Benchmark Random Forests, Tree Ensemble, (AoS and SoA)#

The script compares different implementations for the operator TreeEnsembleRegressor.

  • baseline: RandomForestRegressor from scikit-learn

  • ort: onnxruntime,

  • mlprodict: an implementation based on an array of structures, every structure describes a node,

  • mlprodict2 similar implementation but instead of having an array of structures, it relies on a structure of arrays, it parallelizes by blocks of 128 observations and inside every block, goes through trees then through observations (double loop),

  • mlprodict3: parallelizes by trees, this implementation is faster when the depth is higher than 10.

A structure of arrays has better performance: Case study: Comparing Arrays of Structures and Structures of Arrays Data Layouts for a Compute-Intensive Loop. See also AoS and SoA.

Profile the execution

py-spy can be used to profile the execution of a program. The profile is more informative if the code is compiled with debug information.

py-spy record --native -r 10 -o plot_random_forest_reg.svg -- python plot_random_forest_reg.py

Import#

import warnings
from time import perf_counter as time
from multiprocessing import cpu_count
import numpy
from numpy.random import rand
from numpy.testing import assert_almost_equal
import pandas
import matplotlib.pyplot as plt
from sklearn import config_context
from sklearn.ensemble import RandomForestRegressor
from sklearn.utils._testing import ignore_warnings
from skl2onnx import convert_sklearn
from skl2onnx.common.data_types import FloatTensorType
from onnxruntime import InferenceSession
from mlprodict.onnxrt import OnnxInference

Available optimisation on this machine.

from mlprodict.testing.experimental_c_impl.experimental_c import code_optimisation
print(code_optimisation())

Out:

AVX-omp=8

Versions#

def version():
    from datetime import datetime
    import sklearn
    import numpy
    import onnx
    import onnxruntime
    import skl2onnx
    import mlprodict
    df = pandas.DataFrame([
        {"name": "date", "version": str(datetime.now())},
        {"name": "numpy", "version": numpy.__version__},
        {"name": "scikit-learn", "version": sklearn.__version__},
        {"name": "onnx", "version": onnx.__version__},
        {"name": "onnxruntime", "version": onnxruntime.__version__},
        {"name": "skl2onnx", "version": skl2onnx.__version__},
        {"name": "mlprodict", "version": mlprodict.__version__},
    ])
    return df


version()
name version
0 date 2022-04-05 06:06:45.763336
1 numpy 1.21.5
2 scikit-learn 1.0.2
3 onnx 1.11.0
4 onnxruntime 1.11.0
5 skl2onnx 1.11.1
6 mlprodict 0.8.1762


Implementations to benchmark#

def fcts_model(X, y, max_depth, n_estimators, n_jobs):
    "RandomForestClassifier."
    rf = RandomForestRegressor(max_depth=max_depth, n_estimators=n_estimators,
                               n_jobs=n_jobs)
    rf.fit(X, y)

    initial_types = [('X', FloatTensorType([None, X.shape[1]]))]
    onx = convert_sklearn(rf, initial_types=initial_types)
    sess = InferenceSession(onx.SerializeToString())
    outputs = [o.name for o in sess.get_outputs()]
    oinf = OnnxInference(onx, runtime="python")
    oinf.sequence_[0].ops_._init(numpy.float32, 1)
    name = outputs[0]
    oinf2 = OnnxInference(onx, runtime="python")
    oinf2.sequence_[0].ops_._init(numpy.float32, 2)
    oinf3 = OnnxInference(onx, runtime="python")
    oinf3.sequence_[0].ops_._init(numpy.float32, 3)

    def predict_skl_predict(X, model=rf):
        return rf.predict(X)

    def predict_onnxrt_predict(X, sess=sess):
        return sess.run(outputs[:1], {'X': X})[0]

    def predict_onnx_inference(X, oinf=oinf):
        return oinf.run({'X': X})[name]

    def predict_onnx_inference2(X, oinf2=oinf2):
        return oinf2.run({'X': X})[name]

    def predict_onnx_inference3(X, oinf3=oinf3):
        return oinf3.run({'X': X})[name]

    return {'predict': (
        predict_skl_predict, predict_onnxrt_predict,
        predict_onnx_inference, predict_onnx_inference2,
        predict_onnx_inference3)}

Benchmarks#

def allow_configuration(**kwargs):
    return True


def bench(n_obs, n_features, max_depths, n_estimatorss, n_jobss,
          methods, repeat=10, verbose=False):
    res = []
    for nfeat in n_features:

        ntrain = 50000
        X_train = numpy.empty((ntrain, nfeat)).astype(numpy.float32)
        X_train[:, :] = rand(ntrain, nfeat)[:, :]
        eps = rand(ntrain) - 0.5
        y_train = X_train.sum(axis=1) + eps

        for n_jobs in n_jobss:
            for max_depth in max_depths:
                for n_estimators in n_estimatorss:
                    fcts = fcts_model(X_train, y_train,
                                      max_depth, n_estimators, n_jobs)

                    for n in n_obs:
                        for method in methods:

                            fct1, fct2, fct3, fct4, fct5 = fcts[method]

                            if not allow_configuration(
                                    n=n, nfeat=nfeat, max_depth=max_depth,
                                    n_estimator=n_estimators, n_jobs=n_jobs,
                                    method=method):
                                continue

                            obs = dict(n_obs=n, nfeat=nfeat,
                                       max_depth=max_depth,
                                       n_estimators=n_estimators,
                                       method=method,
                                       n_jobs=n_jobs)

                            # creates different inputs to avoid caching
                            Xs = []
                            for r in range(repeat):
                                x = numpy.empty((n, nfeat))
                                x[:, :] = rand(n, nfeat)[:, :]
                                Xs.append(x.astype(numpy.float32))

                            # measures the baseline
                            with config_context(assume_finite=True):
                                st = time()
                                repeated = 0
                                for X in Xs:
                                    p1 = fct1(X)
                                    repeated += 1
                                    if time() - st >= 1:
                                        break  # stops if longer than a second
                                end = time()
                                obs["time_skl"] = (end - st) / repeated

                            # measures the new implementation
                            st = time()
                            r2 = 0
                            for X in Xs:
                                p2 = fct2(X)
                                r2 += 1
                                if r2 >= repeated:
                                    break
                            end = time()
                            obs["time_ort"] = (end - st) / r2

                            # measures the other new implementation
                            st = time()
                            r2 = 0
                            for X in Xs:
                                p2 = fct3(X)
                                r2 += 1
                                if r2 >= repeated:
                                    break
                            end = time()
                            obs["time_mlprodict"] = (end - st) / r2

                            # measures the other new implementation 2
                            st = time()
                            r2 = 0
                            for X in Xs:
                                p2 = fct4(X)
                                r2 += 1
                                if r2 >= repeated:
                                    break
                            end = time()
                            obs["time_mlprodict2"] = (end - st) / r2

                            # measures the other new implementation 3
                            st = time()
                            r2 = 0
                            for X in Xs:
                                p2 = fct5(X)
                                r2 += 1
                                if r2 >= repeated:
                                    break
                            end = time()
                            obs["time_mlprodict3"] = (end - st) / r2

                            # final
                            res.append(obs)
                            if verbose and (len(res) % 1 == 0 or n >= 10000):
                                print("bench", len(res), ":", obs)

                            # checks that both produce the same outputs
                            if n <= 10000:
                                if len(p1.shape) == 1 and len(p2.shape) == 2:
                                    p2 = p2.ravel()
                                try:
                                    assert_almost_equal(
                                        p1.ravel(), p2.ravel(), decimal=5)
                                except AssertionError as e:
                                    warnings.warn(str(e))
    return res

Graphs#

def plot_rf_models(dfr):

    def autolabel(ax, rects):
        for rect in rects:
            height = rect.get_height()
            ax.annotate('%1.1fx' % height,
                        xy=(rect.get_x() + rect.get_width() / 2, height),
                        xytext=(0, 3),  # 3 points vertical offset
                        textcoords="offset points",
                        ha='center', va='bottom',
                        fontsize=8)

    engines = [_.split('_')[-1] for _ in dfr.columns if _.startswith("time_")]
    engines = [_ for _ in engines if _ != 'skl']
    for engine in engines:
        dfr["speedup_%s" % engine] = dfr["time_skl"] / dfr["time_%s" % engine]
    print(dfr.tail().T)

    ncols = 4
    fig, axs = plt.subplots(len(engines), ncols, figsize=(
        14, 4 * len(engines)), sharey=True)

    row = 0
    for row, engine in enumerate(engines):
        pos = 0
        name = "RandomForestRegressor - %s" % engine
        for max_depth in sorted(set(dfr.max_depth)):
            for nf in sorted(set(dfr.nfeat)):
                for est in sorted(set(dfr.n_estimators)):
                    for n_jobs in sorted(set(dfr.n_jobs)):
                        sub = dfr[(dfr.max_depth == max_depth) &
                                  (dfr.nfeat == nf) &
                                  (dfr.n_estimators == est) &
                                  (dfr.n_jobs == n_jobs)]
                        ax = axs[row, pos]
                        labels = sub.n_obs
                        means = sub["speedup_%s" % engine]

                        x = numpy.arange(len(labels))
                        width = 0.90

                        rects1 = ax.bar(x, means, width, label='Speedup')
                        if pos == 0:
                            ax.set_yscale('log')
                            ax.set_ylim([0.1, max(dfr["speedup_%s" % engine])])

                        if pos == 0:
                            ax.set_ylabel('Speedup')
                        ax.set_title(
                            '%s\ndepth %d - %d features\n %d estimators %d '
                            'jobs' % (name, max_depth, nf, est, n_jobs))
                        if row == len(engines) - 1:
                            ax.set_xlabel('batch size')
                        ax.set_xticks(x)
                        ax.set_xticklabels(labels)
                        autolabel(ax, rects1)
                        for tick in ax.xaxis.get_major_ticks():
                            tick.label.set_fontsize(8)
                        for tick in ax.yaxis.get_major_ticks():
                            tick.label.set_fontsize(8)
                        pos += 1

    fig.tight_layout()
    return fig, ax

Run benchs#

@ignore_warnings(category=FutureWarning)
def run_bench(repeat=100, verbose=False):
    n_obs = [1, 10, 100, 1000, 10000]
    methods = ['predict']
    n_features = [30]
    max_depths = [6, 8, 10, 12]
    n_estimatorss = [100]
    n_jobss = [cpu_count()]

    start = time()
    results = bench(n_obs, n_features, max_depths, n_estimatorss, n_jobss,
                    methods, repeat=repeat, verbose=verbose)
    end = time()

    results_df = pandas.DataFrame(results)
    print("Total time = %0.3f sec cpu=%d\n" % (end - start, cpu_count()))

    # plot the results
    return results_df


name = "plot_random_forest_reg"
df = run_bench(verbose=True)
df.to_csv("%s.csv" % name, index=False)
df.to_excel("%s.xlsx" % name, index=False)
fig, ax = plot_rf_models(df)
fig.savefig("%s.png" % name)
plt.show()
RandomForestRegressor - ort depth 6 - 30 features  100 estimators 8 jobs, RandomForestRegressor - ort depth 8 - 30 features  100 estimators 8 jobs, RandomForestRegressor - ort depth 10 - 30 features  100 estimators 8 jobs, RandomForestRegressor - ort depth 12 - 30 features  100 estimators 8 jobs, RandomForestRegressor - mlprodict depth 6 - 30 features  100 estimators 8 jobs, RandomForestRegressor - mlprodict depth 8 - 30 features  100 estimators 8 jobs, RandomForestRegressor - mlprodict depth 10 - 30 features  100 estimators 8 jobs, RandomForestRegressor - mlprodict depth 12 - 30 features  100 estimators 8 jobs, RandomForestRegressor - mlprodict2 depth 6 - 30 features  100 estimators 8 jobs, RandomForestRegressor - mlprodict2 depth 8 - 30 features  100 estimators 8 jobs, RandomForestRegressor - mlprodict2 depth 10 - 30 features  100 estimators 8 jobs, RandomForestRegressor - mlprodict2 depth 12 - 30 features  100 estimators 8 jobs, RandomForestRegressor - mlprodict3 depth 6 - 30 features  100 estimators 8 jobs, RandomForestRegressor - mlprodict3 depth 8 - 30 features  100 estimators 8 jobs, RandomForestRegressor - mlprodict3 depth 10 - 30 features  100 estimators 8 jobs, RandomForestRegressor - mlprodict3 depth 12 - 30 features  100 estimators 8 jobs

Out:

bench 1 : {'n_obs': 1, 'nfeat': 30, 'max_depth': 6, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.05573259542385737, 'time_ort': 8.528462300697963e-05, 'time_mlprodict': 0.008581353144513236, 'time_mlprodict2': 0.00873697487016519, 'time_mlprodict3': 0.008161399823923906}
bench 2 : {'n_obs': 10, 'nfeat': 30, 'max_depth': 6, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.055635006373955145, 'time_ort': 0.0004133107140660286, 'time_mlprodict': 0.0003797211166885164, 'time_mlprodict2': 0.0002635866403579712, 'time_mlprodict3': 0.015889178030192852}
bench 3 : {'n_obs': 100, 'nfeat': 30, 'max_depth': 6, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.059637165025753135, 'time_ort': 0.00045464889091603896, 'time_mlprodict': 0.008950498941190103, 'time_mlprodict2': 0.008859267567887026, 'time_mlprodict3': 0.005013359162737341}
bench 4 : {'n_obs': 1000, 'nfeat': 30, 'max_depth': 6, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.08179497647170837, 'time_ort': 0.0021723667876078533, 'time_mlprodict': 0.015863266988442495, 'time_mlprodict2': 0.013377069710538937, 'time_mlprodict3': 0.018311162407581624}
bench 5 : {'n_obs': 10000, 'nfeat': 30, 'max_depth': 6, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.10728115364909172, 'time_ort': 0.020656468719244002, 'time_mlprodict': 0.08076727837324142, 'time_mlprodict2': 0.02906704768538475, 'time_mlprodict3': 0.03656003475189209}
bench 6 : {'n_obs': 1, 'nfeat': 30, 'max_depth': 8, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.05526429855901944, 'time_ort': 9.753535452641939e-05, 'time_mlprodict': 0.008471277395361349, 'time_mlprodict2': 0.007965212589816043, 'time_mlprodict3': 0.008332478862844016}
bench 7 : {'n_obs': 10, 'nfeat': 30, 'max_depth': 8, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.05593731440603733, 'time_ort': 0.0008637115566266908, 'time_mlprodict': 0.0007608810232745276, 'time_mlprodict2': 0.0006883084360096189, 'time_mlprodict3': 0.016098013044231467}
bench 8 : {'n_obs': 100, 'nfeat': 30, 'max_depth': 8, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.06037295149529681, 'time_ort': 0.0008827326271463843, 'time_mlprodict': 0.010462815708973828, 'time_mlprodict2': 0.009450215408030678, 'time_mlprodict3': 0.017118924242608687}
bench 9 : {'n_obs': 1000, 'nfeat': 30, 'max_depth': 8, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.08232727775780055, 'time_ort': 0.003556580067827151, 'time_mlprodict': 0.022170927805396225, 'time_mlprodict2': 0.02182657902057354, 'time_mlprodict3': 0.019700800999999046}
bench 10 : {'n_obs': 10000, 'nfeat': 30, 'max_depth': 8, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.10702770799398423, 'time_ort': 0.030052572302520276, 'time_mlprodict': 0.13865551315248012, 'time_mlprodict2': 0.04309481009840965, 'time_mlprodict3': 0.04189431071281433}
bench 11 : {'n_obs': 1, 'nfeat': 30, 'max_depth': 10, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.05514783539662236, 'time_ort': 0.00010079517960548401, 'time_mlprodict': 0.00824174873138729, 'time_mlprodict2': 0.007789983365096544, 'time_mlprodict3': 0.008015583710450875}
bench 12 : {'n_obs': 10, 'nfeat': 30, 'max_depth': 10, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.05737313059055143, 'time_ort': 0.001000885230799516, 'time_mlprodict': 0.0009656976908445358, 'time_mlprodict2': 0.0011074329829878276, 'time_mlprodict3': 0.016058510065906577}
bench 13 : {'n_obs': 100, 'nfeat': 30, 'max_depth': 10, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.06914980423947176, 'time_ort': 0.0015494858225186666, 'time_mlprodict': 0.010059794286886851, 'time_mlprodict2': 0.010474308083454767, 'time_mlprodict3': 0.016736970469355582}
bench 14 : {'n_obs': 1000, 'nfeat': 30, 'max_depth': 10, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.0794465414320047, 'time_ort': 0.005977722601248668, 'time_mlprodict': 0.02757561779939211, 'time_mlprodict2': 0.02897698962344573, 'time_mlprodict3': 0.020387352372591313}
bench 15 : {'n_obs': 10000, 'nfeat': 30, 'max_depth': 10, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.10814805328845978, 'time_ort': 0.04370647370815277, 'time_mlprodict': 0.1910786084830761, 'time_mlprodict2': 0.07201163209974766, 'time_mlprodict3': 0.05348367244005203}
bench 16 : {'n_obs': 1, 'nfeat': 30, 'max_depth': 12, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.05718048423942593, 'time_ort': 0.00010644168489509159, 'time_mlprodict': 0.00390939549025562, 'time_mlprodict2': 0.00028803696235020954, 'time_mlprodict3': 9.197731398873859e-05}
bench 17 : {'n_obs': 10, 'nfeat': 30, 'max_depth': 12, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.05955639119972201, 'time_ort': 0.0011636773672174005, 'time_mlprodict': 0.0012538063832942177, 'time_mlprodict2': 0.0014957304605666328, 'time_mlprodict3': 0.00036632203880478353}
bench 18 : {'n_obs': 100, 'nfeat': 30, 'max_depth': 12, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.06897803073128064, 'time_ort': 0.002500701944033305, 'time_mlprodict': 0.006513191759586335, 'time_mlprodict2': 0.0030381677051385244, 'time_mlprodict3': 0.0020588673651218414}
bench 19 : {'n_obs': 1000, 'nfeat': 30, 'max_depth': 12, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.08466138364747167, 'time_ort': 0.010306961523989836, 'time_mlprodict': 0.02768038958311081, 'time_mlprodict2': 0.03632482405131062, 'time_mlprodict3': 0.018308731727302074}
bench 20 : {'n_obs': 10000, 'nfeat': 30, 'max_depth': 12, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.11488199751410219, 'time_ort': 0.0824148419002692, 'time_mlprodict': 0.23218230447835392, 'time_mlprodict2': 0.10829141487677892, 'time_mlprodict3': 0.05775153429971801}
Total time = 406.643 sec cpu=8

                            15          16         17        18        19
n_obs                        1          10        100      1000     10000
nfeat                       30          30         30        30        30
max_depth                   12          12         12        12        12
n_estimators               100         100        100       100       100
method                 predict     predict    predict   predict   predict
n_jobs                       8           8          8         8         8
time_skl               0.05718    0.059556   0.068978  0.084661  0.114882
time_ort              0.000106    0.001164   0.002501  0.010307  0.082415
time_mlprodict        0.003909    0.001254   0.006513   0.02768  0.232182
time_mlprodict2       0.000288    0.001496   0.003038  0.036325  0.108291
time_mlprodict3       0.000092    0.000366   0.002059  0.018309  0.057752
speedup_ort         537.200104    51.17947  27.583467     8.214  1.393948
speedup_mlprodict    14.626426   47.500469  10.590511  3.058533  0.494792
speedup_mlprodict2   198.51787   39.817596  22.703826  2.330676   1.06086
speedup_mlprodict3  621.680301  162.579329  33.502902  4.624099  1.989246

Total running time of the script: ( 7 minutes 3.140 seconds)

Gallery generated by Sphinx-Gallery