

connectivity{“pairwise”, “knn”}, (default = “knn”)

The type of connectivity matrix to compute. - ‘pairwise’ will compute the entire fully-connected graph of

pairwise distances between each set of points. This is the fastest to compute and can be very fast for smaller datasets but requires O(n^2) space.

• ‘knn’ will sparsify the fully-connected connectivity matrix to save memory and enable much larger inputs. “n_neighbors” will control the amount of memory used and the graph will be connected automatically in the event “n_neighbors” was not large enough to connect it.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

Attributes

children_

Python descriptor object to control getting/setting CumlArray attributes on Base objects. See the Estimator Guide for an in depth guide.

labels_

Python descriptor object to control getting/setting CumlArray attributes on Base objects. See the Estimator Guide for an in depth guide.

Methods

fit(X[, y, convert_dtype])	Agglomerat
fit_predict(X[, y])	Agglomerat
get_param_names(self)

fit(X, y=None, convert_dtype=True) → cuml.cluster.agglomerative.AgglomerativeClustering

Agglomerat

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = True)

When set to True, the train 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.

ering.fit(self, X, y=None, convert_dtype=True) -> u’AgglomerativeClustering’

Fit the hierarchical clustering from features.

fit_predict(X, y=None) → cuml.common.array.CumlArray

Agglomerat

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

ering.fit_predict(self, X, y=None) -> CumlArray

Fit the hierarchical clustering from features and return cluster labels.

Returns

predscuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, 1)

Cluster indexes

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

get_param_names(self)

Dimensionality Reduction and Manifold Learning

Principal Component Analysis

class cuml.PCA(*, copy=True, handle=None, iterated_power=15, n_components=None, random_state=None, svd_solver='auto', tol=1e-07, verbose=False, whiten=False, output_type=None)

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 array-like 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

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.

iterated_powerint (default = 15)

Used in Jacobi solver. The more iterations, the more accurate, but slower.

n_componentsint (default = None)

The number of top K singular vectors / values you want. Must be <= number(columns). If n_components is not set, then all components are kept:

n_components = min(n_samples, n_features)

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 = 1e-7)

Used if algorithm = “jacobi”. Smaller tolerance can increase accuracy, but but will slow down the algorithm’s convergence.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

whitenboolean (default = False)

If True, de-correlates 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 multi-collinearity. It might be beneficial for downstream tasks like LinearRegression where correlated features cause problems.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

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 T-SNE 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. y is currently ignored.
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) → u’PCA’

Fit the model with X. y is currently ignored.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense or sparse matrix containing floats or doubles. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

fit_transform(self, X, y=None) → CumlArray

Fit the model with X and apply the dimensionality reduction on X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense or sparse matrix containing floats or doubles. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

Returns

transcuDF, CuPy or NumPy object depending on cuML’s output type configuration, cupyx.scipy.sparse for sparse output, shape = (n_samples, n_components)

Transformed values

For more information on how to configure cuML’s dense output type, refer to: Output Data Type Configuration.

get_param_names(self)

inverse_transform(self, X, convert_dtype=False, return_sparse=False, sparse_tol=1e-10) → CumlArray

Transform data back to its original space.

In other words, return an input X_original whose transform would be X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense or sparse matrix containing floats or doubles. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = False)

When set to True, the inverse_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.

return_sparsebool, optional (default = False)

Ignored when the model is not fit on a sparse matrix If True, the method will convert the result to a cupyx.scipy.sparse.csr_matrix object. NOTE: Currently, there is a loss of information when converting to csr matrix (cusolver bug). Default will be switched to True once this is solved.

sparse_tolfloat, optional (default = 1e-10)

Ignored when return_sparse=False. If True, values in the inverse transform below this parameter are clipped to 0.

Returns

X_invcuDF, CuPy or NumPy object depending on cuML’s output type configuration, cupyx.scipy.sparse for sparse output, shape = (n_samples, n_features)

Transformed values

For more information on how to configure cuML’s dense output type, refer to: Output Data Type Configuration.

transform(self, X, convert_dtype=False) → CumlArray

Apply dimensionality reduction to X.

X is projected on the first principal components previously extracted from a training set.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense or sparse matrix containing floats or doubles. Acceptable dense formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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.

Returns

transcuDF, CuPy or NumPy object depending on cuML’s output type configuration, cupyx.scipy.sparse for sparse output, shape = (n_samples, n_components)

Transformed values

For more information on how to configure cuML’s dense output type, refer to: Output Data Type Configuration.

Incremental PCA

class cuml.IncrementalPCA(*, handle=None, n_components=None, whiten=False, copy=True, batch_size=None, verbose=False, output_type=None)

Based on sklearn.decomposition.IncrementalPCA from scikit-learn 0.23.1

Incremental principal components analysis (IPCA). Linear dimensionality reduction using Singular Value Decomposition of the data, keeping only the most significant singular vectors to project the data to a lower dimensional space. The input data is centered but not scaled for each feature before applying the SVD. Depending on the size of the input data, this algorithm can be much more memory efficient than a PCA, and allows sparse input. This algorithm has constant memory complexity, on the order of batch_size * n_features, enabling use of np.memmap files without loading the entire file into memory. For sparse matrices, the input is converted to dense in batches (in order to be able to subtract the mean) which avoids storing the entire dense matrix at any one time. The computational overhead of each SVD is O(batch_size * n_features ** 2), but only 2 * batch_size samples remain in memory at a time. There will be n_samples / batch_size SVD computations to get the principal components, versus 1 large SVD of complexity O(n_samples * n_features ** 2) for PCA.

Parameters

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.

n_componentsint or None, (default=None)

Number of components to keep. If n_components is None, then n_components is set to min(n_samples, n_features).

whitenbool, optional

If True, de-correlates 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 multi-collinearity. It might be beneficial for downstream tasks like LinearRegression where correlated features cause problems.

copybool, (default=True)

If False, X will be overwritten. copy=False can be used to save memory but is unsafe for general use.

batch_sizeint or None, (default=None)

The number of samples to use for each batch. Only used when calling fit. If batch_size is None, then batch_size is inferred from the data and set to 5 * n_features, to provide a balance between approximation accuracy and memory consumption.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

Notes

Implements the incremental PCA model from [1]. This model is an extension of the Sequential Karhunen-Loeve Transform from [2]. We have specifically abstained from an optimization used by authors of both papers, a QR decomposition used in specific situations to reduce the algorithmic complexity of the SVD. The source for this technique is [3]. This technique has been omitted because it is advantageous only when decomposing a matrix with n_samples >= 5/3 * n_features where n_samples and n_features are the matrix rows and columns, respectively. In addition, it hurts the readability of the implemented algorithm. This would be a good opportunity for future optimization, if it is deemed necessary.

References

1

D. Ross, J. Lim, R. Lin, M. Yang. Incremental Learning for Robust Visual Tracking, International Journal of Computer Vision, Volume 77, Issue 1-3, pp. 125-141, May 2008.

2

A. Levy and M. Lindenbaum, Sequential Karhunen-Loeve Basis Extraction and its Application to Images, IEEE Transactions on Image Processing, Volume 9, Number 8, pp. 1371-1374, August 2000.

3

G. Golub and C. Van Loan. Matrix Computations, Third Edition, Chapter 5, Section 5.4.4, pp. 252-253.

4

C. Bishop, 1999. “Pattern Recognition and Machine Learning”, Section 12.2.1, pp. 574

Examples

>>> from cuml.decomposition import IncrementalPCA
>>> import cupy as cp
>>> import cupyx
>>>
>>> X = cupyx.scipy.sparse.random(1000, 4, format='csr', density=0.07)
>>> ipca = IncrementalPCA(n_components=2, batch_size=200)
>>> ipca.fit(X)
>>>
>>> # Components:
>>> ipca.components_
array([[-0.02362926,  0.87328851, -0.15971988,  0.45967206],
    [-0.14643883,  0.11414225,  0.97589354,  0.11471273]])
>>>
>>> # Singular Values:
>>> ipca.singular_values_
array([4.90298662, 4.54498226])
>>>
>>> # Explained Variance:
>>> ipca.explained_variance_
array([0.02406334, 0.02067754])
>>>
>>> # Explained Variance Ratio:
>>> ipca.explained_variance_ratio_
array([0.28018011, 0.24075775])
>>>
>>> # Mean:
>>> ipca.mean_
array([0.03249896, 0.03629852, 0.03268694, 0.03216601])
>>>
>>> # Noise Variance:
>>> ipca.noise_variance_.item()
0.003474966583315544

Attributes

components_array, shape (n_components, n_features)

Components with maximum variance.

explained_variance_array, shape (n_components,)

Variance explained by each of the selected components.

explained_variance_ratio_array, shape (n_components,)

Percentage of variance explained by each of the selected components. If all components are stored, the sum of explained variances is equal to 1.0.

singular_values_array, shape (n_components,)

The singular values corresponding to each of the selected components. The singular values are equal to the 2-norms of the n_components variables in the lower-dimensional space.

mean_array, shape (n_features,)

Per-feature empirical mean, aggregate over calls to partial_fit.

var_array, shape (n_features,)

Per-feature empirical variance, aggregate over calls to partial_fit.

noise_variance_float

The estimated noise covariance following the Probabilistic PCA model from [4].

n_components_int

The estimated number of components. Relevant when n_components=None.

n_samples_seen_int

The number of samples processed by the estimator. Will be reset on new calls to fit, but increments across partial_fit calls.

batch_size_int

Inferred batch size from batch_size.

Methods

fit(X[, y])	Fit the model with X, using minibatches of size batch_size.
get_param_names(self)
partial_fit(X[, y, check_input])	Incremental fit with X.
transform(X[, convert_dtype])	Apply dimensionality reduction to X.

fit(X, y=None) → cuml.decomposition.incremental_pca.IncrementalPCA

Fit the model with X, using minibatches of size batch_size.

Parameters

Xarray-like or sparse matrix, shape (n_samples, n_features)

Training data, where n_samples is the number of samples and n_features is the number of features.

yIgnored

Returns

selfobject

Returns the instance itself.

get_param_names(self)

partial_fit(X, y=None, check_input=True) → cuml.decomposition.incremental_pca.IncrementalPCA

Incremental fit with X. All of X is processed as a single batch.

Parameters

Xarray-like or sparse matrix, shape (n_samples, n_features)

Training data, where n_samples is the number of samples and n_features is the number of features.

check_inputbool

Run check_array on X.

yIgnored

Returns

selfobject

Returns the instance itself.

transform(X, convert_dtype=False) → cuml.common.array.CumlArray

Apply dimensionality reduction to X.

X is projected on the first principal components previously extracted from a training set, using minibatches of size batch_size if X is sparse.

Parameters

Xarray-like or sparse matrix, shape (n_samples, n_features)

New data, where n_samples is the number of samples and n_features is the number of features.

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_newarray-like, shape (n_samples, n_components)

Truncated SVD

class cuml.TruncatedSVD(*, algorithm='full', handle=None, n_components=1, n_iter=15, random_state=None, tol=1e-07, verbose=False, output_type=None)

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 array-like 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

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.

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 = 1e-7)

Used if algorithm = “jacobi”. Smaller tolerance can increase accuracy, but but will slow down the algorithm’s convergence.

verboseint or boolean, default=False

Sets logging level. It must be one of cuml.common.logger.level_*. See Verbosity Levels for more info.

output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’, ‘numba’}, default=None

Variable to control output type of the results and attributes of the estimator. If None, it’ll inherit the output type set at the module level, cuml.global_settings.output_type. See Output Data Type Configuration for more info.

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 = 1e-9)
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. y is currently ignored.
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) → u’TruncatedSVD’

Fit LSI model on training cudf DataFrame X. y is currently ignored.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

fit_transform(self, X, y=None) → CumlArray

Fit LSI model to X and perform dimensionality reduction on X. y is currently ignored.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

yarray-like (device or host) shape = (n_samples, 1)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

Returns

transcuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, n_components)

Reduced version of X

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

get_param_names(self)

inverse_transform(self, X, convert_dtype=False) → CumlArray

Transform X back to its original space. Returns X_original whose transform would be X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

convert_dtypebool, optional (default = False)

When set to True, the inverse_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.

Returns

X_originalcuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, n_features)

X in original space

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

transform(self, X, convert_dtype=False) → CumlArray

Perform dimensionality reduction on X.

Parameters

Xarray-like (device or host) shape = (n_samples, n_features)

Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.

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.

Returns

X_newcuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_samples, n_components)

Reduced version of X

For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

UMAP

class cuml.UMAP(*, n_neighbors=15, n_components=2, n_epochs=None, learning_rate=1.0, min_dist=0.1, spread=1.0, set_op_mix_ratio=1.0, local_connectivity=1.0, repulsion_strength=1.0, negative_sample_rate=5, transform_queue_size=4.0, init='spectral', verbose=False, a=None, b=None, target_n_neighbors=- 1, target_weight=0.5, target_metric='categorical', handle=None, hash_input=False, random_state=None, callback=None, output_type=None)

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

The UMAP algorithm is outlined in [1]. This implementation follows the GPU-accelerated version as described in [2].

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 1-skeleton

• ‘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 t-norm. 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 and spread.

b: float (optional, default None)

More specific parameters controlling the embedding. If None these values are set automatically as determined by min_dist and spread.

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.

hash_input: bool, optional (default = False)

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 CPU-based 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.

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 floating-point 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

Used 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(