Skip to content

Latest commit

 

History

History
 
 

cuml_bench

README.md

How to create conda environment for benchmarking

conda create -n bench -c rapidsai -c conda-forge python=3.7 scikit-learn cuml pandas cudf tqdm

Algorithms parameters

You can launch benchmarks for each algorithm separately. The tables below list all supported parameters for each algorithm:

General

Parameter Name Type Default Value Description
num-threads int -1 The number of threads to use
arch str ? Achine architecture, for bookkeeping
batch str ? Batch ID, for bookkeeping
prefix str sklearn Prefix string, for bookkeeping
header action False Output CSV header
verbose action False Output extra debug messages
data-format str numpy Data formats: numpy, pandas or cudf
data-order str C Data order: C (row-major, default) or F (column-major)
dtype np.dtype np.float64 Data type: float64 (default) or float32
check-finiteness action False Check finiteness in sklearn input check(disabled by default)
output-format str csv Output format: csv (default) or json'
time-method str mean_min Method used for time mesurements
box-filter-measurements int 100 Maximum number of measurements in box filter
inner-loops int 100 Maximum inner loop iterations. (we take the mean over inner iterations)
outer-loops int 100 Maximum outer loop iterations. (we take the min over outer iterations)
time-limit float 10 Target time to spend to benchmark
goal-outer-loops int 10 The number of outer loops to aim while automatically picking number of inner loops. If zero, do not automatically decide number of inner loops
seed int 12345 Seed to pass as random_state
dataset-name str None Dataset name

DBSCAN

parameter Name Type default value description
epsilon float 10 Radius of neighborhood of a point
min_samples int 5 The minimum number of samples required in a 'neighborhood to consider a point a core point

RandomForestClassifier

parameter Name Type default value description
criterion str gini gini or entropy. The function to measure the quality of a split
split-algorithm str hist hist or global_quantile. The algorithm to determine how nodes are split in the tree
num-trees int 100 The number of trees in the forest
max-features float_or_int None Upper bound on features used at each split
max-depth int None Upper bound on depth of constructed trees
min-samples-split float_or_int 2 Minimum samples number for node splitting
max-leaf-nodes int None Maximum leaf nodes per tree
min-impurity-decrease float 0 Needed impurity decrease for node splitting
no-bootstrap store_false True Don't control bootstraping

RandomForestRegressor

parameter Name Type default value description
criterion str gini gini or entropy. The function to measure the quality of a split
split-algorithm str hist hist or global_quantile. The algorithm to determine how nodes are split in the tree
num-trees int 100 The number of trees in the forest
max-features float_or_int None Upper bound on features used at each split
max-depth int None Upper bound on depth of constructed trees
min-samples-split float_or_int 2 Minimum samples number for node splitting
max-leaf-nodes int None Maximum leaf nodes per tree
min-impurity-decrease float 0 Needed impurity decrease for node splitting
no-bootstrap action True Don't control bootstraping

KMeans

parameter Name Type default value description
init str Initial clusters
tol float 0 Absolute threshold
maxiter int 100 Maximum number of iterations
samples-per-batch int 32768 The number of samples per batch
n-clusters int The number of clusters

KNeighborsClassifier

parameter Name Type default value description
n-neighbors int 5 The number of neighbors to use
weights str uniform Weight function used in prediction
method str brute Algorithm used to compute the nearest neighbors
metric str euclidean Distance metric to use

LinearRegression

parameter Name Type default value description
no-fit-intercept action True Don't fit intercept (assume data already centered)
solver str eig eig or svd. Solver used for training

LogisticRegression

parameter Name Type default value description
no-fit-intercept action True Don't fit intercept
solver str qn qn, owl. Solver to use
maxiter int 100 Maximum iterations for the iterative solver
C float 1.0 Regularization parameter
tol float None Tolerance for solver

PCA

parameter Name Type default value description
svd-solver str full auto, full or jacobi. SVD solver to use
n-components int None The number of components to find
whiten action False Perform whitening

Ridge

parameter Name Type default value description
no-fit-intercept action True Don't fit intercept (assume data already centered)
solver str eig eig, cd or svd. Solver used for training
alpha float 1.0 Regularization strength

SVC

parameter Name Type default value description
C float 0.01 SVM slack parameter
kernel str linear linear or rbf. SVM kernel function
gamma float None Parameter for kernel="rbf"
max-cache-size int 64 Maximum cache size for SVM.
tol float 1e-16 Tolerance passed to sklearn.svm.SVC
probability action True Use probability for SVC

TSNE

parameter Name Type default value description
n-components int 2 Dimension of the embedded space
early-exaggeration float 12.0 This factor increases the attractive forces between points
and allows points to move around more freely finding their nearest neighbors more easily
learning-rate float 200.0 The learning rate for t-SNE is usually in the range [10.0, 1000.0]
angle float 0.5 Angular size. This is the trade-off between speed and accuracy
min-grad-norm float 1e-7 If the gradient norm is below this threshold, the optimization is stopped
random-state int 1234 Determines the random number generator

train_test_split

parameter Name Type default value description
train-size float 0.75 Size of training subset
test-size float 0.25 Size of testing subset
do-not-shuffle action False Do not perform data shuffle before splitting