cuML API Reference¶
Module Configuration¶
Output Data Type Configuration¶
memory_utils.
set_global_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'>
memory_utils.
using_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.core.core.ndarray'>
Verbosity Levels¶
cuML follows a verbosity model similar to Scikitlearn’s: The verbose parameter can be a boolean, or a numeric value, and higher numeric values mean more verbosity. The exact values can be set directly, or through the cuml.common.logger module, and they are:
Numeric value 
cuml.common.logger value 
Verbosity level 

0 
cuml.common.logger.level_off 
Disables all log messages 
1 
cuml.common.logger.level_critical 
Enables only critical messages 
2 
cuml.common.logger.level_error 
Enables all messages up to and including errors. 
3 
cuml.common.logger.level_warn 
Enables all messages up to and including warnings. 
4 or False 
cuml.common.logger.level_info 
Enables all messages up to and including information messages. 
5 or True 
cuml.common.logger.level_debug 
Enables all messages up to and including debug messages. 
6 
cuml.common.logger.level_trace 
Enables all messages up to and including trace messages. 
Preprocessing, Metrics, and Utilities¶
Model Selection and Data Splitting¶
model_selection.
train_test_split
(y=None, test_size: Union[float, int] = None, train_size: Union[float, int] = None, shuffle: bool = True, random_state: Union[int, cupy.random.generator.RandomState, numpy.random.mtrand.RandomState] = None, seed: Union[int, cupy.random.generator.RandomState, numpy.random.mtrand.RandomState] = None, stratify=None)¶Partitions device data into four collated objects, mimicking Scikitlearn’s train_test_split.
 Parameters
 Xcudf.DataFrame or cuda_array_interface compliant device array
Data to split, has shape (n_samples, n_features)
 ystr, cudf.Series or cuda_array_interface compliant device array
Set of labels for the data, either a series of shape (n_samples) or the string label of a column in X (if it is a cuDF DataFrame) containing the labels
 train_sizefloat or int, optional
If float, represents the proportion [0, 1] of the data to be assigned to the training set. If an int, represents the number of instances to be assigned to the training set. Defaults to 0.8
 shufflebool, optional
Whether or not to shuffle inputs before splitting
 random_stateint, CuPy RandomState or NumPy RandomState optional
If shuffle is true, seeds the generator. Unseeded by default
 seed: random_stateint, CuPy RandomState or NumPy RandomState optional
Deprecated in favor of random_state. If shuffle is true, seeds the generator. Unseeded by default
 stratify: bool, optional
Whether to stratify the input data based on class labels. None by default
 Returns
 X_train, X_test, y_train, y_testcudf.DataFrame or arraylike objects
Partitioned dataframes if X and y were cuDF objects. If y was provided as a column name, the column was dropped from X. Partitioned numba device arrays if X and y were Numba device arrays. Partitioned CuPy arrays for any other input.
Examples
import cudf from cuml.preprocessing.model_selection import train_test_split # Generate some sample data df = cudf.DataFrame({'x': range(10), 'y': [0, 1] * 5}) print(f'Original data: {df.shape[0]} elements') # Suppose we want an 80/20 split X_train, X_test, y_train, y_test = train_test_split(df, 'y', train_size=0.8) print(f'X_train: {X_train.shape[0]} elements') print(f'X_test: {X_test.shape[0]} elements') print(f'y_train: {y_train.shape[0]} elements') print(f'y_test: {y_test.shape[0]} elements') # Alternatively, if our labels are stored separately labels = df['y'] df = df.drop(['y']) # we can also do X_train, X_test, y_train, y_test = train_test_split(df, labels, train_size=0.8)Output:
Original data: 10 elements X_train: 8 elements X_test: 2 elements y_train: 8 elements y_test: 2 elements
Feature and Label Encoding (SingleGPU)¶
 class
cuml.preprocessing.LabelEncoder.
LabelEncoder
(handle_unknown='error')¶An nvcategory based implementation of ordinal label encoding
 Parameters
 handle_unknown{‘error’, ‘ignore’}, default=’error’
Whether to raise an error or ignore if an unknown categorical feature is present during transform (default is to raise). When this parameter is set to ‘ignore’ and an unknown category is encountered during transform or inverse transform, the resulting encoding will be null.
Examples
Converting a categorical implementation to a numerical one
from cudf import DataFrame, Series data = DataFrame({'category': ['a', 'b', 'c', 'd']}) # There are two functionally equivalent ways to do this le = LabelEncoder() le.fit(data.category) # le = le.fit(data.category) also works encoded = le.transform(data.category) print(encoded) # This method is preferred le = LabelEncoder() encoded = le.fit_transform(data.category) print(encoded) # We can assign this to a new column data = data.assign(encoded=encoded) print(data.head()) # We can also encode more data test_data = Series(['c', 'a']) encoded = le.transform(test_data) print(encoded) # After train, ordinal label can be inverse_transform() back to # string labels ord_label = cudf.Series([0, 0, 1, 2, 1]) ord_label = dask_cudf.from_cudf(data, npartitions=2) str_label = le.inverse_transform(ord_label) print(str_label)Output:
0 0 1 1 2 2 3 3 dtype: int64 0 0 1 1 2 2 3 3 dtype: int32 category encoded 0 a 0 1 b 1 2 c 2 3 d 3 0 2 1 0 dtype: int64 0 a 1 a 2 b 3 c 4 b dtype: objectMethods
fit
(y)Fit a LabelEncoder (nvcategory) instance to a set of categories
Simultaneously fit and transform an input
Revert ordinal label to original label
transform
(y)Transform an input into its categorical keys.
fit
(y)¶Fit a LabelEncoder (nvcategory) instance to a set of categories
 Parameters
 ycudf.Series
Series containing the categories to be encoded. It’s elements may or may not be unique
 Returns
 selfLabelEncoder
A fitted instance of itself to allow method chaining
fit_transform
(y: cudf.core.series.Series) → cudf.core.series.Series¶Simultaneously fit and transform an input
This is functionally equivalent to (but faster than) LabelEncoder().fit(y).transform(y)
inverse_transform
(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)¶A multiclass dummy encoder for labels.
Examples
Create an array with labels and dummy encode them
import cupy as cp from cuml.preprocessing import LabelBinarizer labels = cp.asarray([0, 5, 10, 7, 2, 4, 1, 0, 0, 4, 3, 2, 1], dtype=cp.int32) lb = LabelBinarizer() encoded = lb.fit_transform(labels) print(str(encoded) decoded = lb.inverse_transform(encoded) print(str(decoded)Output:
[[1 0 0 0 0 0 0 0] [0 0 0 0 0 1 0 0] [0 0 0 0 0 0 0 1] [0 0 0 0 0 0 1 0] [0 0 1 0 0 0 0 0] [0 0 0 0 1 0 0 0] [0 1 0 0 0 0 0 0] [1 0 0 0 0 0 0 0] [1 0 0 0 0 0 0 0] [0 0 0 0 1 0 0 0] [0 0 0 1 0 0 0 0] [0 0 1 0 0 0 0 0] [0 1 0 0 0 0 0 0]] [ 0 5 10 7 2 4 1 0 0 4 3 2 1]Methods
fit
(y)Fit label binarizer
Fit label binarizer and transform multiclass labels to their dummyencoded representation.
inverse_transform
(y[, threshold])Transform binary labels back to original multiclass labels
transform
(y)Transform multiclass labels to their dummyencoded representation labels.
fit
(y)¶Fit label binarizer
 Parameters
 yarray of shape [n_samples,] or [n_samples, n_classes]
Target values. The 2d matrix should only contain 0 and 1, represents multilabel classification.
 Returns
 selfreturns an instance of self.
fit_transform
(y)¶Fit label binarizer and transform multiclass labels to their dummyencoded representation.
 Parameters
 yarray of shape [n_samples,] or [n_samples, n_classes]
 Returns
 arrarray with encoded labels
inverse_transform
(y, threshold=None)¶Transform binary labels back to original multiclass labels
 Parameters
 yarray of shape [n_samples, n_classes]
 thresholdfloat this value is currently ignored
 Returns
 arrarray with original labels
transform
(y)¶Transform multiclass labels to their dummyencoded representation labels.
 Parameters
 yarray of shape [n_samples,] or [n_samples, n_classes]
 Returns
 arrarray with encoded labels
preprocessing.
label_binarize
(classes, neg_label=0, pos_label=1, sparse_output=False)¶A stateless helper function to dummy encode multiclass labels.
 Parameters
 yarraylike of size [n_samples,] or [n_samples, n_classes]
 classesthe set of unique classes in the input
 neg_labelinteger the negative value for transformed output
 pos_labelinteger the positive value for transformed output
 sparse_outputbool whether to return sparse array
 class
cuml.preprocessing.
OneHotEncoder
(categories='auto', drop=None, sparse=True, dtype=<class 'float'>, handle_unknown='error')¶Encode categorical features as a onehot numeric array. The input to this estimator should be a cuDF.DataFrame or a cupy.ndarray, denoting the unique values taken on by categorical (discrete) features. The features are encoded using a onehot (aka ‘oneofK’ or ‘dummy’) encoding scheme. This creates a binary column for each category and returns a sparse matrix or dense array (depending on the
sparse
parameter). By default, the encoder derives the categories based on the unique values in each feature. Alternatively, you can also specify the categories manually.Note
a onehot encoding of y labels should use a LabelBinarizer instead.
 Parameters
 categories‘auto’ an cupy.ndarray or a cudf.DataFrame, default=’auto’
Categories (unique values) per feature:
‘auto’ : Determine categories automatically from the training data.
DataFrame/ndarray :
categories[col]
holds the categories expected in the feature col. drop‘first’, None, a dict or a list, default=None
Specifies a methodology to use to drop one of the categories per feature. This is useful in situations where perfectly collinear features cause problems, such as when feeding the resulting data into a neural network or an unregularized regression.
None : retain all features (the default).
‘first’ : drop the first category in each feature. If only one category is present, the feature will be dropped entirely.
dict/list :
drop[col]
is the category in feature col that should be dropped. sparsebool, default=False
This feature was deactivated and will give an exception when True. The reason is because sparse matrix are not fully supported by cupy yet, causing incorrect values when computing one hot encodings. See https://github.com/cupy/cupy/issues/3223
 dtypenumber type, default=np.float
Desired datatype of transform’s output.
 handle_unknown{‘error’, ‘ignore’}, default=’error’
Whether to raise an error or ignore if an unknown categorical feature is present during transform (default is to raise). When this parameter is set to ‘ignore’ and an unknown category is encountered during transform, the resulting onehot encoded columns for this feature will be all zeros. In the inverse transform, an unknown category will be denoted as None.
 Attributes
 drop_idx_array of shape (n_features,)
drop_idx_[i]
is the index incategories_[i]
of the category to be dropped for each feature. None if all the transformed features will be retained.Methods
fit
(X)Fit OneHotEncoder to X.
Fit OneHotEncoder to X, then transform X.
Convert the data back to the original representation.
transform
(X)Transform X using onehot 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)¶Fit OneHotEncoder to X, then transform X. Equivalent to fit(X).transform(X).
 Parameters
 Xcudf.DataFrame or cupy.ndarray, shape = (n_samples, n_features)
The data to encode.
 Returns
 X_outsparse matrix if sparse=True else a 2d array
Transformed input.
inverse_transform
(X)¶Convert the data back to the original representation. In case unknown categories are encountered (all zeros in the onehot encoding),
None
is used to represent this category.The return type is the same as the type of the input used by the first call to fit on this estimator instance.
 Parameters
 Xarraylike or sparse matrix, shape [n_samples, n_encoded_features]
The transformed data.
 Returns
 X_trcudf.DataFrame or cupy.ndarray
Inverse transformed array.
transform
(X)¶Transform X using onehot encoding.
 Parameters
 Xcudf.DataFrame or cupy.ndarray
The data to encode.
 Returns
 X_outsparse matrix if sparse=True else a 2d array
Transformed input.
 class
cuml.preprocessing.
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.
 smoothfloat (default=0)
0 <= smooth <= 1 Percentage of samples to smooth the encoding
 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
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
 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
Feature and Label Encoding (Daskbased MultiGPU)¶
 class
cuml.dask.preprocessing.
LabelBinarizer
(client=None, **kwargs)¶A distributed version of LabelBinarizer for onehot encoding a collection of labels.
Examples
Create an array with labels and dummy encode them
import cupy as cp from cuml.dask.preprocessing import LabelBinarizer from dask_cuda import LocalCUDACluster from dask.distributed import Client import dask cluster = LocalCUDACluster() client = Client(cluster) labels = cp.asarray([0, 5, 10, 7, 2, 4, 1, 0, 0, 4, 3, 2, 1], dtype=cp.int32) labels = dask.array.from_array(labels) lb = LabelBinarizer() encoded = lb.fit_transform(labels) print(str(encoded.compute()) decoded = lb.inverse_transform(encoded) print(str(decoded.compute())Output:
[[1 0 0 0 0 0 0 0] [0 0 0 0 0 1 0 0] [0 0 0 0 0 0 0 1] [0 0 0 0 0 0 1 0] [0 0 1 0 0 0 0 0] [0 0 0 0 1 0 0 0] [0 1 0 0 0 0 0 0] [1 0 0 0 0 0 0 0] [1 0 0 0 0 0 0 0] [0 0 0 0 1 0 0 0] [0 0 0 1 0 0 0 0] [0 0 1 0 0 0 0 0] [0 1 0 0 0 0 0 0]] [ 0 5 10 7 2 4 1 0 0 4 3 2 1]Methods
fit
(y)Fit label binarizer
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 2d 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 2d 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 onehot numeric array. The input to this transformer should be a dask_cuDF.DataFrame or cupy dask.Array, denoting the values taken on by categorical features. The features are encoded using a onehot (aka ‘oneofK’ or ‘dummy’) encoding scheme. This creates a binary column for each category and returns a sparse matrix or dense array (depending on the
sparse
parameter). By default, the encoder derives the categories based on the unique values in each feature. Alternatively, you can also specify the categories manually.
 Parameters
 categories‘auto’, cupy.ndarray or cudf.DataFrame, default=’auto’
Categories (unique values) per feature. All categories are expected to fit on one GPU.
‘auto’ : Determine categories automatically from the training data.
DataFrame/ndarray :
categories[col]
holds the categories expected in the feature col. drop‘first’, None or a dict, default=None
Specifies a methodology to use to drop one of the categories per feature. This is useful in situations where perfectly collinear features cause problems, such as when feeding the resulting data into a neural network or an unregularized regression.
None : retain all features (the default).
‘first’ : drop the first category in each feature. If only one category is present, the feature will be dropped entirely.
Dict :
drop[col]
is the category in feature col that should be dropped. sparsebool, default=False
This feature was deactivated and will give an exception when True. The reason is because sparse matrix are not fully supported by cupy yet, causing incorrect values when computing one hot encodings. See https://github.com/cupy/cupy/issues/3223
 dtypenumber type, default=np.float
Desired datatype of transform’s output.
 handle_unknown{‘error’, ‘ignore’}, default=’error’
Whether to raise an error or ignore if an unknown categorical feature is present during transform (default is to raise). When this parameter is set to ‘ignore’ and an unknown category is encountered during transform, the resulting onehot encoded columns for this feature will be all zeros. In the inverse transform, an unknown category will be denoted as None.
Methods
fit
(X)Fit a multinode multigpu 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 onehot encoding.
fit
(X)¶Fit a multinode multigpu OneHotEncoder to X.
 Parameters
 XDask cuDF DataFrame or CuPy backed Dask Array
The data to determine the categories of each feature.
 Returns
 self
fit_transform
(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 onehot encoding),
None
is used to represent this category.
 Parameters
 XCuPy backed Dask Array, shape [n_samples, n_encoded_features]
The transformed data.
 delayedbool (default = True)
Whether to execute as a delayed task or eager.
 Returns
 X_trDask cuDF DataFrame or CuPy backed Dask Array
Distributed object containing the inverse transformed array.
transform
(X, delayed=True)¶Transform X using onehot encoding.
 Parameters
 XDask cuDF DataFrame or CuPy backed Dask Array
The data to encode.
 delayedbool (default = True)
Whether to execute as a delayed task or eager.
 Returns
 outDask cuDF DataFrame or CuPy backed Dask Array
Distributed object containing the transformed input.
Feature Extraction (SingleGPU)¶
 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 apriori 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 ngrams generation steps.
 stop_wordsstring {‘english’}, list, or None (default)
If ‘english’, a builtin 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 nvalues for different word ngrams or char ngrams 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 ngram or character ngrams. Option ‘char_wb’ creates character ngrams only from text inside word boundaries; ngrams 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 (corpusspecific 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 cutoff 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 documentterm matrix.
Array mapping from feature integer indices to feature name.
Return terms per document with nonzero entries in X.
transform
(raw_documents)Transform documents to documentterm 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 documentterm 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)
Documentterm 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
 Xarraylike of shape (n_samples, n_features)
Documentterm matrix.
 Returns
 X_invlist of cudf.Series of shape (n_samples,)
List of Series of terms.
transform
(raw_documents)¶Transform documents to documentterm 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)
Documentterm 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 cupy.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 unpickle 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 inmemory 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 32bit 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 ngrams generation steps.
 stop_wordsstring {‘english’}, list, default=None
If ‘english’, a builtin 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 nvalues for different word ngrams or char ngrams 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 ngram or character ngrams. Option ‘char_wb’ creates character ngrams only from text inside word boundaries; ngrams 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
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 documentterm 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 documentterm 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 documentterm 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)
Documentterm 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 documentterm 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)
Documentterm 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 TFIDF features.
Equivalent to
CountVectorizer
followed byTfidfTransformer
.
 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 ngrams generation steps.
 stop_wordsstring {‘english’}, list, or None (default)
If ‘english’, a builtin 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 nvalues for different word ngrams or char ngrams 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 ngram or character ngrams. Option ‘char_wb’ creates character ngrams only from text inside word boundaries; ngrams 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 (corpusspecific 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 cutoff 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 inversedocumentfrequency 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 scikitlearn 0.23.1’s TfIdfVectorizer code, which is provided under the BSD3 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 documentterm matrix.
transform
(raw_documents)Transform documents to documentterm 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 documentterm 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)
Tfidfweighted documentterm matrix.
transform
(raw_documents)¶Transform documents to documentterm 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)
Tfidfweighted documentterm matrix.
Dataset Generation (SingleGPU)¶
 random_state
Determines random number generation for dataset creation. Pass an int for reproducible output across multiple function calls.
datasets.
make_blobs
(n_features=2, centers=None, cluster_std=1.0, center_box= 10.0, 10.0, shuffle=True, random_state=None, return_centers=False, order='F', dtype='float32')¶Generate isotropic Gaussian blobs for clustering.
 Parameters
 n_samplesint or arraylike, optional (default=100)
If int, it is the total number of points equally divided among clusters. If arraylike, each element of the sequence indicates the number of samples per cluster.
 n_featuresint, optional (default=2)
The number of features for each sample.
 centersint or array of shape [n_centers, n_features], optional
(default=None) The number of centers to generate, or the fixed center locations. If n_samples is an int and centers is None, 3 centers are generated. If n_samples is arraylike, centers must be either None or an array of length equal to the length of n_samples.
 cluster_stdfloat or sequence of floats, optional (default=1.0)
The standard deviation of the clusters.
 center_boxpair of floats (min, max), optional (default=(10.0, 10.0))
The bounding box for each cluster center when centers are generated at random.
 shuffleboolean, optional (default=True)
Shuffle the samples.
 random_stateint, RandomState instance, default=None
Determines random number generation for dataset creation. Pass an int for reproducible output across multiple function calls.
 return_centersbool, optional (default=False)
If True, then return the centers of each cluster
 order: str, optional (default=’F’)
The order of the generated samples
 dtypestr, optional (default=’float32’)
Dtype of the generated samples
 Returns
 Xdevice array of shape [n_samples, n_features]
The generated samples.
 ydevice array of shape [n_samples]
The integer labels for cluster membership of each sample.
 centersdevice array, shape [n_centers, n_features]
The centers of each cluster. Only returned if
return_centers=True
.See also
make_classification
a more intricate variant
Examples
>>> from sklearn.datasets import make_blobs >>> X, y = make_blobs(n_samples=10, centers=3, n_features=2, ... random_state=0) >>> print(X.shape) (10, 2) >>> y array([0, 0, 1, 0, 2, 2, 2, 1, 1, 0]) >>> X, y = make_blobs(n_samples=[3, 3, 4], centers=None, n_features=2, ... random_state=0) >>> print(X.shape) (10, 2) >>> y array([0, 1, 2, 0, 2, 2, 2, 1, 1, 0])
datasets.
make_classification
(n_features=20, n_informative=2, n_redundant=2, n_repeated=0, n_classes=2, n_clusters_per_class=2, weights=None, flip_y=0.01, class_sep=1.0, hypercube=True, shift=0.0, scale=1.0, shuffle=True, random_state=None, order='F', dtype='float32', _centroids=None, _informative_covariance=None, _redundant_covariance=None, _repeated_indices=None)¶Generate a random nclass classification problem. This initially creates clusters of points normally distributed (std=1) about vertices of an
n_informative
dimensional hypercube with sides of length2*class_sep
and assigns an equal number of clusters to each class. It introduces interdependence between these features and adds various types of further noise to the data. Without shuffling,X
horizontally stacks features in the following order: the primaryn_informative
features, followed byn_redundant
linear combinations of the informative features, followed byn_repeated
duplicates, drawn randomly with replacement from the informative and redundant features. The remaining features are filled with random noise. Thus, without shuffling, all useful features are contained in the columnsX[:, :n_informative + n_redundant + n_repeated]
.
 Parameters
 n_samplesint, optional (default=100)
The number of samples.
 n_featuresint, optional (default=20)
The total number of features. These comprise
n_informative
informative features,n_redundant
redundant features,n_repeated
duplicated features andn_featuresn_informativen_redundantn_repeated
useless features drawn at random. n_informativeint, optional (default=2)
The number of informative features. Each class is composed of a number of gaussian clusters each located around the vertices of a hypercube in a subspace of dimension
n_informative
. For each cluster, informative features are drawn independently from N(0, 1) and then randomly linearly combined within each cluster in order to add covariance. The clusters are then placed on the vertices of the hypercube. n_redundantint, optional (default=2)
The number of redundant features. These features are generated as random linear combinations of the informative features.
 n_repeatedint, optional (default=0)
The number of duplicated features, drawn randomly from the informative and the redundant features.
 n_classesint, optional (default=2)
The number of classes (or labels) of the classification problem.
 n_clusters_per_classint, optional (default=2)
The number of clusters per class.
 weightsarraylike of shape (n_classes,) or (n_classes  1,), (default=None)
The proportions of samples assigned to each class. If None, then classes are balanced. Note that if
len(weights) == n_classes  1
, then the last class weight is automatically inferred. More thann_samples
samples may be returned if the sum ofweights
exceeds 1. flip_yfloat, optional (default=0.01)
The fraction of samples whose class is assigned randomly. Larger values introduce noise in the labels and make the classification task harder.
 class_sepfloat, optional (default=1.0)
The factor multiplying the hypercube size. Larger values spread out the clusters/classes and make the classification task easier.
 hypercubeboolean, optional (default=True)
If True, the clusters are put on the vertices of a hypercube. If False, the clusters are put on the vertices of a random polytope.
 shiftfloat, array of shape [n_features] or None, optional (default=0.0)
Shift features by the specified value. If None, then features are shifted by a random value drawn in [class_sep, class_sep].
 scalefloat, array of shape [n_features] or None, optional (default=1.0)
Multiply features by the specified value. If None, then features are scaled by a random value drawn in [1, 100]. Note that scaling happens after shifting.
 shuffleboolean, optional (default=True)
Shuffle the samples and the features.
 random_stateint, RandomState instance or None (default)
Determines random number generation for dataset creation. Pass an int for reproducible output across multiple function calls. See Glossary.
 order: str, optional (default=’F’)
The order of the generated samples
 dtypestr, optional (default=’float32’)
Dtype of the generated samples
 _centroids: array of centroids of shape (n_clusters, n_informative)
 _informative_covariance: array for covariance between informative features
of shape (n_clusters, n_informative, n_informative)
 _redundant_covariance: array for covariance between redundant features
of shape (n_informative, n_redundant)
 _repeated_indices: array of indices for the repeated features
of shape (n_repeated, )
 Returns
 Xdevice array of shape [n_samples, n_features]
The generated samples.
 ydevice array of shape [n_samples]
The integer labels for class membership of each sample.
Notes
The algorithm is adapted from Guyon [1] and was designed to generate the “Madelon” dataset. How we optimized for GPUs:
Firstly, we generate X from a standard univariate instead of zeros. This saves memory as we don’t need to generate univariates each time for each feature class (informative, repeated, etc.) while also providing the added speedup of generating a big matrix on GPU
We generate order=F construction. We exploit the fact that X is a generated from a univariate normal, and covariance is introduced with matrix multiplications. Which means, we can generate X as a 1D array and just reshape it to the desired order, which only updates the metadata and eliminates copies
Lastly, we also shuffle by construction. Centroid indices are permuted for each sample, and then we construct the data for each centroid. This shuffle works for both order=C and order=F and eliminates any need for secondary copies
References
 1
I. Guyon, “Design of experiments for the NIPS 2003 variable selection benchmark”, 2003.
Examples
from cuml.datasets.classification import make_classification X, y = make_classification(n_samples=10, n_features=4, n_informative=2, n_classes=2) print("X:") print(X) print("y:") print(y)Output:
X: [[2.3249989 0.8679415 1.1511791 1.3525577 ] [ 2.2933831 1.3743551 0.63128835 0.84648645] [ 1.6361488 1.3233329 0.807027 0.894092 ] [1.0093077 0.9990691 0.00808992 0.00950443] [ 0.99803793 2.068382 0.49570698 0.8462848 ] [1.2750955 0.9725835 0.2390058 0.28081596] [1.3635055 0.9637669 0.31582272 0.37106958] [ 1.1893625 2.227583 0.48750278 0.8737561 ] [0.05753583 1.0939395 0.8188342 0.9620734 ] [ 0.47910076 0.7648213 0.17165393 0.26144698]] y: [0 1 0 0 1 0 0 1 0 1]
datasets.
make_regression
(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)¶Generate a random regression problem.
See https://scikitlearn.org/stable/modules/generated/sklearn.datasets.make_regression.html
 Parameters
 n_samplesint, optional (default=100)
The number of samples.
 n_featuresint, optional (default=2)
The number of features.
 n_informativeint, optional (default=2)
The number of informative features, i.e., the number of features used to build the linear model used to generate the output.
 n_targetsint, optional (default=1)
The number of regression targets, i.e., the dimension of the y output vector associated with a sample. By default, the output is a scalar.
 biasfloat, optional (default=0.0)
The bias term in the underlying linear model.
 effective_rankint or None, optional (default=None)
 if not None:
The approximate number of singular vectors required to explain most of the input data by linear combinations. Using this kind of singular spectrum in the input allows the generator to reproduce the correlations often observed in practice.
 if None:
The input set is well conditioned, centered and gaussian with unit variance.
 tail_strengthfloat between 0.0 and 1.0, optional (default=0.5)
The relative importance of the fat noisy tail of the singular values profile if effective_rank is not None.
 noisefloat, optional (default=0.0)
The standard deviation of the gaussian noise applied to the output.
 shuffleboolean, optional (default=True)
Shuffle the samples and the features.
 coefboolean, optional (default=False)
If True, the coefficients of the underlying linear model are returned.
 random_stateint, RandomState instance or None (default)
Seed for the random number generator for dataset creation.
 dtype: string or numpy dtype (default: ‘single’)
Type of the data. Possible values: float32, float64, ‘single’, ‘float’ or ‘double’.
 handle: cuml.Handle
If it is None, a new one is created just for this function call
 Returns
 outdevice array of shape [n_samples, n_features]
The input samples.
 valuesdevice array of shape [n_samples, n_targets]
The output values.
 coefdevice array of shape [n_features, n_targets], optional
The coefficient of the underlying linear model. It is returned only if coef is True.
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_)
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', output_type='cupy', 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’
 output_type: {‘cudf’, ‘cupy’, ‘numpy’}
Type of the returned dataset
 handle: cuml.Handle
If it is None, a new one is created just for this function call
 Returns
 out: arraylike, 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 (Daskbased MultiGPU)¶
cuml.dask.datasets.blobs.
make_blobs
(n_samples=100, n_features=2, centers=None, cluster_std=1.0, n_parts=None, center_box= 10, 10, shuffle=True, random_state=None, return_centers=False, verbose=False, order='F', dtype='float32', client=None)¶Makes labeled DaskCupy arrays containing blobs for a randomly generated set of centroids.
This function calls make_blobs from cuml.datasets on each Dask worker and aggregates them into a single Dask Dataframe.
For more information on Scikitlearn’s make_blobs:.
 Parameters
 n_samplesint
number of rows
 n_featuresint
number of features
 centersint or array of shape [n_centers, n_features],
optional (default=None) The number of centers to generate, or the fixed center locations. If n_samples is an int and centers is None, 3 centers are generated. If n_samples is arraylike, centers must be either None or an array of length equal to the length of n_samples.
 cluster_stdfloat (default = 1.0)
standard deviation of points around centroid
 n_partsint (default = None)
number of partitions to generate (this can be greater than the number of workers)
 center_boxtuple (int, int) (default = (10, 10))
the bounding box which constrains all the centroids
 random_stateint (default = None)
sets random seed (or use None to reinitialize each time)
 return_centersbool, optional (default=False)
If True, then return the centers of each cluster
 verboseint or boolean (default = False)
Logging level.
 shufflebool (default=False)
Shuffles the samples on each worker.
 order: str, optional (default=’F’)
The order of the generated samples
 dtypestr, optional (default=’float32’)
Dtype of the generated samples
 clientdask.distributed.Client (optional)
Dask client to use
 Returns
 Xdask.array backed by CuPy array of shape [n_samples, n_features]
The input samples.
 ydask.array backed by CuPy array of shape [n_samples]
The output values.
 centersdask.array backed by CuPy array of shape
[n_centers, n_features], optional The centers of the underlying blobs. It is returned only if return_centers is True.
cuml.dask.datasets.classification.
make_classification
(n_samples=100, n_features=20, n_informative=2, n_redundant=2, n_repeated=0, n_classes=2, n_clusters_per_class=2, weights=None, flip_y=0.01, class_sep=1.0, hypercube=True, shift=0.0, scale=1.0, shuffle=True, random_state=None, order='F', dtype='float32', n_parts=None, client=None)¶Generate a random nclass classification problem.
This initially creates clusters of points normally distributed (std=1) about vertices of an n_informativedimensional 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 columnsX[:, :n_informative + n_redundant + n_repeated]
.
 Parameters
 n_samplesint, optional (default=100)
The number of samples.
 n_featuresint, optional (default=20)
The total number of features. These comprise n_informative informative features, n_redundant redundant features, n_repeated duplicated features and
n_featuresn_informativen_redundantn_repeated
useless features drawn at random. n_informativeint, optional (default=2)
The number of informative features. Each class is composed of a number of gaussian clusters each located around the vertices of a hypercube in a subspace of dimension n_informative. For each cluster, informative features are drawn independently from N(0, 1) and then randomly linearly combined within each cluster in order to add covariance. The clusters are then placed on the vertices of the hypercube.
 n_redundantint, optional (default=2)
The number of redundant features. These features are generated as random linear combinations of the informative features.
 n_repeatedint, optional (default=0)
The number of duplicated features, drawn randomly from the informative and the redundant features.
 n_classesint, optional (default=2)
The number of classes (or labels) of the classification problem.
 n_clusters_per_classint, optional (default=2)
The number of clusters per class.
 weightsarraylike of shape
(n_classes,)
or(n_classes  1,)
, (default=None)The proportions of samples assigned to each class. If None, then classes are balanced. Note that if
len(weights) == n_classes  1
, then the last class weight is automatically inferred. More 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:
We generate centroids of shape
(n_centroids, n_informative)
We generate an informative covariance of shape
(n_centroids, n_informative, n_informative)
We generate a redundant covariance of shape
(n_informative, n_redundant)
We generate the indices for the repeated features We pass along the references to the futures of the above arrays with each part to the single GPU cuml.datasets.classification.make_classification so that each part (and worker) has access to the correct values to generate data from the same covariances
Examples
from dask.distributed import Client from dask_cuda import LocalCUDACluster from cuml.dask.datasets.classification import make_classification cluster = LocalCUDACluster() client = Client(cluster) X, y = make_classification(n_samples=10, n_features=4, n_informative=2, n_classes=2) print("X:") print(X.compute()) print("y:") print(y.compute())Output:
X: [[1.6990056 0.8241044 0.06997631 0.45107925] [1.8105277 1.7829906 0.492909 0.05390119] [0.18290454 0.6155432 0.6667889 1.0053712 ] [2.7530136 0.888528 0.5023055 1.3983376 ] [0.9788184 0.89851004 0.10802134 0.10021686] [0.76883423 1.0689086 0.01249526 0.1404741 ] [1.5676656 0.83082974 0.03072987 0.34499463] [0.9381793 1.0971068 0.07465998 0.02618019] [1.3021476 0.87076336 0.02249984 0.15187258] [ 1.1820307 1.7524253 1.5087451 2.4626074 ]] y: [0 1 0 0 0 0 0 0 0 1]
cuml.dask.datasets.regression.
make_low_rank_matrix
(n_samples=100, n_features=100, effective_rank=10, tail_strength=0.5, random_state=None, n_parts=1, n_samples_per_part=None, dtype='float32')¶Generate a mostly low rank matrix with bellshaped singular values
 Parameters
 n_samplesint, optional (default=100)
The number of samples.
 n_featuresint, optional (default=100)
The number of features.
 effective_rankint, optional (default=10)
The approximate number of singular vectors required to explain most of the data by linear combinations.
 tail_strengthfloat between 0.0 and 1.0, optional (default=0.5)
The relative importance of the fat noisy tail of the singular values profile.
 random_stateint, CuPy RandomState instance, Dask RandomState instance or None (default)
Determines random number generation for dataset creation. Pass an int for reproducible output across multiple function calls.
 n_partsint, optional (default=1)
The number of parts of work.
 dtype: str, optional (default=’float32’)
dtype of generated data
 Returns
 XDaskCuPy array of shape [n_samples, n_features]
The matrix.
cuml.dask.datasets.regression.
make_regression
(n_samples=100, n_features=100, n_informative=10, n_targets=1, bias=0.0, effective_rank=None, tail_strength=0.5, noise=0.0, shuffle=False, coef=False, random_state=None, n_parts=1, n_samples_per_part=None, order='F', dtype='float32', client=None, use_full_low_rank=True)¶Generate a random regression problem.
The input set can either be well conditioned (by default) or have a low rankfat tail singular profile.
The output is generated by applying a (potentially biased) random linear regression model with “n_informative” nonzero regressors to the previously generated input and some gaussian centered noise with some adjustable scale.
 Parameters
 n_samplesint, optional (default=100)
The number of samples.
 n_featuresint, optional (default=100)
The number of features.
 n_informativeint, optional (default=10)
The number of informative features, i.e., the number of features used to build the linear model used to generate the output.
 n_targetsint, optional (default=1)
The number of regression targets, i.e., the dimension of the y output vector associated with a sample. By default, the output is a scalar.
 biasfloat, optional (default=0.0)
The bias term in the underlying linear model.
 effective_rankint or None, optional (default=None)
 if not None:
The approximate number of singular vectors required to explain most of the input data by linear combinations. Using this kind of singular spectrum in the input allows the generator to reproduce the correlations often observed in practice.
 if None:
The input set is well conditioned, centered and gaussian with unit variance.
 tail_strengthfloat between 0.0 and 1.0, optional (default=0.5)
The relative importance of the fat noisy tail of the singular values profile if “effective_rank” is not None.
 noisefloat, optional (default=0.0)
The standard deviation of the gaussian noise applied to the output.
 shuffleboolean, optional (default=False)
Shuffle the samples and the features.
 coefboolean, optional (default=False)
If True, the coefficients of the underlying linear model are returned.
 random_stateint, CuPy RandomState instance, Dask RandomState instance or None (default)
Determines random number generation for dataset creation. Pass an int for reproducible output across multiple function calls.
 n_partsint, optional (default=1)
The number of parts of work.
 orderstr, optional (default=’F’)
Rowmajor or Colmajor
 dtype: str, optional (default=’float32’)
dtype of generated data
 use_full_low_rankboolean (default=True)
Whether to use the entire dataset to generate the low rank matrix. If False, it creates a low rank covariance and uses the corresponding covariance to generate a multivariate normal distribution on the remaining chunks
 Returns
 XDaskCuPy array of shape [n_samples, n_features]
The input samples.
 yDaskCuPy array of shape [n_samples] or [n_samples, n_targets]
The output values.
 coefDaskCuPy array of shape [n_features] or [n_features, n_targets], optional
The coefficient of the underlying linear model. It is returned only if coef is True.
Notes
 Known Performance Limitations:
When effective_rank is set and use_full_low_rank is True, we cannot generate order F by construction, and an explicit transpose is performed on each part. This may cause memory to spike (other parameters make order F by construction)
When n_targets > 1 and order = ‘F’ as above, we have to explicity transpose the y array. If coef = True, then we also explicity transpose the ground_truth array
When shuffle = True and order = F, there are memory spikes to shuffle the F order arrays
Note
If outofmemory errors are encountered in any of the above configurations, try increasing the n_parts parameter.
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 arraylike object or integer representing a device or host pointer to preallocated 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.
 dtypedatatype, 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 Fmajor or Cmajor array.
Notes
cuml Array is not meant as an enduser 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 Fortranmajor or Cmajor 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.
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.
serialize

classmethod
empty
(shape, dtype, order='F')¶ Create an empty Array with an allocated but uninitialized DeviceBuffer
 Parameters
 dtypedatatype, 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 Fmajor or Cmajor array.

classmethod
full
(shape, value, dtype, order='F')¶ Create an Array with an allocated DeviceBuffer initialized to value.
 Parameters
 dtypedatatype, 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 Fmajor or Cmajor array.

classmethod
ones
(shape, dtype='float32', order='F')¶ Create an Array with an allocated DeviceBuffer initialized to zeros.
 Parameters
 dtypedatatype, 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 Fmajor or Cmajor 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
 dtypedatatype, 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 Fmajor or Cmajor 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_truearraylike (device or host) shape = (n_samples,)
or (n_samples, n_outputs) Ground truth (correct) target values.
 y_predarraylike (device or host) shape = (n_samples,)
or (n_samples, n_outputs) Estimated target values.
 sample_weightarraylike (device or host) shape = (n_samples,), optional
Sample weights.
 multioutputstring in [‘raw_values’, ‘uniform_average’]
or arraylike of shape (n_outputs) Defines aggregating of multiple output values. Arraylike value defines weights used to average errors. ‘raw_values’ : Returns a full set of errors in case of multioutput input. ‘uniform_average’ : Errors of all outputs are averaged with uniform weight.
 Returns
 lossfloat or ndarray of floats
If multioutput is ‘raw_values’, then mean absolute error is returned for each output separately. If multioutput is ‘uniform_average’ or an ndarray of weights, then the weighted average of all output errors is returned.
MAE output is nonnegative floating point. The best value is 0.0.
cuml.metrics.regression.
mean_squared_error
(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_truearraylike (device or host) shape = (n_samples,)
or (n_samples, n_outputs) Ground truth (correct) target values.
 y_predarraylike (device or host) shape = (n_samples,)
or (n_samples, n_outputs) Estimated target values.
 sample_weightarraylike (device or host) shape = (n_samples,), optional
Sample weights.
 multioutputstring in [‘raw_values’, ‘uniform_average’]
or arraylike of shape (n_outputs) Defines aggregating of multiple output values. Arraylike value defines weights used to average errors. ‘raw_values’ : Returns a full set of errors in case of multioutput input. ‘uniform_average’ : Errors of all outputs are averaged with uniform weight.
 squaredboolean value, optional (default = True)
If True returns MSE value, if False returns RMSE value.
 Returns
 lossfloat or ndarray of floats
A nonnegative floating point value (the best value is 0.0), or an array of floating point values, one for each individual target.
cuml.metrics.regression.
mean_squared_log_error
(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_truearraylike (device or host) shape = (n_samples,)
or (n_samples, n_outputs) Ground truth (correct) target values.
 y_predarraylike (device or host) shape = (n_samples,)
or (n_samples, n_outputs) Estimated target values.
 sample_weightarraylike (device or host) shape = (n_samples,), optional
Sample weights.
 multioutputstring in [‘raw_values’, ‘uniform_average’]
or arraylike of shape (n_outputs) Defines aggregating of multiple output values. Arraylike value defines weights used to average errors. ‘raw_values’ : Returns a full set of errors in case of multioutput input. ‘uniform_average’ : Errors of all outputs are averaged with uniform weight.
 squaredboolean value, optional (default = True)
If True returns MSE value, if False returns RMSE value.
 Returns
 lossfloat or ndarray of floats
A nonnegative floating point value (the best value is 0.0), or an array of floating point values, one for each individual target.
cuml.metrics.regression.
r2_score
(y, y_hat, convert_dtype=False, handle=None)¶Calculates r2 score between y and y_hat
 Parameters
 yarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 y_hatarraylike (device or host) shape = (n_samples, 1)
Dense vector (floats or doubles) of shape (n_samples, 1). Acceptable formats: cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 convert_dtypebool, optional (default = False)
When set to True, the fit method will, when necessary, convert y_hat to be the same data type as y if they differ. This will increase memory used for the method.
 Returns
 trustworthiness scoredouble
Trustworthiness of the lowdimensional embedding
cuml.metrics.accuracy.
accuracy_score
(ground_truth, predictions, handle=None, convert_dtype=True)¶Calcuates the accuracy score of a classification model.
 Parameters
 handlecuml.Handle
 predictionNumPy ndarray or Numba device
The labels predicted by the model for the test dataset
 ground_truthNumPy ndarray, Numba device
The ground truth labels of the test dataset
 Returns
 float
The accuracy of the model used for prediction
metrics.
confusion_matrix
(y_pred, labels=None, sample_weight=None, normalize=None)¶Compute confusion matrix to evaluate the accuracy of a classification.
 Parameters
 y_truearraylike (device or host) shape = (n_samples,)
or (n_samples, n_outputs) Ground truth (correct) target values.
 y_predarraylike (device or host) shape = (n_samples,)
or (n_samples, n_outputs) Estimated target values.
 labelsarraylike (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_weightarraylike (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.
 Returns
 Carraylike (device or host) shape = (n_classes, n_classes)
Confusion matrix.
metrics.
roc_auc_score
(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_truearraylike of shape (n_samples,)
True labels. The binary cases expect labels with shape (n_samples,)
 y_scorearraylike of shape (n_samples,)
Target scores. In the binary cases, these can be either probability estimates or nonthresholded 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
metrics.
precision_recall_curve
(probs_pred)¶Compute precisionrecall pairs for different probability thresholds
Note
this implementation is restricted to the binary classification task. The precision is the ratio
tp / (tp + fp)
wheretp
is the number of true positives andfp
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)
wheretp
is the number of true positives andfn
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 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, output_type=None, **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 scikitlearn: [‘cityblock’, ‘cosine’, ‘euclidean’, ‘l1’, ‘l2’, ‘manhattan’].
Sparse matrices are not supported.
 From scipy.spatial.distance: [‘sqeuclidean’]
See the documentation for scipy.spatial.distance for details on this metric. Sparse matrices are not supported.
 Parameters
 Xarraylike (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
 Yarraylike (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.
 output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’}, optional
Variable to control output type of the results of the function. If None, it’ll inherit the output type set at the module level, cuml.output_type. If set, the function will temporarily override the global option.
 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, MultiInput: >>> pairwise_distances(X, Y, metric='cosine') array([[0.4452998 , 0.13175686], [0.48550424, 0.15633851], [0.47000106, 0.14671817]]) >>> >>> # Manhattan Pairwise Distance, MultiInput: >>> pairwise_distances(X, Y, metric='manhattan') array([[ 4., 2.], [ 7., 5.], [12., 10.]])
Metrics (clustering and trustworthiness)¶
cuml.metrics.trustworthiness.
trustworthiness
(X, X_embedded, handle=None, n_neighbors=5, metric='euclidean', should_downcast=True, convert_dtype=False, batch_size=512)¶Expresses to what extent the local structure is retained in embedding. The score is defined in the range [0, 1].
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 X_embeddedarraylike (device or host) shape= (n_samples, n_features)
Acceptable formats: cuDF DataFrame, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy
 n_neighborsint, optional (default: 5)
Number of neighbors considered
 convert_dtypebool, optional (default = False)
When set to True, the trustworthiness method will automatically convert the inputs to np.float32.
 Returns
 trustworthiness scoredouble
Trustworthiness of the lowdimensional embedding
cuml.metrics.cluster.adjustedrandindex.
adjusted_rand_score
(labels_true, labels_pred, handle=None, convert_dtype=True)¶Adjusted_rand_score is a clustering similarity metric based on the Rand index and is corrected for chance.
 Parameters
 labels_trueGround truth labels to be used as a reference
labels_pred : Array of predicted labels used to evaluate the model
handle : cuml.Handle
 Returns
 float
The adjusted rand index value between 1.0 and 1.0
cuml.metrics.cluster.entropy.
cython_entropy
(clustering, base=None, handle=None)¶Computes the entropy of a distribution for given probability values.
 Parameters
 clusteringarraylike (device or host) shape = (n_samples,)
Clustering of labels. Probabilities are computed based on occurrences of labels. For instance, to represent a fair coin (2 equally possible outcomes), the clustering could be [0,1]. For a biased coin with 2/3 probability for tail, the clustering could be [0, 0, 1].
 base: float, optional
The logarithmic base to use, defaults to e (natural logarithm).
 handlecuml.Handle
Specifies the cuml.handle that holds internal CUDA state for computations in this model. Most importantly, this specifies the CUDA stream that will be used for the model’s computations, so users can run different models concurrently in different streams by creating handles in several streams. If it is None, a new one is created.
 Returns
 Sfloat
The calculated entropy.
cuml.metrics.cluster.homogeneity_score.
homogeneity_score
(labels_true, labels_pred, handle=None)¶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_predarraylike (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_truearraylike (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.completeness_score.
completeness_score
(labels_true, labels_pred, handle=None)¶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_predarraylike (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_truearraylike (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.
mutual_info_score
(labels_true, labels_pred, handle=None)¶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_predarraylike (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_truearraylike (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 nonnegative 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 cpubased algorithm (typically scikitlearn, but does not need to be as long as it offers fit and predict or transform methods). Provides mechanisms to run each version with default arguments. If no CPUbased version of the algorithm is available, pass None for the cpu_class when instantiating
 Parameters
 cpu_classclass
Class for CPU version of algorithm. Set to None if not available.
 cuml_classclass
Class for cuML algorithm
 shared_argsdict
Arguments passed to both implementations’s initializer
 cuml_argsdict
Arguments only passed to cuml’s initializer
 cpu_args dict
Arguments only passed to sklearn’s initializer
 accepts_labelsboolean
If True, the fit methods expects both X and y inputs. Otherwise, it expects only an X input.
 data_prep_hookfunction (data > data)
Optional function to run on input data before passing to fit
 accuracy_functionfunction (y_test, y_pred)
Function that returns a scalar representing accuracy
 bench_funccustom function to perform fit/predict/transform
calls.
Methods
run_cpu
(data, **override_args)Runs the cpubased algorithm’s fit method on specified data
run_cuml
(data, **override_args)Runs the cumlbased algorithm’s fit method on specified data
setup_cpu
setup_cuml
run_cpu
(data, **override_args)¶Runs the cpubased algorithm’s fit method on specified data
run_cuml
(data, **override_args)¶Runs the cumlbased algorithm’s fit method on specified data
cuml.benchmark.algorithms.
algorithm_by_name
(name)¶Returns the algorithm pair with the name ‘name’ (caseinsensitive)
cuml.benchmark.algorithms.
all_algorithms
()¶Returns all defined AlgorithmPair objects
Wrappers to run ML benchmarks
 class
cuml.benchmark.runners.
AccuracyComparisonRunner
(bench_rows, bench_dims, dataset_name='blobs', input_type='numpy', test_fraction=0.1, n_reps=1)¶Wrapper to run an algorithm with multiple dataset sizes and compute accuracy and speedup of cuml relative to sklearn baseline.
 class
cuml.benchmark.runners.
BenchmarkTimer
(reps=1)¶Provides a context manager that runs a code block reps times and records results to the instance variable timings. Use like:
timer = BenchmarkTimer(rep=5) for _ in timer.benchmark_runs(): ... do something ... print(np.min(timer.timings))Methods
benchmark_runs
 class
cuml.benchmark.runners.
SpeedupComparisonRunner
(bench_rows, bench_dims, dataset_name='blobs', input_type='numpy', n_reps=1)¶Wrapper to run an algorithm with multiple dataset sizes and compute speedup of cuml relative to sklearn baseline.
Methods
run
cuml.benchmark.runners.
run_variations
(algos, dataset_name, bench_rows, bench_dims, param_override_list=[{}], cuml_param_override_list=[{}], cpu_param_override_list=[{}], dataset_param_override_list=[{}], input_type='numpy', test_fraction=0.1, run_cpu=True, raise_on_error=False, n_reps=1)¶Runs each algo in algos once per bench_rows X bench_dims X params_override_list X cuml_param_override_list combination and returns a dataframe containing timing and accuracy data.
 Parameters
 algosstr or list
Name of algorithms to run and evaluate
 dataset_namestr
Name of dataset to use
 bench_rowslist of int
Dataset row counts to test
 bench_dimslist of int
Dataset column counts to test
 param_override_listlist of dict
Dicts containing parameters to pass to __init__. Each dict specifies parameters to override in one run of the algorithm.
 cuml_param_override_listlist of dict
Dicts containing parameters to pass to __init__ of the cuml algo only.
 cpu_param_override_listlist of dict
Dicts containing parameters to pass to __init__ of the cpu algo only.
 dataset_param_override_listdict
Dicts containing parameters to pass to dataset generator function
 test_fractionfloat
The fraction of data to use for testing.
 run_cpuboolean
If True, run the cpubased algorithm for comparison
Data generators for cuML benchmarks
The main entry point for consumers is gen_data, which wraps the underlying data generators.
Notes when writing new generators:
 Each generator is a function that accepts:
n_samples (set to 0 for ‘default’)
n_features (set to 0 for ‘default’)
random_state
(and optional generatorspecific parameters)
The function should return a 2tuple (X, y), where X is a Pandas dataframe and y is a Pandas series. If the generator does not produce labels, it can return (X, None)
A set of helper functions (convert_*) can convert these to alternative formats. Future revisions may support generating cudf dataframes or GPU arrays directly instead.
cuml.benchmark.datagen.
gen_data
(dataset_name, dataset_format, n_samples=0, n_features=0, random_state=42, test_fraction=0.0, **kwargs)¶Returns a tuple of data from the specified generator.
 Parameters
 dataset_namestr
Dataset to use. Can be a synthetic generator (blobs or regression) or a specified dataset (higgs currently, others coming soon)
 dataset_formatstr
Type of data to return. (One of cudf, numpy, pandas, gpuarray)
 n_samplesint
Number of samples to include in training set (regardless of test split)
 test_fractionfloat
Fraction of the dataset to partition randomly into the test set. If this is 0.0, no test set will be created.
cuml.benchmark.datagen.
load_higgs
()¶Returns the Higgs Boson dataset as an X, y tuple of dataframes.
Regression and Classification¶
Linear Regression¶

class
cuml.
LinearRegression
(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.
Notes
LinearRegression suffers from multicollinearity (when columns are correlated with each other), and variance explosions from outliers. Consider using Ridge Regression to fix the multicollinearity problem, and consider maybe first DBSCAN to remove the outliers, or statistical analysis to filter possible outliers.
Applications of LinearRegression
LinearRegression is used in regression tasks where one wants to predict say sales or house prices. It is also used in extrapolation or time series tasks, dynamic systems modelling and many other machine learning tasks. This model should be first tried if the machine learning problem is a regression task (predicting a continuous variable).
For additional information, see scikitlearn’s OLS documentation.
For an additional example see the OLS notebook.
Examples
import numpy as np import cudf # Both import methods supported from cuml import LinearRegression from cuml.linear_model import LinearRegression lr = LinearRegression(fit_intercept = True, normalize = False, algorithm = "eig") X = cudf.DataFrame() X['col1'] = np.array([1,1,2,2], dtype = np.float32) X['col2'] = np.array([1,2,2,3], dtype = np.float32) y = cudf.Series( np.array([6.0, 8.0, 9.0, 11.0], dtype = np.float32) ) reg = lr.fit(X,y) print("Coefficients:") print(reg.coef_) print("Intercept:") print(reg.intercept_) X_new = cudf.DataFrame() X_new['col1'] = np.array([3,2], dtype = np.float32) X_new['col2'] = np.array([5,5], dtype = np.float32) preds = lr.predict(X_new) print("Predictions:") print(preds)
Output:
Coefficients: 0 1.0000001 1 1.9999998 Intercept: 3.0 Predictions: 0 15.999999 1 14.999999
 Attributes
 coef_array, shape (n_features)
The estimated coefficients for the linear regression model.
 intercept_array
The independent term. If fit_intercept is False, will be 0.
Methods
fit
(self, X, y[, convert_dtype])Fit the model with X and y.
get_param_names
(self)predict
(self, X[, convert_dtype])Predicts y values for X.

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

decision_function
(self, X, convert_dtype=False)¶ Gives confidence score for X
 Parameters
 Xarraylike (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 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, convert_dtype=True)¶ Fit the model with X and y.
 Parameters
 Xarraylike (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.
 yarraylike (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 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)¶ Predicts the y for X.
 Parameters
 Xarraylike (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.

predict_log_proba
(self, X, convert_dtype=False)¶ Predicts the log class probabilities for each class in X
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix 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_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=False)¶ Predicts the class probabilities for each class in X
 Parameters
 Xarraylike (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_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.
Ridge Regression¶

class
cuml.
Ridge
(alpha=1.0, solver='eig', fit_intercept=True, normalize=False, handle=None, output_type=None)¶ Ridge extends LinearRegression by providing L2 regularization on the coefficients when predicting response y with a linear combination of the predictors in X. It can reduce the variance of the predictors, and improves the conditioning of the problem.
cuML’s Ridge can take arraylike objects, either in host as NumPy arrays or in device (as Numba or __cuda_array_interface__ compliant), in addition to cuDF objects. It provides 3 algorithms: SVD, Eig and CD to fit a linear model. In general SVD uses significantly more memory and is slower than Eig. If using CUDA 10.1, the memory difference is even bigger than in the other supported CUDA versions. However, SVD is more stable than Eig (default). CD uses Coordinate Descent and can be faster when data is large.
 Parameters
 alphafloat (default = 1.0)
Regularization strength  must be a positive float. Larger values specify stronger regularization. Array input will be supported later.
 solver{‘eig’, ‘svd’, ‘cd’} (default = ‘eig’)
Eig uses a eigendecomposition of the covariance matrix, and is much faster. SVD is slower, but guaranteed to be stable. CD or Coordinate Descent is very fast and is suitable for large problems.
 fit_interceptboolean (default = True)
If True, Ridge tries to correct for the global mean of y. If False, the model expects that you have centered the data.
 normalizeboolean (default = False)
If True, the predictors in X will be normalized by dividing by it’s L2 norm. If False, no scaling will be done.
Notes
Ridge provides L2 regularization. This means that the coefficients can shrink to become very small, but not zero. This can cause issues of interpretability on the coefficients. Consider using Lasso, or thresholding small coefficients to zero.
Applications of Ridge
Ridge Regression is used in the same way as LinearRegression, but does not suffer from multicollinearity issues. Ridge is used in insurance premium prediction, stock market analysis and much more.
For additional docs, see Scikitlearn’s Ridge Regression.
Examples
import numpy as np import cudf # Both import methods supported from cuml import Ridge from cuml.linear_model import Ridge alpha = np.array([1e5]) ridge = Ridge(alpha = alpha, fit_intercept = True, normalize = False, solver = "eig") X = cudf.DataFrame() X['col1'] = np.array([1,1,2,2], dtype = np.float32) X['col2'] = np.array([1,2,2,3], dtype = np.float32) y = cudf.Series( np.array([6.0, 8.0, 9.0, 11.0], dtype = np.float32) ) result_ridge = ridge.fit(X, y) print("Coefficients:") print(result_ridge.coef_) print("Intercept:") print(result_ridge.intercept_) X_new = cudf.DataFrame() X_new['col1'] = np.array([3,2], dtype = np.float32) X_new['col2'] = np.array([5,5], dtype = np.float32) preds = result_ridge.predict(X_new) print("Predictions:") print(preds)
Output:
Coefficients: 0 1.0000001 1 1.9999998 Intercept: 3.0 Preds: 0 15.999999 1 14.999999
 Attributes
 coef_array, shape (n_features)
The estimated coefficients for the linear regression model.
 intercept_array
The independent term. If fit_intercept is False, will be 0.
Methods
fit
(self, X, y[, convert_dtype])Fit the model with X and y.
get_param_names
(self)predict
(self, X[, convert_dtype])Predicts the y for X.

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

fit
(self, X, y, convert_dtype=True)¶ Fit the model with X and y.
 Parameters
 Xarraylike (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.
 yarraylike (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 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_params
(self, deep=True)¶ Scikitlearn style function that returns the estimator parameters.
 Parameters
 deepboolean (default = True)

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

fit
(self, X, y, convert_dtype=True)¶ Fit the model with X and y.
 Parameters
 Xarraylike (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.
 yarraylike (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 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_params
(self, deep=True)¶ Scikitlearn style function that returns the estimator parameters.
 Parameters
 deepboolean (default = True)

predict
(self, X, convert_dtype=False)¶ Predicts y values for X.
 Parameters
 Xarraylike (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)¶ Sklearn style set parameter state to dictionary of params.
 Parameters
 paramsdict of new 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 minibatch 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 = 1e3)
The training process will stop if current_loss > previous_loss  tol
 shuffleboolean (default = True)
True, shuffles the training data after each epoch False, does not shuffle the training data after each epoch
 eta0float (default = 0.001)
Initial learning rate
 power_tfloat (default = 0.5)
The exponent used for calculating the invscaling learning rate
 learning_rate{‘optimal’, ‘constant’, ‘invscaling’, ‘adaptive’}
(default = ‘constant’)
optimal option will be supported in a future version
constant keeps the learning rate constant
adaptive changes the learning rate if the training loss or the validation accuracy does not improve for n_iter_no_change epochs. The old learning rate is generally divided by 5
 n_iter_no_changeint (default = 5)
the number of epochs to train without any imporvement in the model
 output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’}, optional
Variable to control output type of the results and attributes of the estimators. If None, it’ll inherit the output type set at the module level, cuml.output_type. If set, the estimator will override the global option for its behavior.
Notes
For additional docs, see scikitlearn’s SGDClassifier.
Examples
import numpy as np import cudf from cuml.linear_model import MBSGDClassifier as cumlMBSGDClassifier X = cudf.DataFrame() X['col1'] = np.array([1,1,2,2], dtype = np.float32) X['col2'] = np.array([1,2,2,3], dtype = np.float32) y = cudf.Series(np.array([1, 1, 2, 2], dtype=np.float32)) pred_data = cudf.DataFrame() pred_data['col1'] = np.asarray([3, 2], dtype=np.float32) pred_data['col2'] = np.asarray([5, 5], dtype=np.float32) cu_mbsgd_classifier = cumlMBSGClassifier(learning_rate='constant', eta0=0.05, epochs=2000, fit_intercept=True, batch_size=1, tol=0.0, penalty='l2', loss='squared_loss', alpha=0.5) cu_mbsgd_classifier.fit(X, y) cu_pred = cu_mbsgd_classifier.predict(pred_data).to_array() print(" cuML intercept : ", cu_mbsgd_classifier.intercept_) print(" cuML coef : ", cu_mbsgd_classifier.coef_) print("cuML predictions : ", cu_pred)
Output:
cuML intercept : 0.7150013446807861 cuML coef : 0 0.27320495 1 0.1875956 dtype: float32 cuML predictions : [1. 1.]
Methods
fit
(self, X, y[, convert_dtype])Fit the model with X and y.
get_params
(self[, deep])Scikitlearn style function that returns the estimator parameters.
predict
(self, X[, convert_dtype])Predicts the y for X.
set_params
(self, **params)Sklearn style set parameter state to dictionary of params.

fit
(self, X, y, convert_dtype=True)¶ Fit the model with X and y.
 Parameters
 Xarraylike (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.
 yarraylike (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 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_params
(self, deep=True)¶ Scikitlearn style function that returns the estimator parameters.
 Parameters
 deepboolean (default = True)

predict
(self, X, convert_dtype=False)¶ Predicts the y for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix 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)¶ Sklearn style set parameter state to dictionary of params.
 Parameters
 paramsdict of new 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 minibatch 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 = 1e3)
The training process will stop if current_loss > previous_loss  tol
 shuffleboolean (default = True)
True, shuffles the training data after each epoch False, does not shuffle the training data after each epoch
 eta0float (default = 0.001)
Initial learning rate
 power_tfloat (default = 0.5)
The exponent used for calculating the invscaling learning rate
 learning_rate{‘optimal’, ‘constant’, ‘invscaling’, ‘adaptive’}
(default = ‘constant’)
optimal option will be supported in a future version
constant keeps the learning rate constant
adaptive changes the learning rate if the training loss or the validation accuracy does not improve for n_iter_no_change epochs. The old learning rate is generally divided by 5
 n_iter_no_changeint (default = 5)
the number of epochs to train without any imporvement in the model
 output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’}, optional
Variable to control output type of the results and attributes of the estimators. If None, it’ll inherit the output type set at the module level, cuml.output_type. If set, the estimator will override the global option for its behavior.
Notes
For additional docs, see scikitlearn’s SGDRegressor.
Examples
import numpy as np import cudf from cuml.linear_model import MBSGDRegressor as cumlMBSGDRegressor X = cudf.DataFrame() X['col1'] = np.array([1,1,2,2], dtype = np.float32) X['col2'] = np.array([1,2,2,3], dtype = np.float32) y = cudf.Series(np.array([1, 1, 2, 2], dtype=np.float32)) pred_data = cudf.DataFrame() pred_data['col1'] = np.asarray([3, 2], dtype=np.float32) pred_data['col2'] = np.asarray([5, 5], dtype=np.float32) cu_mbsgd_regressor = cumlMBSGDRegressor(learning_rate='constant', eta0=0.05, epochs=2000, fit_intercept=True, batch_size=1, tol=0.0, penalty='l2', loss='squared_loss', alpha=0.5) cu_mbsgd_regressor.fit(X, y) cu_pred = cu_mbsgd_regressor.predict(pred_data).to_array() print(" cuML intercept : ", cu_mbsgd_regressor.intercept_) print(" cuML coef : ", cu_mbsgd_regressor.coef_) print("cuML predictions : ", cu_pred)
Output:
cuML intercept : 0.7150013446807861 cuML coef : 0 0.27320495 1 0.1875956 dtype: float32 cuML predictions : [2.4725943 2.1993892]
Methods
fit
(self, X, y[, convert_dtype])Fit the model with X and y.
get_params
(self[, deep])Scikitlearn style function that returns the estimator parameters.
predict
(self, X[, convert_dtype])Predicts the y for X.
set_params
(self, **params)Sklearn style set parameter state to dictionary of params.

fit
(self, X, y, convert_dtype=True)¶ Fit the model with X and y.
 Parameters
 Xarraylike (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.
 yarraylike (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 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_params
(self, deep=True)¶ Scikitlearn style function that returns the estimator parameters.
 Parameters
 deepboolean (default = True)

predict
(self, X, convert_dtype=False)¶ Predicts the y for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix 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)¶ Sklearn style set parameter state to dictionary of params.
 Parameters
 paramsdict of new params
Mutinomial Naive Bayes¶

class
cuml.
MultinomialNB
(**kwargs)¶ 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 tfidf may also work.
NOTE: While cuML only provides the multinomial version currently, the other variants are planned to be included soon. Refer to the corresponding Github issue for updates: https://github.com/rapidsai/cuml/issues/1666
Examples
Load the 20 newsgroups dataset from Scikitlearn 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
fit
(X, y[, sample_weight])Fit Naive Bayes classifier according to X, y
partial_fit
(X, y[, classes, sample_weight])Incremental fit on a batch of samples.
predict
(X)Perform classification on an array of test vectors X.
Return logprobability estimates for the test vector X.
Return probability estimates for the test vector X.
score
(X, y[, sample_weight])Return the mean accuracy on the given test data and labels.
Updates the log probabilities.

fit
(X, y, sample_weight=None)¶ Fit Naive Bayes classifier according to X, y
 Parameters
 Xarraylike (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.
 yarraylike (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.
 sample_weightarraylike (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.

partial_fit
(X, y, classes=None, sample_weight=None)¶ Incremental fit on a batch of samples.
This method is expected to be called several times consecutively on different chunks of a dataset so as to implement outofcore or online learning.
This is especially useful when the whole dataset is too big to fit in memory at once.
This method has some performance overhead hence it is better to call partial_fit on chunks of data that are as large as possible (as long as fitting in the memory budget) to hide the overhead.
 Parameters
 X{arraylike, cupy sparse matrix} of shape (n_samples, n_features)
Training vectors, where n_samples is the number of samples and n_features is the number of features
 yarraylike of shape (n_samples) Target values.
 classesarraylike of shape (n_classes)
List of all the classes that can possibly appear in the y vector. Must be provided at the first call to partial_fit, can be omitted in subsequent calls.
 sample_weightarraylike of shape (n_samples)
Weights applied to individual samples (1. for unweighted). Currently sample weight is ignored
 Returns
 selfobject

predict
(X)¶ Perform classification on an array of test vectors X.
 Parameters
 Xarraylike (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.
 Returns
 y_hatcuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_rows, 1)
Predicted values
For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

predict_log_proba
(X)¶ Return logprobability estimates for the test vector X.
 Parameters
 Xarraylike (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.
 Returns
 CcuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_rows, 1)
Returns the logprobability of the samples for each class in the model. The columns correspond to the classes in sorted order, as they appear in the attribute classes_.
For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

predict_proba
(X)¶ Return probability estimates for the test vector X.
 Parameters
 Xarraylike (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.
 Returns
 CcuDF, CuPy or NumPy object depending on cuML’s output type configuration, shape = (n_rows, 1)
Returns the probability of the samples for each class in the model. The columns correspond to the classes in sorted order, as they appear in the attribute classes_.
For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.

score
(X, y, sample_weight=None)¶ Return the mean accuracy on the given test data and labels.
In multilabel classification, this is the subset accuracy which is a harsh metric since you require for each sample that each label set be correctly predicted.
Currently, sample weight is ignored
 Parameters
 Xarraylike (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.
 yarraylike (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.
 sample_weightarraylike (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
 scorefloat
Mean accuracy of self.predict(X) with respect to y.

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

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

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

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

fit
(self, X, y, convert_dtype=True)¶ Perform Random Forest Classification on the input data
 Parameters
 Xarraylike (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.
 yarraylike (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.
 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_params
(self, deep=True)¶ Returns the value of all parameters required to configure this estimator as a dictionary. Parameters ———– deep : boolean (default = True)

predict
(self, X, predict_model='GPU', output_class=True, threshold=0.5, algo='auto', num_classes=None, convert_dtype=True, fil_sparse_format='auto')¶ Predicts the labels for X.
 Parameters
 Xarraylike (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 binary classification problems.
 output_classboolean (default = True)
This is optional and required only while performing the predict operation on the GPU. If true, return a 1 or 0 depending on whether the raw prediction exceeds the threshold. If False, just return the raw prediction.
 algostring (default = ‘auto’)
This is optional and required only while performing the predict operation on the GPU. ‘naive’  simple inference using shared memory ‘tree_reorg’  similar to naive but trees rearranged to be more coalescingfriendly ‘batch_tree_reorg’  similar to tree_reorg but predicting multiple rows per thread block auto  choose the algorithm automatically. Currently ‘batch_tree_reorg’ is used for dense storage and ‘naive’ for sparse storage
 thresholdfloat (default = 0.5)
Threshold used for classification. Optional and required only while performing the predict operation on the GPU. It is applied if output_class == True, else it is ignored
 num_classesint (default = None)
number of different classes present in the dataset. This variable will be deprecated in 0.16. 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, output_class=True, threshold=0.5, algo='auto', num_classes=None, convert_dtype=True, fil_sparse_format='auto')¶ Predicts class probabilites for X. This function uses the GPU implementation of predict. Therefore, data with ‘dtype = np.float32’ should be used with this function.
 Parameters
 Xarraylike (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.
 output_class: boolean (default = True)
This is optional and required only while performing the predict operation on the GPU. If true, return a 1 or 0 depending on whether the raw prediction exceeds the threshold. If False, just return the raw prediction.
 algostring (default = ‘auto’)
This is optional and required only while performing the predict operation on the GPU. ‘naive’  simple inference using shared memory ‘tree_reorg’  similar to naive but trees rearranged to be more coalescingfriendly ‘batch_tree_reorg’  similar to tree_reorg but predicting multiple rows per thread block auto  choose the algorithm automatically. Currently ‘batch_tree_reorg’ is used for dense storage and ‘naive’ for sparse storage
 thresholdfloat (default = 0.5)
Threshold used for classification. Optional and required only while performing the predict operation on the GPU. It is applied if output_class == True, else it is ignored
 num_classesint (default = None)
number of different classes present in the dataset. This variable will be deprecated in 0.16. 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)

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

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

score
(self, X, y, threshold=0.5, algo='auto', num_classes=None, predict_model='GPU', convert_dtype=True, fil_sparse_format='auto')¶ Calculates the accuracy metric score of the model for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
 Dense matrix containing floats or doubles.
 Acceptable formats: CUDA array interface compliant objects like
 CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas
 DataFrame/Series.
y : arraylike (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 coalescingfriendly ‘batch_tree_reorg’  similar to tree_reorg but predicting multiple rows per thread block auto  choose the algorithm automatically. Currently ‘batch_tree_reorg’ is used for dense storage and ‘naive’ for sparse storage
 thresholdfloat
threshold is used to for classification This is optional and required only while performing the predict operation on the GPU.
 num_classesint (default = None)
number of different classes present in the dataset. This variable will be deprecated in 0.16. 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 binary classification problems.
 fil_sparse_formatboolean or string (default = auto)
This variable is used to choose the type of forest that will be created in the Forest Inference Library. It is not required while using predict_model=’CPU’. ‘auto’  choose the storage type automatically (currently True is chosen by auto) False  create a dense forest True  create a sparse forest, requires algo=’naive’ or algo=’auto’
 Returns
 accuracyfloat
Accuracy of the model [0.0  1.0]

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

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

convert_to_fil_model
(self, output_class=False, algo='auto', fil_sparse_format='auto')¶ Create a Forest Inference (FIL) model from the trained cuML Random Forest model.
 Parameters
 output_classboolean (default = 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 coalescingfriendly ‘batch_tree_reorg’  similar to tree_reorg but predicting multiple rows per thread block auto  choose the algorithm automatically. Currently ‘batch_tree_reorg’ is used for dense storage and ‘naive’ for sparse storage
 fil_sparse_formatboolean or string (default = ‘auto’)
This variable is used to choose the type of forest that will be created in the Forest Inference Library. It is not required while using predict_model=’CPU’. ‘auto’  choose the storage type automatically (currently True is chosen by auto) False  create a dense forest True  create a sparse forest, requires algo=’naive’ or algo=’auto’
 Returns
 fil_model
A Forest Inference model which can be used to perform inferencing on the random forest model.

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

fit
(self, X, y, convert_dtype=True)¶ Perform Random Forest Regression on the input data
 Parameters
 Xarraylike (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.
 yarraylike (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 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_params
(self, deep=True)¶ Returns the value of all parameters required to configure this estimator as a dictionary. Parameters ———– deep : boolean (default = True)

predict
(self, X, predict_model='GPU', algo='auto', convert_dtype=True, fil_sparse_format='auto')¶ Predicts the labels for X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
 Dense matrix 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 coalescingfriendly ‘batch_tree_reorg’  similar to tree_reorg but predicting multiple rows per thread block auto  choose the algorithm automatically. Currently ‘batch_tree_reorg’ is used for dense storage and ‘naive’ for sparse storage
 convert_dtypebool, optional (default = True)
When set to True, the predict method will, when necessary, convert the input to the data type which was used to train the model. This will increase memory used for the method.
 fil_sparse_formatboolean or string (default = auto)
This variable is used to choose the type of forest that will be created in the Forest Inference Library. It is not required while using predict_model=’CPU’. ‘auto’  choose the storage type automatically (currently True is chosen by auto) False  create a dense forest True  create a sparse forest, requires algo=’naive’ or algo=’auto’
 Returns
 ycuDF, CuPy or NumPy object depending on cuML’s output typeconfiguration, shape =(n_samples, 1)

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

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

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

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

class
cuml.
ForestInference
(handle=None, output_type=None)¶ ForestInference provides GPUaccelerated inference (prediction) for random forest and boosted decision tree models.
This module does not support training models. Rather, users should train a model in another package and save it in a treelitecompatible format. (See https://github.com/dmlc/treelite) Currently, LightGBM, XGBoost and SKLearn GBDT and random forest models are supported.
Users typically create a ForestInference object by loading a saved model file with ForestInference.load. It is also possible to create it from an SKLearn model using ForestInference.load_from_sklearn. The resulting object provides a predict method for carrying out inference.
 Known limitations:
A single row of data should fit into the shared memory of a thread block, which means that more than 12288 features are not supported.
From sklearn.ensemble, only {RandomForest,GradientBoosting}{Classifier,Regressor} models are supported; other sklearn.ensemble models are currently not supported.
Importing large SKLearn models can be slow, as it is done in Python.
LightGBM categorical features are not supported.
Inference uses a dense matrix format, which is efficient for many problems but can be suboptimal for sparse datasets.
Only binary classification and regression are supported.
 Parameters
 handlecuml.Handle
If it is None, a new one is created just for this class.
Notes
For additional usage examples, see the sample notebook at https://github.com/rapidsai/cuml/blob/branch0.15/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
load
(filename[, output_class, threshold, …])Returns a FIL instance containing the forest saved in filename This uses Treelite to load the saved model.
load_from_sklearn
(skl_model[, output_class, …])Creates a FIL model using the scikitlearn model passed to the function.
load_from_treelite_model
(self, model[, …])Creates a FIL model using the treelite model passed to the function.
load_using_treelite_handle
(self, model_handle)Returns a FIL instance by converting a treelite model to FIL model by using the treelite ModelHandle passed.
predict
(self, X[, preds])Predicts the labels for X with the loaded forest model.
predict_proba
(self, X[, preds])Predicts the class probabilities for X with the loaded forest model.

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

static
load_from_sklearn
(skl_model, output_class=False, threshold=0.5, algo='auto', storage_type='auto', handle=None)¶ Creates a FIL model using the scikitlearn model passed to the function. This function requires Treelite 0.90 to be installed.
 Parameters
 skl_model
The scikitlearn 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 coalescingfriendly
‘BATCH_TREE_REORG’ or ‘batch_tree_reorg’  similar to TREE_REORG but predicting multiple rows per thread block
 thresholdfloat (default=0.5)
Threshold is used to for classification. It is applied only if
output_class == True
, else it is ignored. storage_typestring or boolean (default=’auto’)
 Inmemory 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’
 Returns
 fil_model
A Forest Inference model created from the scikitlearn model passed.

load_from_treelite_model
(self, model, output_class=False, algo='auto', threshold=0.5, storage_type='auto')¶ Creates a FIL model using the treelite model passed to the function.
 Parameters
 model
the trained model information in the treelite format loaded from a saved model using the treelite API https://treelite.readthedocs.io/en/latest/treeliteapi.html
 output_class: boolean (default=False)
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 coalescingfriendly
‘BATCH_TREE_REORG’ or ‘batch_tree_reorg’  similar to TREE_REORG but predicting multiple rows per thread block
 thresholdfloat (default=0.5)
Threshold is used to for classification. It is applied only if output_class == True, else it is ignored.
 storage_typestring or boolean (default=’auto’)
 Inmemory 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’
 ‘sparse8’  (experimental) create a sparse forest
with 8byte nodes; requires algo=’NAIVE’ or algo=’AUTO’; can fail if 8byte nodes are not enough to store the forest, e.g. if there are too many nodes in a tree or too many features
 Returns
 fil_model
A Forest Inference model which can be used to perform inferencing on the random forest/ XGBoost model.

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

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

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

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

fit
(self, X, y, convert_dtype=False)¶ Fit the model with X and y.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.
 yarraylike (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 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=False)¶ Predicts the y for X.
 Parameters
 Xarraylike (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.
QuasiNewton¶

class
cuml.
QN
(loss='sigmoid', fit_intercept=True, l1_strength=0.0, l2_strength=0.0, max_iter=1000, tol=0.001, linesearch_max_iter=50, lbfgs_memory=5, verbose=False, handle=None, output_type=None)¶ QuasiNewton methods are used to either find zeroes or local maxima and minima of functions, and used by this class to optimize a cost function.
Two algorithms are implemented underneath cuML’s QN class, and which one is executed depends on the following rule:
OrthantWise Limited Memory QuasiNewton (OWLQN) if there is l1 regularization
Limited Memory BFGS (LBFGS) otherwise.
cuML’s QN class can take arraylike objects, either in host as NumPy arrays or in device (as Numba or __cuda_array_interface__ compliant).
 Parameters
 loss: ‘sigmoid’, ‘softmax’, ‘squared_loss’ (default = ‘squared_loss’)
‘sigmoid’ loss used for single class logistic regression ‘softmax’ loss used for multiclass logistic regression ‘normal’ used for normal/square loss
 fit_intercept: boolean (default = True)
If True, the model tries to correct for the global mean of y. If False, the model expects that you have centered the data.
 l1_strength: float (default = 0.0)
l1 regularization strength (if nonzero, will run OWLQN, else LBFGS). Note, that as in Scikitlearn, the bias will not be regularized.
 l2_strength: float (default = 0.0)
l2 regularization strength. Note, that as in Scikitlearn, the bias will not be regularized.
 max_iter: int (default = 1000)
Maximum number of iterations taken for the solvers to converge.
 tol: float (default = 1e3)
The training process will stop if current_loss > previous_loss  tol
 linesearch_max_iter: int (default = 50)
Max number of linesearch iterations per outer iteration of the algorithm.
 lbfgs_memory: int (default = 5)
Rank of the lbfgs inverseHessian approximation. Method will use O(lbfgs_memory * D) memory.
 verboseint or boolean (default = False)
Controls verbose level of logging.
Notes
This class contains implementations of two popular QuasiNewton methods:
Limitedmemory Broyden Fletcher Goldfarb Shanno (LBFGS) [Nocedal, Wright  Numerical Optimization (1999)]
Orthantwise limitedmemory quasinewton (OWLQN) [Andrew, Gao  ICML 2007] <https://www.microsoft.com/enus/research/publication/scalabletrainingofl1regularizedloglinearmodels/>
Examples
import cudf import numpy as np # Both import methods supported # from cuml import QN from cuml.solvers import QN X = cudf.DataFrame() X['col1'] = np.array([1,1,2,2], dtype = np.float32) X['col2'] = np.array([1,2,2,3], dtype = np.float32) y = cudf.Series( np.array([0.0, 0.0, 1.0, 1.0], dtype = np.float32) ) solver = QN() solver.fit(X,y) # Note: for now, the coefficients also include the intercept in the # last position if fit_intercept=True print("Coefficients:") print(solver.coef_) 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)
The estimated coefficients for the linear regression model. Note: shape is (n_classes, n_features + 1) if fit_intercept = True.
 intercept_array (n_classes, 1)
The independent term. If fit_intercept is False, will be 0.
Methods
fit
(self, X, y[, convert_dtype])Fit the model with X and y.
get_param_names
(self)predict
(self, X[, convert_dtype])Predicts the y for X.
score
(self, X, y)
fit
(self, X, y, convert_dtype=False)¶ Fit the model with X and y.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.
 yarraylike (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 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)¶ Predicts the y for X.
 Parameters
 Xarraylike (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.

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

class
cuml.svm.
SVC
(CSupport Vector Classification)¶ Construct an SVC classifier for training and predictions.
Note
This implementation has the following known limitations:
Currently only binary classification is supported.
 Parameters
 handlecuml.Handle
If it is None, a new one is created for this class
 Cfloat (default = 1.0)
Penalty parameter C
 kernelstring (default=’rbf’)
Specifies the kernel function. Possible options: ‘linear’, ‘poly’, ‘rbf’, ‘sigmoid’. Currently precomputed kernels are not supported.
 degreeint (default=3)
Degree of polynomial kernel function.
 gammafloat or string (default = ‘scale’)
Coefficient for rbf, poly, and sigmoid kernels. You can specify the numeric value, or use one of the following options:  ‘auto’: gamma will be set to 1 / n_features  ‘scale’: gamma will be se to 1 / (n_features * X.var())
 coef0float (default = 0.0)
Independent term in kernel function, only signifficant for poly and sigmoid
 tolfloat (default = 1e3)
Tolerance for stopping criterion.
 cache_sizefloat (default = 200.0)
Size of the kernel cache during training in MiB. The default is a conservative value, increase it to improve the training time, at the cost of higher memory footprint. After training the kernel cache is deallocated. During prediction, we also need a temporary space to store kernel matrix elements (this can be signifficant if n_support is large). The cache_size variable sets an upper limit to the prediction buffer as well.
 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
 nochange_stepsint (default = 1000)
We monitor how much our stopping criteria changes during outer iterations. If it does not change (changes less then 1e3*tol) for nochange_steps consecutive steps, then we stop training.
 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)
verbosity level
Notes
The solver uses the SMO method to fit the classifier. We use the Optimized Hierarchical Decomposition [1] variant of the SMO algorithm, similar to [2].
For additional docs, see scikitlearn’s SVC.
References
 1
J. Vanek et al. A GPUArchitecture Optimized Hierarchical Decomposition Algorithm for Support VectorMachine Training, IEEE Transactions on Parallel and Distributed Systems, vol 28, no 12, 3330, (2017)
 2
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_int
The constant in the decision function
 fit_status_int
0 if SVM is correctly fitted
coef_
float, shape (1, n_cols)SVMBase.coef_(self)
 classes_: shape (n_classes_,)
Array of class labels.
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)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)¶ Calculates the decision function values for X.
 Parameters
 Xarraylike (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)¶ Fit the model with X and y.
 Parameters
 Xarraylike (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.
 yarraylike (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_weightarraylike (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)¶ Predicts the class labels for X. The returned y values are the class labels associated to sign(decision_function(X)).
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix 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.

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

fit
(self, X, y, sample_weight=None, convert_dtype=True)¶ Fit the model with X and y.
 Parameters
 Xarraylike (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.
 yarraylike (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_weightarraylike (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)¶ Predicts the values for X.
 Parameters
 Xarraylike (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.
Nearest Neighbors Classification¶

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

fit
(self, X, y, convert_dtype=True) Fit a GPU index for knearest neighbors classifier model.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.
 yarraylike (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.

get_param_names
(self)

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

fit
(self, X, y, convert_dtype=True) Fit a GPU index for knearest neighbors regression model.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.
 yarraylike (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.

get_param_names
(self)

predict
(self, X, convert_dtype=True) Use the trained knearest neighbors regression model to predict the labels for X
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix 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¶
KMeans Clustering¶

class
cuml.
KMeans
(handle=None, n_clusters=8, max_iter=300, tol=0.0001, verbose=False, random_state=1, init='scalablekmeans++', 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 arraylike object or cuDF DataFrame, and supports the scalable KMeans++ initialization method. This method is more stable than randomly selecting K points.
 Parameters
 handlecuml.Handle
If it is None, a new one is created just for this class.
 n_clustersint (default = 8)
The number of centroids or clusters you want.
 max_iterint (default = 300)
The more iterations of EM, the more accurate, but slower.
 tolfloat64 (default = 1e4)
Stopping criterion when centroid means do not change much.
 verboseint or boolean (default = False)
Logging level.
 random_stateint (default = 1)
If you want results to be the same when you restart Python, select a state.
 init‘scalablekmeans++’, ‘kmeans’ , ‘random’ or an ndarray (default = ‘scalablekmeans++’) # noqa
‘scalablekmeans++’ or ‘kmeans’: Uses fast and stable scalable kmeans++ initialization. ‘random’: Choose ‘n_cluster’ observations (rows) at random from data for the initial centroids. If an ndarray is passed, it should be of shape (n_clusters, n_features) and gives the initial centers.
 n_init: int (default = 1)
Number of instances the kmeans algorithm will be called with different seeds. The final results will be from the instance that produces lowest inertia out of n_init instances.
 oversampling_factorfloat64
scalable kmeans oversampling factor
 max_samples_per_batchint (default=1<<15)
maximum number of samples to use for each batch of the pairwise distance computation.
 oversampling_factorint (default = 2)
The amount of points to sample in scalable kmeans++ initialization for potential centroids. Increasing this value can lead to better initial centroids at the cost of memory. The total number of centroids sampled in scalable kmeans++ is oversampling_factor * n_clusters * 8.
 max_samples_per_batchint (default = 32768)
The number of data samples to use for batches of the pairwise distance computation. This computation is done throughout both fit predict. The default should suit most cases. The total number of elements in the batched pairwise distance computation is max_samples_per_batch * n_clusters. It might become necessary to lower this number when n_clusters becomes prohibitively large.
Notes
KMeans requires n_clusters to be specified. This means one needs to approximately guess or know how many clusters a dataset has. If one is not sure, one can start with a small number of clusters, and visualize the resulting clusters with PCA, UMAP or TSNE, and verify that they look appropriate.
Applications of KMeans
The biggest advantage of KMeans is its speed and simplicity. That is why KMeans is many practitioner’s first choice of a clustering algorithm. KMeans has been extensively used when the number of clusters is approximately known, such as in big data clustering tasks, image segmentation and medical clustering.
For additional docs, see scikitlearn’s Kmeans.
Examples
# Both import methods supported from cuml import KMeans from cuml.cluster import KMeans import cudf import numpy as np import pandas as pd def np2cudf(df): # convert numpy array to cuDF dataframe df = pd.DataFrame({'fea%d'%i:df[:,i] for i in range(df.shape[1])}) pdf = cudf.DataFrame() for c,column in enumerate(df): pdf[str(c)] = df[column] return pdf a = np.asarray([[1.0, 1.0], [1.0, 2.0], [3.0, 2.0], [4.0, 3.0]], dtype=np.float32) b = np2cudf(a) print("input:") print(b) print("Calling fit") kmeans_float = KMeans(n_clusters=2) kmeans_float.fit(b) print("labels:") print(kmeans_float.labels_) print("cluster_centers:") print(kmeans_float.cluster_centers_)
Output:
input: 0 1 0 1.0 1.0 1 1.0 2.0 2 3.0 2.0 3 4.0 3.0 Calling fit labels: 0 0 1 0 2 1 3 1 cluster_centers: 0 1 0 1.0 1.5 1 3.5 2.5
 Attributes
 cluster_centers_array
The coordinates of the final clusters. This represents of “mean” of each data cluster.
 labels_array
Which cluster each datapoint belongs to.
Methods
fit
(self, X[, sample_weight])Compute kmeans clustering with X.
fit_predict
(self, X[, sample_weight])Compute cluster centers and predict cluster index for each sample.
fit_transform
(self, X[, convert_dtype])Compute clustering and transform X to clusterdistance space.
get_param_names
(self)predict
(self, X[, convert_dtype, sample_weight])Predict the closest cluster each sample in X belongs to.
score
(self, X[, y, sample_weight, convert_dtype])Opposite of the value of X on the Kmeans objective.
transform
(self, X[, convert_dtype])Transform X to a clusterdistance space.

fit
(self, X, sample_weight=None)¶ Compute kmeans clustering with X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.
 sample_weightarraylike (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)¶ Compute cluster centers and predict cluster index for each sample.
 Parameters
 Xarraylike (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_weightarraylike (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)¶ Compute clustering and transform X to clusterdistance space.
 Parameters
 Xarraylike (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.
 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)¶ Predict the closest cluster each sample in X belongs to.
 Parameters
 Xarraylike (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_weightarraylike (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 Kmeans objective.
 Parameters
 Xarraylike (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.
 yarraylike (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_weightarraylike (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 Kmeans objective.

transform
(self, X, convert_dtype=False)¶ Transform X to a clusterdistance space.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix 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, 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 arraylike object or cuDF DataFrame, and constructs an adjacency graph to compute the distances between close neighbours.
 Parameters
 epsfloat (default = 0.5)
The maximum distance between 2 points such they reside in the same neighborhood.
 handlecuml.Handle
If it is None, a new one is created just for this class
 min_samplesint (default = 5)
The number of samples in a neighborhood such that this group can be considered as an important core point (including the point itself).
 verboseint or boolean (default = False)
Logging level
 max_mbytes_per_batch(optional) int64
Calculate batch size using no more than this number of megabytes for the pairwise distance computation. This enables the tradeoff between runtime and memory usage for making the N^2 pairwise distance computations more tractable for large numbers of samples. If you are experiencing out of memory errors when running DBSCAN, you can set this value based on the memory size of your device. Note: this option does not set the maximum total memory used in the DBSCAN computation and so this value will not be able to be set to the total memory available on the device.
 output_type(optional) {‘input’, ‘cudf’, ‘cupy’, ‘numpy’} default = None
Use it to control output type of the results and attributes. If None it’ll inherit the output type set at the module level, cuml.output_type. If that has not been changed, by default the estimator will mirror the type of the data used for each fit or predict call. If set, the estimator will override the global option for its behavior.
 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 nonlinearly shaped clusters. This also allows DBSCAN to be robust to noise. DBSCAN has been applied to analyzing particle collisions in the Large Hadron Collider, customer segmentation in marketing analyses, and much more.
For additional docs, see scikitlearn’s DBSCAN.
Examples
# Both import methods supported from cuml import DBSCAN from cuml.cluster import DBSCAN import cudf import numpy as np gdf_float = cudf.DataFrame() gdf_float['0'] = np.asarray([1.0,2.0,5.0], dtype = np.float32) gdf_float['1'] = np.asarray([4.0,2.0,1.0], dtype = np.float32) gdf_float['2'] = np.asarray([4.0,2.0,1.0], dtype = np.float32) dbscan_float = DBSCAN(eps = 1.0, min_samples = 1) dbscan_float.fit(gdf_float) print(dbscan_float.labels_)
Output:
0 0 1 1 2 2
 Attributes
 labels_arraylike or cuDF series
Which cluster each datapoint belongs to. Noisy samples are labeled as 1. Format depends on cuml global output type and estimator output_type.
 core_sample_indices_arraylike 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')¶ Perform DBSCAN clustering from features.
 Parameters
 Xarraylike (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')¶ Performs clustering on X and returns cluster labels.
 Parameters
 Xarraylike (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)¶
Dimensionality Reduction and Manifold Learning¶
Principal Component Analysis¶

class
cuml.
PCA
(copy=True, handle=None, iterated_power=15, n_components=1, random_state=None, svd_solver='auto', tol=1e07, 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 arraylike object or cuDF DataFrame, and provides 2 algorithms Full and Jacobi. Full (default) uses a full eigendecomposition then selects the top K eigenvectors. The Jacobi algorithm is much faster as it iteratively tries to correct the top K eigenvectors, but might be less accurate.
 Parameters
 copyboolean (default = True)
If True, then copies data then removes mean from data. False might cause data to be overwritten with its mean centered version.
 handlecuml.Handle
If it is None, a new one is created just for this class
 iterated_powerint (default = 15)
Used in Jacobi solver. The more iterations, the more accurate, but slower.
 n_componentsint (default = 1)
The number of top K singular vectors / values you want. Must be <= number(columns).
 random_stateint / None (default = None)
If you want results to be the same when you restart Python, select a state.
 svd_solver‘full’ or ‘jacobi’ or ‘auto’ (default = ‘full’)
Full uses a eigendecomposition of the covariance matrix then discards components. Jacobi is much faster as it iteratively corrects, but is less accurate.
 tolfloat (default = 1e7)
Used if algorithm = “jacobi”. Smaller tolerance can increase accuracy, but but will slow down the algorithm’s convergence.
 verboseint or boolean (default = False)
Logging level
 whitenboolean (default = False)
If True, decorrelates the components. This is done by dividing them by the corresponding singular values then multiplying by sqrt(n_samples). Whitening allows each component to have unit variance and removes multicollinearity. It might be beneficial for downstream tasks like LinearRegression where correlated features cause problems.
Notes
PCA considers linear combinations of features, specifically those that maximize global variance structure. This means PCA is fantastic for global structure analyses, but weak for local relationships. Consider UMAP or TSNE for a locally important embedding.
Applications of PCA
PCA is used extensively in practice for data visualization and data compression. It has been used to visualize extremely large word embeddings like Word2Vec and GloVe in 2 or 3 dimensions, large datasets of everyday objects and images, and used to distinguish between cancerous cells from healthy cells.
For additional docs, see scikitlearn’s PCA.
Examples
# Both import methods supported from cuml import PCA from cuml.decomposition import PCA import cudf import numpy as np gdf_float = cudf.DataFrame() gdf_float['0'] = np.asarray([1.0,2.0,5.0], dtype = np.float32) gdf_float['1'] = np.asarray([4.0,2.0,1.0], dtype = np.float32) gdf_float['2'] = np.asarray([4.0,2.0,1.0], dtype = np.float32) pca_float = PCA(n_components = 2) pca_float.fit(gdf_float) print(f'components: {pca_float.components_}') print(f'explained variance: {pca_float._explained_variance_}') exp_var = pca_float._explained_variance_ratio_ print(f'explained variance ratio: {exp_var}') print(f'singular values: {pca_float._singular_values_}') print(f'mean: {pca_float._mean_}') print(f'noise variance: {pca_float._noise_variance_}') trans_gdf_float = pca_float.transform(gdf_float) print(f'Inverse: {trans_gdf_float}') input_gdf_float = pca_float.inverse_transform(trans_gdf_float) print(f'Input: {input_gdf_float}')
Output:
components: 0 1 2 0 0.69225764 0.5102837 0.51028395 1 0.72165036 0.48949987 0.4895003 explained variance: 0 8.510402 1 0.48959687 explained variance ratio: 0 0.9456003 1 0.054399658 singular values: 0 4.1256275 1 0.9895422 mean: 0 2.6666667 1 2.3333333 2 2.3333333 noise variance: 0 0.0 transformed matrix: 0 1 0 2.8547091 0.42891636 1 0.121316016 0.80743366 2 2.9760244 0.37851727 Input Matrix: 0 1 2 0 1.0000001 3.9999993 4.0 1 2.0 2.0000002 1.9999999 2 4.9999995 1.0000006 1.0
 Attributes
 components_array
The top K components (VT.T[:,:n_components]) in U, S, VT = svd(X)
 explained_variance_array
How much each component explains the variance in the data given by S**2
 explained_variance_ratio_array
How much in % the variance is explained given by S**2/sum(S**2)
 singular_values_array
The top K singular values. Remember all singular values >= 0
 mean_array
The column wise mean of X. Used to mean  center the data first.
 noise_variance_float
From Bishop 1999’s Textbook. Used in later tasks like calculating the estimated covariance of X.
Methods
fit
(self, X[, y])Fit the model with X. 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)¶ Fit the model with X. y is currently ignored.
 Parameters
 Xarraylike (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.
 yarraylike (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.

fit_transform
(self, X, y=None)¶ Fit the model with X and apply the dimensionality reduction on X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
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.
 yarraylike (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.
 Returns
 transcuDF, CuPy or NumPy object depending on cuML’s output type configuration, cupy.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=1e10)¶ Transform data back to its original space.
In other words, return an input X_original whose transform would be X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
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 cupy.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 = 1e10)
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, cupy.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)¶ Apply dimensionality reduction to X.
X is projected on the first principal components previously extracted from a training set.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
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, cupy.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.
Truncated SVD¶

class
cuml.
TruncatedSVD
(algorithm='full', handle=None, n_components=1, n_iter=15, random_state=None, tol=1e07, 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 arraylike object or cuDF DataFrame, and provides 2 algorithms Full and Jacobi. Full (default) uses a full eigendecomposition then selects the top K singular vectors. The Jacobi algorithm is much faster as it iteratively tries to correct the top K singular vectors, but might be less accurate.
 Parameters
 algorithm‘full’ or ‘jacobi’ or ‘auto’ (default = ‘full’)
Full uses a eigendecomposition of the covariance matrix then discards components. Jacobi is much faster as it iteratively corrects, but is less accurate.
 handlecuml.Handle
If it is None, a new one is created just for this class
 n_componentsint (default = 1)
The number of top K singular vectors / values you want. Must be <= number(columns).
 n_iterint (default = 15)
Used in Jacobi solver. The more iterations, the more accurate, but slower.
 random_stateint / None (default = None)
If you want results to be the same when you restart Python, select a state.
 tolfloat (default = 1e7)
Used if algorithm = “jacobi”. Smaller tolerance can increase accuracy, but but will slow down the algorithm’s convergence.
 verboseint or boolean (default = False)
Logging level
Notes
TruncatedSVD (the randomized version [Jacobi]) is fantastic when the number of components you want is much smaller than the number of features. The approximation to the largest singular values and vectors is very robust, however, this method loses a lot of accuracy when you want many, many components.
Applications of TruncatedSVD
TruncatedSVD is also known as Latent Semantic Indexing (LSI) which tries to find topics of a word count matrix. If X previously was centered with mean removal, TruncatedSVD is the same as TruncatedPCA. TruncatedSVD is also used in information retrieval tasks, recommendation systems and data compression.
For additional documentation, see scikitlearn’s TruncatedSVD docs.
Examples
# Both import methods supported from cuml import TruncatedSVD from cuml.decomposition import TruncatedSVD import cudf import numpy as np gdf_float = cudf.DataFrame() gdf_float['0'] = np.asarray([1.0,2.0,5.0], dtype = np.float32) gdf_float['1'] = np.asarray([4.0,2.0,1.0], dtype = np.float32) gdf_float['2'] = np.asarray([4.0,2.0,1.0], dtype = np.float32) tsvd_float = TruncatedSVD(n_components = 2, algorithm = "jacobi", n_iter = 20, tol = 1e9) tsvd_float.fit(gdf_float) print(f'components: {tsvd_float.components_}') print(f'explained variance: {tsvd_float._explained_variance_}') exp_var = tsvd_float._explained_variance_ratio_ print(f'explained variance ratio: {exp_var}') print(f'singular values: {tsvd_float._singular_values_}') trans_gdf_float = tsvd_float.transform(gdf_float) print(f'Transformed matrix: {trans_gdf_float}') input_gdf_float = tsvd_float.inverse_transform(trans_gdf_float) print(f'Input matrix: {input_gdf_float}')
Output:
components: 0 1 2 0 0.58725953 0.57233137 0.5723314 1 0.80939883 0.41525528 0.4152552 explained variance: 0 55.33908 1 16.660923 explained variance ratio: 0 0.7685983 1 0.23140171 singular values: 0 7.439024 1 4.0817795 Transformed Matrix: 0 1 2 0 5.1659107 2.512643 1 3.4638448 0.042223275 2 4.0809603 3.2164836 Input matrix: 0 1 2 0 1.0 4.000001 4.000001 1 2.0000005 2.0000005 2.0000007 2 5.000001 0.9999999 1.0000004
 Attributes
 components_array
The top K components (VT.T[:,:n_components]) in U, S, VT = svd(X)
 explained_variance_array
How much each component explains the variance in the data given by S**2
 explained_variance_ratio_array
How much in % the variance is explained given by S**2/sum(S**2)
 singular_values_array
The top K singular values. Remember all singular values >= 0
Methods
fit
(self, X[, y])Fit LSI model on training cudf DataFrame X. 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)¶ Fit LSI model on training cudf DataFrame X. y is currently ignored.
 Parameters
 Xarraylike (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.
 yarraylike (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.

fit_transform
(self, X, y=None)¶ Fit LSI model to X and perform dimensionality reduction on X. y is currently ignored.
 Parameters
 Xarraylike (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.
 yarraylike (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
 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)¶ Transform X back to its original space. Returns X_original whose transform would be X.
 Parameters
 Xarraylike (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)¶ Perform dimensionality reduction on X.
 Parameters
 Xarraylike (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_weights=0.5, target_metric='categorical', handle=None, hash_input=False, random_state=None, optim_batch_size=0, 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 GPUaccelerated 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 1skeleton
‘random’: assign initial embedding positions at random.
 min_dist: float (optional, default 0.1)
The effective minimum distance between embedded points. Smaller values will result in a more clustered/clumped embedding where nearby points on the manifold are drawn closer together, while larger values will result on a more even dispersal of points. The value should be set relative to the
spread
value, which determines the scale at which embedded points will be spread out. spread: float (optional, default 1.0)
The effective scale of embedded points. In combination with
min_dist
this determines how clustered/clumped the embedded points are. set_op_mix_ratio: float (optional, default 1.0)
Interpolate between (fuzzy) union and intersection as the set operation used to combine local fuzzy simplicial sets to obtain a global fuzzy simplicial sets. Both fuzzy set operations use the product tnorm. The value of this parameter should be between 0.0 and 1.0; a value of 1.0 will use a pure fuzzy union, while 0.0 will use a pure fuzzy intersection.
 local_connectivity: int (optional, default 1)
The local connectivity required – i.e. the number of nearest neighbors that should be assumed to be connected at a local level. The higher this value the more connected the manifold becomes locally. In practice this should be not more than the local intrinsic dimension of the manifold.
 repulsion_strength: float (optional, default 1.0)
Weighting applied to negative samples in low dimensional embedding optimization. Values higher than one will result in greater weight being given to negative samples.
 negative_sample_rate: int (optional, default 5)
The number of negative samples to select per positive sample in the optimization process. Increasing this value will result in greater repulsive force being applied, greater optimization cost, but slightly more accuracy.
 transform_queue_size: float (optional, default 4.0)
For transform operations (embedding new points using a trained model this will control how aggressively to search for nearest neighbors. Larger values will result in slower performance but more accurate nearest neighbor evaluation.
 a: float (optional, default None)
More specific parameters controlling the embedding. If None these values are set automatically as determined by
min_dist
andspread
. b: float (optional, default None)
More specific parameters controlling the embedding. If None these values are set automatically as determined by
min_dist
andspread
. hash_input: 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 callingmodel.fit(X).transform(X)
. Not that the CPUbased UMAP reference implementation does this by default. This feature is made optional in the GPU version due to the significant overhead in copying memory to the host for computing the hash. random_stateint, RandomState instance or None, optional (default=None)
random_state is the seed used by the random number generator during embedding initialization and during sampling used by the optimizer. Note: Unfortunately, achieving a high amount of parallelism during the optimization stage often comes at the expense of determinism, since many floatingpoint additions are being made in parallel without a deterministic ordering. This causes slightly different results across training sessions, even when the same seed is used for random number generation. Setting a random_state will enable consistency of trained embeddings, allowing for reproducible results to 3 digits of precision, but will do so at the expense of potentially slower training and increased memory usage.
 optim_batch_size: int (optional, default 100000 / n_components)
Used to maintain the consistency of embeddings for large datasets. The optimization step will be processed with at most optim_batch_size edges at once preventing inconsistencies. A lower batch size will yield more consistently repeatable embeddings at the cost of speed.
 callback: An instance of GraphBasedDimRedCallback class
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(self, embeddings): print(embeddings.copy_to_host()) def on_epoch_end(self, embeddings): print(embeddings.copy_to_host()) def on_train_end(self, embeddings): print(embeddings.copy_to_host())
 verboseint or boolean (default = False)
Controls verbosity of logging.
Notes
This module is heavily based on Leland McInnes’ reference UMAP package. However, there are a number of differences and features that are not yet implemented in cuml.umap:
Using a precomputed pairwise distance matrix (under consideration for future releases)
Manual initialization of initial embedding positions
In addition to these missing features, you should expect to see the final embeddings differing between cuml.umap and the reference UMAP. In particular, the reference UMAP uses an approximate kNN algorithm for large data sizes while cuml.umap always uses exact kNN.
References
 1
 2
Methods
find_ab_params
(spread, min_dist)Function taken from UMAPlearn : https://github.com/lmcinnes/umap Fit a, b params for the differentiable curve used in lower dimensional fuzzy simplicial complex construction.
fit
(self, X[, y, convert_dtype, knn_graph])Fit X into an embedded space.
fit_transform
(self, X[, y, convert_dtype, …])Fit X into an embedded space and return that transformed
transform
(self, X[, convert_dtype, knn_graph])Transform X into the existing embedded space and return that
validate_hyperparams
(self)
static
find_ab_params
(spread, min_dist)¶ Function taken from UMAPlearn : https://github.com/lmcinnes/umap Fit a, b params for the differentiable curve used in lower dimensional fuzzy simplicial complex construction. We want the smooth curve (from a predefined family with simple gradient) that best matches an offset exponential decay.

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

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

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

validate_hyperparams
(self)¶
Random Projections¶

class
cuml.random_projection.
GaussianRandomProjection
(handle=None, n_components='auto', eps=0.1, random_state=None, verbose=False)¶ Gaussian Random Projection method derivated from BaseRandomProjection class.
Random projection is a dimensionality reduction technique. Random projection methods are powerful methods known for their simplicity, computational efficiency and restricted model size. This algorithm also has the advantage to preserve distances well between any two samples and is thus suitable for methods having this requirement.
The components of the random matrix are drawn from N(0, 1 / n_components).
 Parameters
 handlecuml.Handle
If it is None, a new one is created just for this class
 n_componentsint (default = ‘auto’)
Dimensionality of the target projection space. If set to ‘auto’, the parameter is deducted thanks to Johnson–Lindenstrauss lemma. The automatic deduction make use of the number of samples and the eps parameter.
The Johnson–Lindenstrauss lemma can produce very conservative n_components parameter as it makes no assumption on dataset structure.
 epsfloat (default = 0.1)
Error tolerance during projection. Used by Johnson–Lindenstrauss automatic deduction when n_components is set to ‘auto’.
 random_stateint (default = None)
Seed used to initilize random generator
Notes
Inspired by Scikitlearn’s implementation : https://scikitlearn.org/stable/modules/random_projection.html
Examples
from cuml.random_projection import GaussianRandomProjection from sklearn.datasets.samples_generator import make_blobs from sklearn.svm import SVC # dataset generation data, target = make_blobs(n_samples=800, centers=400, n_features=3000, random_state=42) # model fitting model = GaussianRandomProjection(n_components=5, random_state=42).fit(data) # dataset transformation transformed_data = model.transform(data) # classifier training classifier = SVC(gamma=0.001).fit(transformed_data, target) # classifier scoring score = classifier.score(transformed_data, target) # measure information preservation print("Score: {}".format(score))
Output:
Score: 1.0
 Attributes
 gaussian_methodboolean
To be passed to base class in order to determine random matrix generation method

class
cuml.random_projection.
SparseRandomProjection
(handle=None, n_components='auto', density='auto', eps=0.1, dense_output=True, random_state=None, verbose=False)¶ Sparse Random Projection method derivated from BaseRandomProjection class.
Random projection is a dimensionality reduction technique. Random projection methods are powerful methods known for their simplicity, computational efficiency and restricted model size. This algorithm also has the advantage to preserve distances well between any two samples and is thus suitable for methods having this requirement.
Sparse random matrix is an alternative to dense random projection matrix (e.g. Gaussian) that guarantees similar embedding quality while being much more memory efficient and allowing faster computation of the projected data (with sparse enough matrices). If we note
s = 1 / density
the components of the random matrix are drawn from:sqrt(s) / sqrt(n_components)
 with probability1 / 2s
0
 with probability1  1 / s
+sqrt(s) / sqrt(n_components)
 with probability1 / 2s
 Parameters
 handlecuml.Handle
If it is None, a new one is created just for this class
 n_componentsint (default = ‘auto’)
Dimensionality of the target projection space. If set to ‘auto’, the parameter is deducted thanks to Johnson–Lindenstrauss lemma. The automatic deduction make use of the number of samples and the eps parameter. The Johnson–Lindenstrauss lemma can produce very conservative n_components parameter as it makes no assumption on dataset structure.
 densityfloat in range (0, 1] (default = ‘auto’)
Ratio of nonzero component in the random projection matrix. If density = ‘auto’, the value is set to the minimum density as recommended by Ping Li et al.: 1 / sqrt(n_features).
 epsfloat (default = 0.1)
Error tolerance during projection. Used by Johnson–Lindenstrauss automatic deduction when n_components is set to ‘auto’.
 dense_outputboolean (default = True)
If set to True transformed matrix will be dense otherwise sparse.
 random_stateint (default = None)
Seed used to initilize random generator
Notes
Inspired by Scikitlearn’s implementation.
Examples
from cuml.random_projection import SparseRandomProjection from sklearn.datasets.samples_generator import make_blobs from sklearn.svm import SVC # dataset generation data, target = make_blobs(n_samples=800, centers=400, n_features=3000, random_state=42) # model fitting model = SparseRandomProjection(n_components=5, random_state=42).fit(data) # dataset transformation transformed_data = model.transform(data) # classifier training classifier = SVC(gamma=0.001).fit(transformed_data, target) # classifier scoring score = classifier.score(transformed_data, target) # measure information preservation print("Score: {}".format(score))
Output:
Score: 1.0
 Attributes
 gaussian_methodboolean
To be passed to base class in order to determine random matrix generation method

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

class
cuml.
TSNE
(n_components=2, perplexity=30.0, early_exaggeration=12.0, learning_rate=200.0, n_iter=1000, n_iter_without_progress=300, min_grad_norm=1e07, metric='euclidean', init='random', verbose=False, random_state=None, method='barnes_hut', angle=0.5, learning_rate_method='adaptive', n_neighbors=90, perplexity_max_iter=100, exaggeration_iter=250, pre_momentum=0.5, post_momentum=0.8, handle=None)¶ TSNE (TDistributed Stochastic Neighbor Embedding) is an extremely powerful dimensionality reduction technique that aims to maintain local distances between data points. It is extremely robust to whatever dataset you give it, and is used in many areas including cancer research, music analysis and neural network weight visualizations.
Currently, cuML’s TSNE supports the fast Barnes Hut O(NlogN) TSNE approximation (derived from CannyLabs’ BH open source CUDA code). This allows TSNE to produce extremely fast embeddings when n_components = 2. cuML defaults to this algorithm. A slower but more accurate Exact algorithm is also provided.
 Parameters
 n_componentsint (default 2)
The output dimensionality size. Currently only size=2 is tested, but the ‘exact’ algorithm will support greater dimensionality in future.
 perplexityfloat (default 30.0)
Larger datasets require a larger value. Consider choosing different perplexity values from 5 to 50 and see the output differences.
 early_exaggerationfloat (default 12.0)
Controls the space between clusters. Not critical to tune this.
 learning_ratefloat (default 200.0)
The learning rate usually between (10, 1000). If this is too high, TSNE could look like a cloud / ball of points.
 n_iterint (default 1000)
The more epochs, the more stable/accurate the final embedding.
 n_iter_without_progressint (default 300)
Currently unused. When the KL Divergence becomes too small after some iterations, terminate TSNE early.
 min_grad_normfloat (default 1e07)
The minimum gradient norm for when TSNE will terminate early.
 metricstr ‘euclidean’ only (default ‘euclidean’)
Currently only supports euclidean distance. Will support cosine in a future release.
 initstr ‘random’ (default ‘random’)
Currently supports random intialization.
 verboseint or boolean (default = False) (default logger.level_info)
Level of verbosity. Most messages will be printed inside the Python Console.
 random_stateint (default None)
Setting this can allow future runs of TSNE to look mostly the same. It is known that TSNE tends to have vastly different outputs on many runs. Try using PCA intialization (upcoming with change #1098) to possibly counteract this problem. It is known that small perturbations can directly change the result of the embedding for parallel TSNE implementations.
 methodstr ‘barnes_hut’ or ‘exact’ (default ‘barnes_hut’)
Options are either barnes_hut or exact. It is recommended that you use the barnes hut approximation for superior O(nlogn) complexity.
 anglefloat (default 0.5)
Tradeoff between accuracy and speed. Choose between (0,2 0.8) where closer to one indicates full accuracy but slower speeds.
 learning_rate_methodstr ‘adaptive’, ‘none’ or None (default ‘adaptive’)
Either adaptive or None. Uses a special adpative method that tunes the learning rate, early exaggeration and perplexity automatically based on input size.
 n_neighborsint (default 90)
The number of datapoints you want to use in the attractive forces. Smaller values are better for preserving local structure, whilst larger values can improve global structure preservation. Default is 3 * 30 (perplexity)
 perplexity_max_iterint (default 100)
The number of epochs the best gaussian bands are found for.
 exaggeration_iterint (default 250)
To promote the growth of clusters, set this higher.
 pre_momentumfloat (default 0.5)
During the exaggeration iteration, more forcefully apply gradients.
 post_momentumfloat (default 0.8)
During the late phases, less forcefully apply gradients.
 handle(cuML Handle, default None)
You can pass in a past handle that was initialized, or we will create one for you anew!
References
 1
van der Maaten, L.J.P. tDistributed Stochastic Neighbor Embedding
 2
van der Maaten, L.J.P.; Hinton, G.E. Visualizing HighDimensional Data Using tSNE. Journal of Machine Learning Research 9:25792605, 2008.
 3
George C. Linderman, Manas Rachh, Jeremy G. Hoskins, Stefan Steinerberger, Yuval Kluger Efficient Algorithms for tdistributed Stochastic Neighborhood Embedding
Tip
Maaten and Linderman showcased how TSNE can be very sensitive to both the starting conditions (ie random initialization), and how parallel versions of TSNE can generate vastly different results. It has been suggested that you run TSNE a few times to settle on the best configuration. Notice specifying random_state and fixing it across runs can help, but TSNE does not guarantee similar results each time.
As suggested, PCA (upcoming with change #1098) can also help to alleviate this issue.
Note
The CUDA implementation is derived from the excellent CannyLabs open source implementation here: https://github.com/CannyLab/tsnecuda/. The CannyLabs code is licensed according to the conditions in cuml/cpp/src/tsne/ cannylabs_tsne_license.txt. A full description of their approach is available in their article tSNECUDA: GPUAccelerated tSNE and its Applications to Modern Data (https://arxiv.org/abs/1807.11824).
Methods
fit
(self, X[, convert_dtype])Fit X into an embedded space.
fit_transform
(self, X[, convert_dtype])Fit X into an embedded space and return that transformed output.

fit
(self, X, convert_dtype=True)¶ Fit X into an embedded space.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
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.

fit_transform
(self, X, convert_dtype=True)¶ Fit X into an embedded space and return that transformed output.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
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_components)
Embedding of the training data in lowdimensional space.
For more information on how to configure cuML’s output type, refer to: Output Data Type Configuration.
Neighbors¶
Nearest Neighbors¶

class
cuml.neighbors.
NearestNeighbors
(n_neighbors=5, verbose=False, handle=None, algorithm='brute', metric='euclidean', p=2, metric_params=None, output_type=None)¶ NearestNeighbors is an queries neighborhoods from a given set of datapoints. Currently, cuML supports kNN queries, which define the neighborhood as the closest k neighbors to each query point.
 Parameters
 n_neighborsint (default=5)
Default number of neighbors to query
 verboseint or boolean (default = False)
Logging level
 handlecumlHandle
The cumlHandle resources to use
 algorithmstring (default=’brute’)
The query algorithm to use. Currently, only ‘brute’ is supported.
 metricstring (default=’euclidean’).
Distance metric to use. Supported distances are [‘l1, ‘cityblock’, ‘taxicab’, ‘manhattan’, ‘euclidean’, ‘l2’, ‘braycurtis’, ‘canberra’, ‘minkowski’, ‘chebyshev’, ‘jensenshannon’, ‘cosine’, ‘correlation’]
 pfloat (default=2) Parameter for the Minkowski metric. When p = 1, this
is equivalent to manhattan distance (l1), and euclidean distance (l2) for p = 2. For arbitrary p, minkowski distance (lp) is used.
 metric_expandedbool
Can increase performance in Minkowskibased (Lp) metrics (for p > 1) by using the expanded form and not computing the nth roots.
 metric_paramsdict, optional (default = None) This is currently ignored.
Notes
For an additional example see the NearestNeighbors notebook.
For additional docs, see scikitlearn’s NearestNeighbors.
Examples
import cudf from cuml.neighbors import NearestNeighbors from cuml.datasets import make_blobs X, _ = make_blobs(n_samples=25, centers=5, n_features=10, random_state=42) # build a cudf Dataframe X_cudf = cudf.DataFrame(X) # fit model model = NearestNeighbors(n_neighbors=3) model.fit(X) # get 3 nearest neighbors distances, indices = model.kneighbors(X_cudf) # print results print(indices) print(distances)
Output:
indices: 0 1 2 0 0 14 21 1 1 19 8 2 2 9 23 3 3 14 21 ... 22 22 18 11 23 23 16 9 24 24 17 10 distances: 0 1 2 0 0.0 4.883116 5.570006 1 0.0 3.047896 4.105496 2 0.0 3.558557 3.567704 3 0.0 3.806127 3.880100 ... 22 0.0 4.210738 4.227068 23 0.0 3.357889 3.404269 24 0.0 3.428183 3.818043
Methods
fit
(self, X[, convert_dtype])Fit GPU index for performing nearest neighbor queries.
get_param_names
(self)kneighbors
(self[, X, n_neighbors, …])Query the GPU index for the k nearest neighbors of column vectors in X.
kneighbors_graph
(self[, X, n_neighbors, mode])Find the k nearest neighbors of column vectors in X and return as

fit
(self, X, convert_dtype=True)¶ Fit GPU index for performing nearest neighbor queries.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix 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)¶

kneighbors
(self, X=None, n_neighbors=None, return_distance=True, convert_dtype=True)¶ Query the GPU index for the k nearest neighbors of column vectors in X.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
 Dense matrix containing floats or doubles.
 Acceptable formats: CUDA array interface compliant objects like
 CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas
 DataFrame/Series.
 n_neighborsInteger
Number of neighbors to search. If not provided, the n_neighbors from the model instance is used (default=10)
 return_distance: Boolean
If False, distances will not be returned
 convert_dtypebool, optional (default = True)
When set to True, the kneighbors method will automatically convert the inputs to np.float32.
 Returns
 distancescuDF, CuPy or NumPy object depending on cuML’s output typeconfiguration, shape =(n_samples, n_features)
The distances of the knearest neighbors for each column vector in X
 indicescuDF, CuPy or NumPy object depending on cuML’s output typeconfiguration, shape =(n_samples, n_features)
The indices of the knearest neighbors for each column vector in X

kneighbors_graph
(self, X=None, n_neighbors=None, mode='connectivity')¶ Find the k nearest neighbors of column vectors in X and return as a sparse matrix in CSR format.
 Parameters
 Xarraylike (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.
 n_neighborsInteger
Number of neighbors to search. If not provided, the n_neighbors from the model instance is used
 modestring (default=’connectivity’)
Values in connectivity matrix: ‘connectivity’ returns the connectivity matrix with ones and zeros, ‘distance’ returns the edges as the distances between points with the requested metric.
 Returns
 Asparse graph in CSR format, shape = (n_samples, n_samples_fit)
n_samples_fit is the number of samples in the fitted data where A[i, j] is assigned the weight of the edge that connects i to j. Values will either be ones/zeros or the selected distance metric. Return types are either cupy’s CSR sparse graph (device) or numpy’s CSR sparse graph (host)
Nearest Neighbors Classification¶

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

fit
(self, X, y, convert_dtype=True)¶ Fit a GPU index for knearest neighbors classifier model.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.
 yarraylike (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.

get_param_names
(self)¶

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

fit
(self, X, y, convert_dtype=True)¶ Fit a GPU index for knearest neighbors regression model.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix containing floats or doubles. Acceptable formats: CUDA array interface compliant objects like CuPy, cuDF DataFrame/Series, NumPy ndarray and Pandas DataFrame/Series.
 yarraylike (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.

get_param_names
(self)¶

predict
(self, X, convert_dtype=True)¶ Use the trained knearest neighbors regression model to predict the labels for X
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Dense matrix 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.
Time Series¶
HoltWinters¶

class
cuml.
ExponentialSmoothing
(endog, seasonal='additive', seasonal_periods=2, start_periods=2, ts_num=1, eps=0.00224, handle=None)¶ Implements a HoltWinters time series analysis model which is used in both forecasting future entries in a time series as well as in providing exponential smoothing, where weights are assigned against historical data with exponentially decreasing impact. This is done by analyzing three components of the data: level, trend, and seasonality.
 Parameters
 endogarraylike (device or host)
Acceptable formats: cuDF DataFrame, cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy. Note: cuDF.DataFrame types assumes data is in columns, while all other datatypes assume data is in rows. The endogenous dataset to be operated on.
 seasonal‘additive’, ‘add’, ‘multiplicative’, ‘mul’ (default = ‘additive’)
Whether the seasonal trend should be calculated additively or multiplicatively.
 seasonal_periodsint (default=2)
The seasonality of the data (how often it repeats). For monthly data this should be 12, for weekly data, this should be 7.
 start_periodsint (default=2)
Number of seasons to be used for seasonal seed values
 ts_numint (default=1)
The number of different time series that were passed in the endog param.
 epsnp.number > 0 (default=2.24e3)
The accuracy to which gradient descent should achieve. Note that changing this value may affect the forecasted results.
 handlecuml.Handle (default=None)
If it is None, a new one is created just for this class.
Notes
Known Limitations: This version of ExponentialSmoothing currently provides only a limited number of features when compared to the statsmodels.holtwinters.ExponentialSmoothing model. Noticeably, it lacks:
 predictno support for insample prediction.
 hessianno support for returning Hessian matrix.
 informationno support for returning Fisher matrix.
 loglikeno support for returning Loglikelihood.
Additionally, be warned that there may exist floating point instability issues in this model. Small values in endog may lead to faulty results. See https://github.com/rapidsai/cuml/issues/888 for more information.
Known Differences: This version of ExponentialSmoothing differs from statsmodels in some other minor ways:
Cannot pass trend component or damped trend component
this version can take additional parameters eps, start_periods, ts_num, and handle
Score returns SSE rather than gradient logL https://github.com/rapidsai/cuml/issues/876
This version provides get_level(), get_trend(), get_season()
Examples
from cuml import ExponentialSmoothing import cudf import numpy as np data = cudf.Series([1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14], dtype=np.float64) cu_hw = ExponentialSmoothing(data, seasonal_periods=12) cu_hw.fit() cu_pred = cu_hw.forecast(4) print('Forecasted points:', cu_pred)
Output:
Forecasted points : 0 4.000143766093652 1 5.000000163513641 2 6.000000000174092 3 7.000000000000178
Methods
fit
(self)Perform fitting on the given endog dataset.
forecast
(self[, h, index])Forecasts future points based on the fitted model.
get_level
(self[, index])Returns the level component of the model.
get_season
(self[, index])Returns the season component of the model.
get_trend
(self[, index])Returns the trend component of the model.
score
(self[, index])Returns the score of the model.

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

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

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

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

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

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

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

property
aic
¶ Akaike Information Criterion

property
aicc
¶ Corrected Akaike Information Criterion

property
bic
¶ Bayesian Information Criterion

property
complexity
¶ Model complexity (number of parameters)

fit
(self, start_params: Optional[Mapping[str, object]] = None, opt_disp: int = 1, double h: float = 1e8, maxiter: int = 1000, method=u'ml', truncate: int = 0)¶ Fit the ARIMA model to each time series.
 Parameters
 start_paramsMapping[str, object] (optional)
A mapping (e.g dictionary) of parameter names and associated arrays The key names are in {“mu”, “ar”, “ma”, “sar”, “sma”, “sigma2”} The shape of the arrays are (batch_size,) for mu parameters and (n, batch_size) for any other type, where n is the corresponding number of parameters of this type. Pass None for automatic estimation (recommended)
 opt_dispint
Fit diagnostic level (for LBFGS solver):
1 for no output (default)
0<n<100 for output every n steps
n>100 for more detailed output
 hfloat
Finitedifferencing step size. The gradient is computed using forward finite differencing: \(g = \frac{f(x + \mathtt{h})  f(x)}{\mathtt{h}} + O(\mathtt{h})\) # noqa
 maxiterint
Maximum number of iterations of LBFGSB
 methodstr
Estimation method  “css”, “cssml” or “ml”. CSS uses a sumofsquares approximation. ML estimates the loglikelihood with statespace methods. CSSML starts with CSS and refines with ML.
 truncateint
When using CSS, start the sum of squares after a given number of observations

forecast
(self, nsteps: int, level=None)¶ Forecast the given model nsteps into the future.
 Parameters
 nstepsint
The number of steps to forecast beyond end of the given series
 level: float or None (default = None)
Confidence level for prediction intervals, or None to return only the point forecasts. 0 < level < 1
 Returns
 y_fcarraylike
Forecasts. Shape = (nsteps, batch_size)
 lower: arraylike (device) (optional)
Lower limit of the prediction interval if level != None Shape = (end  start, batch_size)
 upper: arraylike (device) (optional)
Upper limit of the prediction interval if level != None Shape = (end  start, batch_size)
Examples
from cuml.tsa.arima import ARIMA ... model = ARIMA(ys, (1,1,1)) model.fit() y_fc = model.forecast(10)

get_params
(self) → Dict[str, np.ndarray]¶ Get the parameters of the model
 Returns
 params: Dict[str, np.ndarray]
A dictionary of parameter names and associated arrays The key names are in {“mu”, “ar”, “ma”, “sar”, “sma”, “sigma2”} The shape of the arrays are (batch_size,) for mu parameters and (n, batch_size) for any other type, where n is the corresponding number of parameters of this type.

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

pack
(self) → np.ndarray¶ Pack parameters of the model into a linearized vector x
 Returns
 xarraylike
Packed parameter array, grouped by series. Shape: (n_params * batch_size,)

predict
(self, start=0, end=None, level=None)¶ Compute insample and/or outofsample prediction for each series
 Parameters
 start: int (default = 0)
Index where to start the predictions (0 <= start <= num_samples)
 end: int (default = None)
Index where to end the predictions, excluded (end > start), or
None
to predict until the last observation level: float or None (default = None)
Confidence level for prediction intervals, or None to return only the point forecasts.
0 < level < 1
 Returns
 y_parraylike (device)
Predictions. Shape = (end  start, batch_size)
 lower: arraylike (device) (optional)
Lower limit of the prediction interval if
level != None
Shape = (end  start, batch_size) upper: arraylike (device) (optional)
Upper limit of the prediction interval if
level != None
Shape = (end  start, batch_size)
Examples
from cuml.tsa.arima import ARIMA model = ARIMA(ys, (1,1,1)) model.fit() y_pred = model.predict()

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

class
cuml.tsa.auto_arima.
AutoARIMA
(endog, handle=None, simple_differencing=True, verbose=4, output_type=None)¶ Implements a batched autoARIMA model for in and outofsample timesseries prediction.
This interface offers a highly customizable search, with functionality similar to the forecast and fable packages in R. It provides an abstraction around the underlying ARIMA models to predict and forecast as if using a single model.
 Parameters
 endogdataframe or arraylike (device or host)
The time series data, assumed to have each time series in columns. Acceptable formats: cuDF DataFrame, cuDF Series, NumPy ndarray, Numba device ndarray, cuda array interface compliant array like CuPy.
 handlecuml.Handle
If it is None, a new one is created just for this instance
 simple_differencing: bool or int (default = True)
If True, the data is differenced before being passed to the Kalman filter. If False, differencing is part of the statespace model. See additional notes in the ARIMA docs
 verboseint
Logging level. It must be one of cuml.common.logger.level_*
 output_type{‘input’, ‘cudf’, ‘cupy’, ‘numpy’}, optional
Variable to control output type of the results and attributes. If None, it’ll inherit the output type set at the module level, cuml.output_type. If set, it will override the global option.
References
The interface was influenced by the R fable package: See https://fable.tidyverts.org/reference/ARIMA.html
A useful (though outdated) reference is the paper: “Automatic Time Series Forecasting: The forecast Package for R”, Rob J. Hyndman & Yeasmin Khandakar (2008), Journal of Statistical Software 27, https://doi.org/10.18637/jss.v027.i03
Examples
from cuml.tsa.auto_arima import AutoARIMA model = AutoARIMA(y) model.search(s=12, d=(0, 1), D=(0, 1), p=(0, 2, 4), q=(0, 2, 4), P=range(2), Q=range(2), method="css", truncate=100) model.fit(method="cssml") fc = model.forecast(20)
Methods
fit
(self, double h, maxiter[, method])Fits the selected models for their respective series
forecast
(self, nsteps[, level])Forecast nsteps into the future.
predict
(self[, start, end, level])Compute insample and/or outofsample prediction for each series
search
(self[, s, d, D, p, q, P, Q, …])Searches through the specified model space and associates each series to the most appropriate model.
summary
(self)Display a quick summary of the models selected by search

fit
(self, double h: float = 1e8, maxiter: int = 1000, method=u'ml', truncate: int = 0)¶ Fits the selected models for their respective series
 Parameters
 hfloat
Finitedifferencing step size used to compute gradients in ARIMA
 maxiterint
Maximum number of iterations of LBFGSB
 methodstr
Estimation method  “css”, “cssml” or “ml”. CSS uses a fast sumofsquares approximation. ML estimates the loglikelihood with statespace methods. CSSML starts with CSS and refines with ML.
 truncateint
When using CSS, start the sum of squares after a given number of observations for better performance (but often a worse fit)

forecast
(self, nsteps: int, level=None)¶ Forecast nsteps into the future.
 nstepsint
The number of steps to forecast beyond end of the given series
 level: float or None (default = None)
Confidence level for prediction intervals, or None to return only the point forecasts. 0 < level < 1
 Returns
 y_fcarraylike
Forecasts. Shape = (nsteps, batch_size)
 lower: arraylike (device) (optional)
Lower limit of the prediction interval if level != None Shape = (end  start, batch_size)
 upper: arraylike (device) (optional)
Upper limit of the prediction interval if level != None Shape = (end  start, batch_size)

predict
(self, start=0, end=None, level=None)¶ Compute insample and/or outofsample prediction for each series
 Returns
 y_parraylike (device)
Predictions. Shape = (end  start, batch_size)
 lower: arraylike (device) (optional)
Lower limit of the prediction interval if level != None Shape = (end  start, batch_size)
 upper: arraylike (device) (optional)
Upper limit of the prediction interval if level != None Shape = (end  start, batch_size)

search
(self, s=None, d=range(3), D=range(2), p=range(1, 4), q=range(1, 4), P=range(3), Q=range(3), fit_intercept=u'auto', ic=u'aicc', test=u'kpss', seasonal_test=u'seas', double h: float = 1e8, maxiter: int = 1000, method=u'auto', truncate: int = 0)¶ Searches through the specified model space and associates each series to the most appropriate model.
 Parameters
 sint
Seasonal period. None or 0 for nonseasonal time series
 dint, sequence or generator
Possible values for d (simple difference)
 Dint, sequence or generator
Possible values for D (seasonal difference)
 pint, sequence or generator
Possible values for p (AR order)
 qint, sequence or generator
Possible values for q (MA order)
 Pint, sequence or generator
Possible values for P (seasonal AR order)
 Qint, sequence or generator
Possible values for Q (seasonal MA order)
 fit_interceptint, sequence, generator or “auto”
Whether to fit an intercept. “auto” chooses based on the model parameters: it uses an incercept iff d + D <= 1
 icstr
Which information criterion to use for the model selection. Currently supported: AIC, AICc, BIC
 teststr
Which stationarity test to use to choose d. Currently supported: KPSS
 seasonal_teststr
Which seasonality test to use to choose D. Currently supported: seas
 hfloat
Finitedifferencing step size used to compute gradients in ARIMA
 maxiterint
Maximum number of iterations of LBFGSB
 methodstr
Estimation method  “auto”, “css”, “cssml” or “ml”. CSS uses a fast sumofsquares approximation. ML estimates the loglikelihood with statespace methods. CSSML starts with CSS and refines with ML. “auto” will use CSS for long seasonal time series, ML otherwise.
 truncateint
When using CSS, start the sum of squares after a given number of observations for better performance. Recommended for long time series when truncating doesn’t lose too much information.

summary
(self)¶ Display a quick summary of the models selected by search
MultiNode, MultiGPU Algorithms¶
KMeans Clustering¶

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

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

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

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

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

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

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

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

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

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

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


class
cuml.dask.neighbors.
KNeighborsRegressor
(client=None, streams_per_handle=0, verbose=False, **kwargs)¶ Multinode MultiGPU KNearest Neighbors Regressor Model.
KNearest Neighbors Regressor is an instancebased learning technique, that keeps training samples around for prediction, rather than trying to learn a generalizable set of model parameters.
Methods
fit
(X, y)Fit a multinode multiGPU KNearest Neighbors Regressor index
predict
(X[, convert_dtype])Predict outputs for a query from previously stored index and outputs.
score
(X, y)Provide score by comparing predictions and ground truth.

fit
(X, y)¶ Fit a multinode multiGPU KNearest Neighbors Regressor index
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Index data. Acceptable formats: dask CuPy/NumPy/Numba Array
 yarraylike (device or host) shape = (n_samples, n_features)
Index output data. Acceptable formats: dask CuPy/NumPy/Numba Array
 Returns
 selfKNeighborsRegressor model

predict
(X, convert_dtype=True)¶ Predict outputs for a query from previously stored index and outputs. The process is done in a multinode multiGPU fashion.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Query data. Acceptable formats: dask cuDF, dask CuPy/NumPy/Numba Array
 convert_dtypebool, optional (default = True)
When set to True, the predict method will automatically convert the data to the right formats.
 Returns
 predictionsDask futures or Dask CuPy Arrays

score
(X, y)¶ Provide score by comparing predictions and ground truth.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Query test data. Acceptable formats: dask CuPy/NumPy/Numba Array
 yarraylike (device or host) shape = (n_samples, n_features)
Outputs test data. Acceptable formats: dask CuPy/NumPy/Numba Array
 Returns
 score


class
cuml.dask.neighbors.
KNeighborsClassifier
(client=None, streams_per_handle=0, verbose=False, **kwargs)¶ Multinode MultiGPU KNearest Neighbors Classifier Model.
KNearest Neighbors Classifier is an instancebased learning technique, that keeps training samples around for prediction, rather than trying to learn a generalizable set of model parameters.
Methods
fit
(X, y)Fit a multinode multiGPU KNearest Neighbors Classifier index
predict
(X[, convert_dtype])Predict labels for a query from previously stored index and index labels.
predict_proba
(X[, convert_dtype])Provide score by comparing predictions and ground truth.
score
(X, y)Predict labels for a query from previously stored index and index labels.

fit
(X, y)¶ Fit a multinode multiGPU KNearest Neighbors Classifier index
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Index data. Acceptable formats: dask CuPy/NumPy/Numba Array
 yarraylike (device or host) shape = (n_samples, n_features)
Index labels data. Acceptable formats: dask CuPy/NumPy/Numba Array
 Returns
 selfKNeighborsClassifier model

predict
(X, convert_dtype=True)¶ Predict labels for a query from previously stored index and index labels. The process is done in a multinode multiGPU fashion.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Query data. Acceptable formats: dask cuDF, dask CuPy/NumPy/Numba Array
 convert_dtypebool, optional (default = True)
When set to True, the predict method will automatically convert the data to the right formats.
 Returns
 predictionsDask futures or Dask CuPy Arrays

predict_proba
(X, convert_dtype=True)¶ Provide score by comparing predictions and ground truth.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Query data. Acceptable formats: dask cuDF, dask CuPy/NumPy/Numba Array
 convert_dtypebool, optional (default = True)
When set to True, the predict method will automatically convert the data to the right formats.
 Returns
 probabilitiesDask futures or Dask CuPy Arrays

score
(X, y)¶ Predict labels for a query from previously stored index and index labels. The process is done in a multinode multiGPU fashion.
 Parameters
 Xarraylike (device or host) shape = (n_samples, n_features)
Query test data. Acceptable formats: dask CuPy/NumPy/Numba Array
 yarraylike (device or host) shape = (n_samples, n_features)
Labels test data. Acceptable formats: dask CuPy/NumPy/Numba Array
 Returns
 score

Principal Component Analysis¶

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

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

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

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

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

class
cuml.dask.ensemble.
RandomForestClassifier
(workers=None, client=None, verbose=False, n_estimators=10, seed=None, ignore_empty_partitions=False, **kwargs)¶ Experimental API implementing a multiGPU Random Forest classifier model which fits multiple decision tree classifiers in an ensemble. This uses Dask to partition data over multiple GPUs (possibly on different nodes).
 Currently, this API makes the following assumptions:
The set of Dask workers used between instantiation, fit, and predict are all consistent
Training data comes in the form of cuDF dataframes or Dask Arrays distributed so that each worker has at least one partition.
The print_summary and print_detailed functions print the information of the forest on the worker.
Future versions of the API will support more flexible data distribution and additional input types.
The distributed algorithm uses an embarrassinglyparallel approach. For a forest with N trees being built on w workers, each worker simply builds N/w trees on the data it has available locally. In many cases, partitioning the data so that each worker builds trees on a subset of the total dataset works well, but it generally requires the data to be wellshuffled in advance. Alternatively, callers can replicate all of the data across workers so that rf.fit receives w partitions, each containing the same data. This would produce results approximately identical to singleGPU fitting.
Please check the singleGPU implementation of Random Forest classifier for more information about the underlying algorithm.
 Parameters
 n_estimatorsint (default = 10)
total number of trees in the forest (not perworker)
 handlecuml.Handle
If it is None, a new one is created just for this class.
 split_criterionThe criterion used to split nodes.
0 for GINI, 1 for ENTROPY, 4 for CRITERION_END. 2 and 3 not valid for classification (default = 0)
 split_algo0 for HIST and 1 for GLOBAL_QUANTILE (default = 1)
the algorithm to determine how nodes are split in the tree.
 split_criterionThe criterion used to split nodes.
0 for GINI, 1 for ENTROPY, 4 for CRITERION_END. 2 and 3 not valid for classification (default = 0)
 bootstrapboolean (default = True)
Control bootstrapping. If set, each tree in the forest is built on a bootstrapped sample with replacement. If false, sampling without replacement is done.
 bootstrap_featuresboolean (default = False)
Control bootstrapping for features. If features are drawn with or without replacement
 rows_samplefloat (default = 1.0)
Ratio of dataset rows used while fitting each tree.
 max_depthint (default = 1)
Maximum tree depth. Unlimited (i.e, until leaves are pure), if 1.
 max_leavesint (default = 1)
Maximum leaf nodes per tree. Soft constraint. Unlimited, if 1.
 max_featuresfloat (default = ‘auto’)
Ratio of number of features (columns) to consider per node split.
 n_binsint (default = 8)
Number of bins used by the split algorithm.
 min_rows_per_nodeint (default = 2)
The minimum number of samples (rows) needed to split a node.
 quantile_per_treeboolean (default = False)
Whether quantile is computed for individual RF trees. Only relevant for GLOBAL_QUANTILE split_algo.
 n_streamsint (default = 4 )
Number of parallel streams used for forest building
 workersoptional, list of strings
Dask addresses of workers to use for computation. If None, all available Dask workers will be used.
 seedint (default = None)
Base seed for the random number generator. Unseeded by default. Does not currently fully guarantee the exact same results.
 ignore_empty_partitions: Boolean (default = False)
Specify behavior when a worker does not hold any data while splitting. When True, it returns the results from workers with data (the number of trained estimators will be less than n_estimators) When False, throws a RuntimeError. This is an experiemental parameter, and may be removed in the future.
Examples
For usage examples, please see the RAPIDS notebooks repository: https://github.com/rapidsai/cuml/blob/branch0.15/notebooks/random_forest_mnmg_demo.ipynb
Methods
fit
(X, y[, convert_dtype])Fit the input data with a Random Forest classifier
get_params
([deep])Returns the value of all parameters required to configure this estimator as a dictionary.
predict
(X[, output_class, algo, threshold, …])Predicts the labels for X.
predict_model_on_cpu
(X[, convert_dtype])Predicts the labels for X.
predict_proba
(X[, delayed])Predicts the probability of each class for X.
Print detailed information of the forest used to train the model on each worker.
Print the summary of the forest used to train the model on each worker.
set_params
(**params)Sets the value of parameters required to configure this estimator, it functions similar to the sklearn set_params.
predict_using_fil

fit
(X, y, convert_dtype=False)¶ Fit the input data with a Random Forest classifier
IMPORTANT: X is expected to be partitioned with at least one partition on each Dask worker being used by the forest (self.workers).
If a worker has multiple data partitions, they will be concatenated before fitting, which will lead to additional memory usage. To minimize memory consumption, ensure that each worker has exactly one partition.
When persisting data, you can use cuml.dask.common.utils.persist_across_workers to simplify this:
X_dask_cudf = dask_cudf.from_cudf(X_cudf, npartitions=n_workers) y_dask_cudf = dask_cudf.from_cudf(y_cudf, npartitions=n_workers) X_dask_cudf, y_dask_cudf = persist_across_workers(dask_client, [X_dask_cudf, y_dask_cudf])
This is equivalent to calling persist with the data and workers:
X_dask_cudf, y_dask_cudf = dask_client.persist([X_dask_cudf, y_dask_cudf], workers={ X_dask_cudf=workers, y_dask_cudf=workers })
 Parameters
 XDask cuDF dataframe or CuPy backed Dask Array (n_rows, n_features)
Distributed dense matrix (floats or doubles) of shape (n_samples, n_features).
 yDask cuDF dataframe or CuPy backed Dask Array (n_rows, 1)
Labels of training examples. y must be partitioned the same way as X
 convert_dtypebool, optional (default = False)
When set to True, the fit method will, when necessary, convert y to be of dtype int32. This will increase memory used for the method.

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

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

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

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

print_detailed
()¶ Print detailed information of the forest used to train the model on each worker. This information is displayed on the workers and not the client.

print_summary
()¶ Print the summary of the forest used to train the model on each worker. This information is displayed on the individual workers and not the client.

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

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

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

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

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

print_detailed
()¶ Print detailed information of the forest used to train the model on each worker. This information is displayed on the workers and not the client.

print_summary
()¶ Print the summary of the forest used to train the model on each worker. This information is displayed on the individual workers and not the client.

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

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

class
cuml.dask.naive_bayes.
MultinomialNB
(client=None, verbose=False, **kwargs)¶ Distributed Naive Bayes classifier for multinomial models
Examples
Load the 20 newsgroups dataset from Scikitlearn and train a Naive Bayes classifier.
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 # 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) # Train model model = MultinomialNB() model.fit(X, y) # Compute accuracy on training set model.score(X, y)
Output:
0.9244298934936523
Methods
fit
(X, y[, classes])Fit distributed Naive Bayes classifier model
predict
(X)Use distributed Naive Bayes model to predict the classes for a given set of data samples.
score
(X, y)Compute accuracy score

fit
(X, y, classes=None)¶ Fit distributed Naive Bayes classifier model
 Parameters
 Xdask.Array with blocks containing dense or sparse cupy arrays
 ydask.Array with blocks containing cupy.ndarray
 classesarraylike containing unique class labels
 Returns
 cuml.dask.naive_bayes.MultinomialNB current model instance

predict
(X)¶ Use distributed Naive Bayes model to predict the classes for a given set of data samples.
 Parameters
 Xdask.Array with blocks containing dense or sparse cupy arrays
 Returns
 dask.Array containing predicted classes

score
(X, y)¶ Compute accuracy score
 Parameters
 XDask.Array
Features to predict. Note it is assumed that chunk sizes and shape of X are known. This can be done for a fully delayed Array by calling X.compute_chunks_sizes()
 yDask.Array
Labels to use for computing accuracy. Note it is assumed that chunk sizes and shape of X are known. This can be done for a fully delayed Array by calling X.compute_chunks_sizes()
 Returns
 scorefloat the resulting accuracy score
