Note
Click here to download the full example code
Benchmark Random Forests, Tree Ensemble, Multi-Classification#
The script compares different implementations for the operator TreeEnsembleRegressor for a multi-regression. It replicates the benchmark Benchmark Random Forests, Tree Ensemble, (AoS and SoA) for multi-classification.
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 RandomForestClassifier
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())
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()
Implementations to benchmark#
def fcts_model(X, y, max_depth, n_estimators, n_jobs):
"RandomForestClassifier."
rf = RandomForestClassifier(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,
options={id(rf): {'zipmap': False}})
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[1]
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_proba(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_f = X_train.sum(axis=1) + eps
y_train = (y_train_f > 12).astype(numpy.int64)
y_train[y_train_f > 15] = 2
y_train[y_train_f < 10] = 3
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(f'{height:1.1f}x',
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[f"speedup_{engine}"] = dfr["time_skl"] / dfr[f"time_{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 = f"RandomForestClassifier - {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[f"speedup_{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[f"speedup_{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_cls_multi"
df = run_bench(verbose=True)
df.to_csv(f"{name}.csv", index=False)
df.to_excel(f"{name}.xlsx", index=False)
fig, ax = plot_rf_models(df)
fig.savefig(f"{name}.png")
plt.show()

bench 1 : {'n_obs': 1, 'nfeat': 30, 'max_depth': 6, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.04765076647024779, 'time_ort': 7.470637293798583e-05, 'time_mlprodict': 0.00248402861567835, 'time_mlprodict2': 0.00010005799343898183, 'time_mlprodict3': 8.814433766972451e-05}
bench 2 : {'n_obs': 10, 'nfeat': 30, 'max_depth': 6, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.04812958252261437, 'time_ort': 0.000464777151743571, 'time_mlprodict': 0.0005142051903974442, 'time_mlprodict2': 0.0004033286656652178, 'time_mlprodict3': 0.00010933933247412954}
bench 3 : {'n_obs': 100, 'nfeat': 30, 'max_depth': 6, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.05467034694983771, 'time_ort': 0.0006469648919607463, 'time_mlprodict': 0.004160436997680287, 'time_mlprodict2': 0.0014160474835845985, 'time_mlprodict3': 0.0002504979431825249}
bench 4 : {'n_obs': 1000, 'nfeat': 30, 'max_depth': 6, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.09081739125152428, 'time_ort': 0.0023643614258617163, 'time_mlprodict': 0.012102204918240508, 'time_mlprodict2': 0.009387286244115481, 'time_mlprodict3': 0.0016603594995103776}
bench 5 : {'n_obs': 10000, 'nfeat': 30, 'max_depth': 6, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.12503014874528162, 'time_ort': 0.022492028743727133, 'time_mlprodict': 0.09427896048873663, 'time_mlprodict2': 0.08740275274612941, 'time_mlprodict3': 0.015035680873552337}
bench 6 : {'n_obs': 1, 'nfeat': 30, 'max_depth': 8, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.04778254699582855, 'time_ort': 0.000332275139433997, 'time_mlprodict': 0.0031919755773352726, 'time_mlprodict2': 0.00013669332977206934, 'time_mlprodict3': 8.068342382709186e-05}
bench 7 : {'n_obs': 10, 'nfeat': 30, 'max_depth': 8, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.048955388189781276, 'time_ort': 0.0007574107564453568, 'time_mlprodict': 0.0008130259035776058, 'time_mlprodict2': 0.000798524623470647, 'time_mlprodict3': 0.0006977413404023363}
bench 8 : {'n_obs': 100, 'nfeat': 30, 'max_depth': 8, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.05498143637209738, 'time_ort': 0.0009036658968972532, 'time_mlprodict': 0.004169576360206855, 'time_mlprodict2': 0.0015404967370590097, 'time_mlprodict3': 0.0005466243693310963}
bench 9 : {'n_obs': 1000, 'nfeat': 30, 'max_depth': 8, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.0923984640023925, 'time_ort': 0.003965903539210558, 'time_mlprodict': 0.016983262623067607, 'time_mlprodict2': 0.013369255466386676, 'time_mlprodict3': 0.0022972208151424475}
bench 10 : {'n_obs': 10000, 'nfeat': 30, 'max_depth': 8, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.12652424213592894, 'time_ort': 0.03095408625085838, 'time_mlprodict': 0.1365099489921704, 'time_mlprodict2': 0.12974844872951508, 'time_mlprodict3': 0.019627700879937038}
bench 11 : {'n_obs': 1, 'nfeat': 30, 'max_depth': 10, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.047797711566090584, 'time_ort': 0.00025866161810145493, 'time_mlprodict': 0.00301460976091524, 'time_mlprodict2': 0.00018026481293851422, 'time_mlprodict3': 9.063099111829485e-05}
bench 12 : {'n_obs': 10, 'nfeat': 30, 'max_depth': 10, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.05004141420358792, 'time_ort': 0.0009473536978475749, 'time_mlprodict': 0.0010013921419158578, 'time_mlprodict2': 0.0011273278505541384, 'time_mlprodict3': 0.0003045463585294783}
bench 13 : {'n_obs': 100, 'nfeat': 30, 'max_depth': 10, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.0641665641887812, 'time_ort': 0.0014383788220584393, 'time_mlprodict': 0.0055228866840479895, 'time_mlprodict2': 0.002110758810886182, 'time_mlprodict3': 0.0008604667527833953}
bench 14 : {'n_obs': 1000, 'nfeat': 30, 'max_depth': 10, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.09443969372659922, 'time_ort': 0.005848687734793533, 'time_mlprodict': 0.019459723367948423, 'time_mlprodict2': 0.018619762098586016, 'time_mlprodict3': 0.0035390679047188974}
bench 15 : {'n_obs': 10000, 'nfeat': 30, 'max_depth': 10, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.13157362985657528, 'time_ort': 0.043568309629336, 'time_mlprodict': 0.15533498051809147, 'time_mlprodict2': 0.17656907974742353, 'time_mlprodict3': 0.026038672134745866}
bench 16 : {'n_obs': 1, 'nfeat': 30, 'max_depth': 12, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.04808341723955458, 'time_ort': 9.259476237708615e-05, 'time_mlprodict': 0.0026264214289507697, 'time_mlprodict2': 0.00023082571680701914, 'time_mlprodict3': 9.850805093135153e-05}
bench 17 : {'n_obs': 10, 'nfeat': 30, 'max_depth': 12, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.051831495692022146, 'time_ort': 0.0011683639022521675, 'time_mlprodict': 0.001267771900165826, 'time_mlprodict2': 0.0015561394044198095, 'time_mlprodict3': 0.001023954397533089}
bench 18 : {'n_obs': 100, 'nfeat': 30, 'max_depth': 12, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.06584856180415954, 'time_ort': 0.0021204155636951327, 'time_mlprodict': 0.00532802494126372, 'time_mlprodict2': 0.002713993191719055, 'time_mlprodict3': 0.001159551742603071}
bench 19 : {'n_obs': 1000, 'nfeat': 30, 'max_depth': 12, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.09783531227995726, 'time_ort': 0.009001969634978608, 'time_mlprodict': 0.022340936374596575, 'time_mlprodict2': 0.024623551829294724, 'time_mlprodict3': 0.005195507187057625}
bench 20 : {'n_obs': 10000, 'nfeat': 30, 'max_depth': 12, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.1335054574883543, 'time_ort': 0.07088223350001499, 'time_mlprodict': 0.1967973895079922, 'time_mlprodict2': 0.24313312163576484, 'time_mlprodict3': 0.040155497379601}
Total time = 213.657 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 0.048083 0.051831 0.065849 0.097835 0.133505
time_ort 0.000093 0.001168 0.00212 0.009002 0.070882
time_mlprodict 0.002626 0.001268 0.005328 0.022341 0.196797
time_mlprodict2 0.000231 0.001556 0.002714 0.024624 0.243133
time_mlprodict3 0.000099 0.001024 0.00116 0.005196 0.040155
speedup_ort 519.288737 44.362459 31.054555 10.868212 1.883483
speedup_mlprodict 18.307579 40.883928 12.358906 4.379195 0.67839
speedup_mlprodict2 208.310486 33.307746 24.262611 3.973241 0.549104
speedup_mlprodict3 488.116624 50.618949 56.787946 18.830753 3.324712
somewhere/workspace/mlprodict/mlprodict_UT_39_std/_doc/examples/plot_opml_random_forest_cls_multi.py:299: MatplotlibDeprecationWarning: The label function was deprecated in Matplotlib 3.1 and will be removed in 3.8. Use Tick.label1 instead.
tick.label.set_fontsize(8)
somewhere/workspace/mlprodict/mlprodict_UT_39_std/_doc/examples/plot_opml_random_forest_cls_multi.py:301: MatplotlibDeprecationWarning: The label function was deprecated in Matplotlib 3.1 and will be removed in 3.8. Use Tick.label1 instead.
tick.label.set_fontsize(8)
Total running time of the script: ( 3 minutes 47.792 seconds)