Prediction Latency

This is an example showing the prediction latency of various scikit-learn estimators.

The goal is to measure the latency one can expect when doing predictions either in bulk or atomic (i.e. one by one) mode.

The plots represent the distribution of the prediction latency as a boxplot.

# Authors: Eustache Diemert <eustache@diemert.fr>
# License: BSD 3 clause
 
from __future__ import print_function
from collections import defaultdict
 
import time
import gc
import numpy as np
import matplotlib.pyplot as plt
 
from sklearn.preprocessing import StandardScaler
from sklearn.model_selection import train_test_split
from scipy.stats import scoreatpercentile
from sklearn.datasets.samples_generator import make_regression
from sklearn.ensemble.forest import RandomForestRegressor
from sklearn.linear_model.ridge import Ridge
from sklearn.linear_model.stochastic_gradient import SGDRegressor
from sklearn.svm.classes import SVR
from sklearn.utils import shuffle
 
 
def _not_in_sphinx():
    # Hack to detect whether we are running by the sphinx builder
    return '__file__' in globals()
 
 
def atomic_benchmark_estimator(estimator, X_test, verbose=False):
    """Measure runtime prediction of each instance."""
    n_instances = X_test.shape[0]
    runtimes = np.zeros(n_instances, dtype=np.float)
    for i in range(n_instances):
        instance = X_test[[i], :]
        start = time.time()
        estimator.predict(instance)
        runtimes[i] = time.time() - start
    if verbose:
        print("atomic_benchmark runtimes:", min(runtimes), scoreatpercentile(
            runtimes, 50), max(runtimes))
    return runtimes
 
 
def bulk_benchmark_estimator(estimator, X_test, n_bulk_repeats, verbose):
    """Measure runtime prediction of the whole input."""
    n_instances = X_test.shape[0]
    runtimes = np.zeros(n_bulk_repeats, dtype=np.float)
    for i in range(n_bulk_repeats):
        start = time.time()
        estimator.predict(X_test)
        runtimes[i] = time.time() - start
    runtimes = np.array(list(map(lambda x: x / float(n_instances), runtimes)))
    if verbose:
        print("bulk_benchmark runtimes:", min(runtimes), scoreatpercentile(
            runtimes, 50), max(runtimes))
    return runtimes
 
 
def benchmark_estimator(estimator, X_test, n_bulk_repeats=30, verbose=False):
    """
    Measure runtimes of prediction in both atomic and bulk mode.
 
    Parameters
    ----------
    estimator : already trained estimator supporting `predict()`
    X_test : test input
    n_bulk_repeats : how many times to repeat when evaluating bulk mode
 
    Returns
    -------
    atomic_runtimes, bulk_runtimes : a pair of `np.array` which contain the
    runtimes in seconds.
 
    """
    atomic_runtimes = atomic_benchmark_estimator(estimator, X_test, verbose)
    bulk_runtimes = bulk_benchmark_estimator(estimator, X_test, n_bulk_repeats,
                                             verbose)
    return atomic_runtimes, bulk_runtimes
 
 
def generate_dataset(n_train, n_test, n_features, noise=0.1, verbose=False):
    """Generate a regression dataset with the given parameters."""
    if verbose:
        print("generating dataset...")
 
    X, y, coef = make_regression(n_samples=n_train + n_test,
                                 n_features=n_features, noise=noise, coef=True)
 
    random_seed = 13
    X_train, X_test, y_train, y_test = train_test_split(
        X, y, train_size=n_train, random_state=random_seed)
    X_train, y_train = shuffle(X_train, y_train, random_state=random_seed)
 
    X_scaler = StandardScaler()
    X_train = X_scaler.fit_transform(X_train)
    X_test = X_scaler.transform(X_test)
 
    y_scaler = StandardScaler()
    y_train = y_scaler.fit_transform(y_train[:, None])[:, 0]
    y_test = y_scaler.transform(y_test[:, None])[:, 0]
 
    gc.collect()
    if verbose:
        print("ok")
    return X_train, y_train, X_test, y_test
 
 
def boxplot_runtimes(runtimes, pred_type, configuration):
    """
    Plot a new `Figure` with boxplots of prediction runtimes.
 
    Parameters
    ----------
    runtimes : list of `np.array` of latencies in micro-seconds
    cls_names : list of estimator class names that generated the runtimes
    pred_type : 'bulk' or 'atomic'
 
    """
 
    fig, ax1 = plt.subplots(figsize=(10, 6))
    bp = plt.boxplot(runtimes, )
 
    cls_infos = ['%s\n(%d %s)' % (estimator_conf['name'],
                                  estimator_conf['complexity_computer'](
                                      estimator_conf['instance']),
                                  estimator_conf['complexity_label']) for
                 estimator_conf in configuration['estimators']]
    plt.setp(ax1, xticklabels=cls_infos)
    plt.setp(bp['boxes'], color='black')
    plt.setp(bp['whiskers'], color='black')
    plt.setp(bp['fliers'], color='red', marker='+')
 
    ax1.yaxis.grid(True, linestyle='-', which='major', color='lightgrey',
                   alpha=0.5)
 
    ax1.set_axisbelow(True)
    ax1.set_title('Prediction Time per Instance - %s, %d feats.' % (
        pred_type.capitalize(),
        configuration['n_features']))
    ax1.set_ylabel('Prediction Time (us)')
 
    plt.show()
 
 
def benchmark(configuration):
    """Run the whole benchmark."""
    X_train, y_train, X_test, y_test = generate_dataset(
        configuration['n_train'], configuration['n_test'],
        configuration['n_features'])
 
    stats = {}
    for estimator_conf in configuration['estimators']:
        print("Benchmarking", estimator_conf['instance'])
        estimator_conf['instance'].fit(X_train, y_train)
        gc.collect()
        a, b = benchmark_estimator(estimator_conf['instance'], X_test)
        stats[estimator_conf['name']] = {'atomic': a, 'bulk': b}
 
    cls_names = [estimator_conf['name'] for estimator_conf in configuration[
        'estimators']]
    runtimes = [1e6 * stats[clf_name]['atomic'] for clf_name in cls_names]
    boxplot_runtimes(runtimes, 'atomic', configuration)
    runtimes = [1e6 * stats[clf_name]['bulk'] for clf_name in cls_names]
    boxplot_runtimes(runtimes, 'bulk (%d)' % configuration['n_test'],
                     configuration)
 
 
def n_feature_influence(estimators, n_train, n_test, n_features, percentile):
    """
    Estimate influence of the number of features on prediction time.
 
    Parameters
    ----------
 
    estimators : dict of (name (str), estimator) to benchmark
    n_train : nber of training instances (int)
    n_test : nber of testing instances (int)
    n_features : list of feature-space dimensionality to test (int)
    percentile : percentile at which to measure the speed (int [0-100])
 
    Returns:
    --------
 
    percentiles : dict(estimator_name,
                       dict(n_features, percentile_perf_in_us))
 
    """
    percentiles = defaultdict(defaultdict)
    for n in n_features:
        print("benchmarking with %d features" % n)
        X_train, y_train, X_test, y_test = generate_dataset(n_train, n_test, n)
        for cls_name, estimator in estimators.items():
            estimator.fit(X_train, y_train)
            gc.collect()
            runtimes = bulk_benchmark_estimator(estimator, X_test, 30, False)
            percentiles[cls_name][n] = 1e6 * scoreatpercentile(runtimes,
                                                               percentile)
    return percentiles
 
 
def plot_n_features_influence(percentiles, percentile):
    fig, ax1 = plt.subplots(figsize=(10, 6))
    colors = ['r', 'g', 'b']
    for i, cls_name in enumerate(percentiles.keys()):
        x = np.array(sorted([n for n in percentiles[cls_name].keys()]))
        y = np.array([percentiles[cls_name][n] for n in x])
        plt.plot(x, y, color=colors[i], )
    ax1.yaxis.grid(True, linestyle='-', which='major', color='lightgrey',
                   alpha=0.5)
    ax1.set_axisbelow(True)
    ax1.set_title('Evolution of Prediction Time with #Features')
    ax1.set_xlabel('#Features')
    ax1.set_ylabel('Prediction Time at %d%%-ile (us)' % percentile)
    plt.show()
 
 
def benchmark_throughputs(configuration, duration_secs=0.1):
    """benchmark throughput for different estimators."""
    X_train, y_train, X_test, y_test = generate_dataset(
        configuration['n_train'], configuration['n_test'],
        configuration['n_features'])
    throughputs = dict()
    for estimator_config in configuration['estimators']:
        estimator_config['instance'].fit(X_train, y_train)
        start_time = time.time()
        n_predictions = 0
        while (time.time() - start_time) < duration_secs:
            estimator_config['instance'].predict(X_test[[0]])
            n_predictions += 1
        throughputs[estimator_config['name']] = n_predictions / duration_secs
    return throughputs
 
 
def plot_benchmark_throughput(throughputs, configuration):
    fig, ax = plt.subplots(figsize=(10, 6))
    colors = ['r', 'g', 'b']
    cls_infos = ['%s\n(%d %s)' % (estimator_conf['name'],
                                  estimator_conf['complexity_computer'](
                                      estimator_conf['instance']),
                                  estimator_conf['complexity_label']) for
                 estimator_conf in configuration['estimators']]
    cls_values = [throughputs[estimator_conf['name']] for estimator_conf in
                  configuration['estimators']]
    plt.bar(range(len(throughputs)), cls_values, width=0.5, color=colors)
    ax.set_xticks(np.linspace(0.25, len(throughputs) - 0.75, len(throughputs)))
    ax.set_xticklabels(cls_infos, fontsize=10)
    ymax = max(cls_values) * 1.2
    ax.set_ylim((0, ymax))
    ax.set_ylabel('Throughput (predictions/sec)')
    ax.set_title('Prediction Throughput for different estimators (%d '
                 'features)' % configuration['n_features'])
    plt.show()

main code

start_time = time.time()
 
# benchmark bulk/atomic prediction speed for various regressors
configuration = {
    'n_train': int(1e3),
    'n_test': int(1e2),
    'n_features': int(1e2),
    'estimators': [
        {'name': 'Linear Model',
         'instance': SGDRegressor(penalty='elasticnet', alpha=0.01,
                                  l1_ratio=0.25, fit_intercept=True),
         'complexity_label': 'non-zero coefficients',
         'complexity_computer': lambda clf: np.count_nonzero(clf.coef_)},
        {'name': 'RandomForest',
         'instance': RandomForestRegressor(),
         'complexity_label': 'estimators',
         'complexity_computer': lambda clf: clf.n_estimators},
        {'name': 'SVR',
         'instance': SVR(kernel='rbf'),
         'complexity_label': 'support vectors',
         'complexity_computer': lambda clf: len(clf.support_vectors_)},
    ]
}
benchmark(configuration)
 
# benchmark n_features influence on prediction speed
percentile = 90
percentiles = n_feature_influence({'ridge': Ridge()},
                                  configuration['n_train'],
                                  configuration['n_test'],
                                  [100, 250, 500], percentile)
plot_n_features_influence(percentiles, percentile)
 
# benchmark throughput
throughputs = benchmark_throughputs(configuration)
plot_benchmark_throughput(throughputs, configuration)
 
stop_time = time.time()
print("example run in %.2fs" % (stop_time - start_time))

Out:

  Benchmarking SGDRegressor(alpha=0.01, average=False, epsilon=0.1, eta0=0.01,
       fit_intercept=True, l1_ratio=0.25, learning_rate='invscaling',
       loss='squared_loss', n_iter=5, penalty='elasticnet', power_t=0.25,
       random_state=None, shuffle=True, verbose=0, warm_start=False)
Benchmarking RandomForestRegressor(bootstrap=True, criterion='mse', max_depth=None,
           max_features='auto', max_leaf_nodes=None,
           min_impurity_split=1e-07, min_samples_leaf=1,
           min_samples_split=2, min_weight_fraction_leaf=0.0,
           n_estimators=10, n_jobs=1, oob_score=False, random_state=None,
           verbose=0, warm_start=False)
Benchmarking SVR(C=1.0, cache_size=200, coef0=0.0, degree=3, epsilon=0.1, gamma='auto',
  kernel='rbf', max_iter=-1, shrinking=True, tol=0.001, verbose=False)
benchmarking with 100 features
benchmarking with 250 features
benchmarking with 500 features
example run in 3.53s

Total running time of the script: (0 minutes 3.532 seconds)

Download Python source code: plot_prediction_latency.py

Download IPython notebook: plot_prediction_latency.ipynb

Links:

http://scikit-learn.org/stable/auto_examples/applications/plot_prediction_latency.html

doc_scikit_learn

2025-01-10 15:47:30

Comments