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': 0.046824001273224974, 'time_ort': 8.112872654402798e-05, 'time_mlprodict': 0.00659822559365156, 'time_mlprodict2': 0.00011165159098296003, 'time_mlprodict3': 0.012211281679232012}
bench 2 : {'n_obs': 10, 'nfeat': 30, 'max_depth': 6, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.047844640810840895, 'time_ort': 0.00043181847736594223, 'time_mlprodict': 0.0004922768622193308, 'time_mlprodict2': 0.0003770684790132301, 'time_mlprodict3': 0.01223254852396037}
bench 3 : {'n_obs': 100, 'nfeat': 30, 'max_depth': 6, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.054239802264706476, 'time_ort': 0.0019677445280218593, 'time_mlprodict': 0.0073170982068404555, 'time_mlprodict2': 0.007085977949349112, 'time_mlprodict3': 0.012463190739876345}
bench 4 : {'n_obs': 1000, 'nfeat': 30, 'max_depth': 6, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.08971791425331806, 'time_ort': 0.002335551088132585, 'time_mlprodict': 0.014738206499411413, 'time_mlprodict2': 0.01276848916313611, 'time_mlprodict3': 0.01461859032860957}
bench 5 : {'n_obs': 10000, 'nfeat': 30, 'max_depth': 6, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.12816017463046592, 'time_ort': 0.022938828740734607, 'time_mlprodict': 0.13100947038037702, 'time_mlprodict2': 0.1356982148718089, 'time_mlprodict3': 0.03585509386903141}
bench 6 : {'n_obs': 1, 'nfeat': 30, 'max_depth': 8, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.04706538513577967, 'time_ort': 8.58868165364997e-05, 'time_mlprodict': 0.006275462997357615, 'time_mlprodict2': 0.00013614131603389978, 'time_mlprodict3': 0.01229259363820099}
bench 7 : {'n_obs': 10, 'nfeat': 30, 'max_depth': 8, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.04836662452379685, 'time_ort': 0.0007203845251795082, 'time_mlprodict': 0.0007675139987397762, 'time_mlprodict2': 0.0007580103341578727, 'time_mlprodict3': 0.012285356049514598}
bench 8 : {'n_obs': 100, 'nfeat': 30, 'max_depth': 8, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.054783546102331264, 'time_ort': 0.0011283658432627195, 'time_mlprodict': 0.008060571100366743, 'time_mlprodict2': 0.005916006207515143, 'time_mlprodict3': 0.01293493589190276}
bench 9 : {'n_obs': 1000, 'nfeat': 30, 'max_depth': 8, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.09015101817203686, 'time_ort': 0.0035813540841142335, 'time_mlprodict': 0.022949589911149815, 'time_mlprodict2': 0.022991571992558118, 'time_mlprodict3': 0.015744912835846964}
bench 10 : {'n_obs': 10000, 'nfeat': 30, 'max_depth': 8, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.12918856274336576, 'time_ort': 0.02982376936415676, 'time_mlprodict': 0.17005822800274473, 'time_mlprodict2': 0.17033570000785403, 'time_mlprodict3': 0.03915524261537939}
bench 11 : {'n_obs': 1, 'nfeat': 30, 'max_depth': 10, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.046970076091714545, 'time_ort': 7.950827288864689e-05, 'time_mlprodict': 0.006540170272769915, 'time_mlprodict2': 0.00017123731826855376, 'time_mlprodict3': 0.003804267410569909}
bench 12 : {'n_obs': 10, 'nfeat': 30, 'max_depth': 10, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.04925348875778062, 'time_ort': 0.0009040169506555511, 'time_mlprodict': 0.0009855637160528982, 'time_mlprodict2': 0.001074648523215382, 'time_mlprodict3': 0.012479823570521105}
bench 13 : {'n_obs': 100, 'nfeat': 30, 'max_depth': 10, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.06369584306230536, 'time_ort': 0.0017485551870777272, 'time_mlprodict': 0.00862945475091692, 'time_mlprodict2': 0.008667226815305185, 'time_mlprodict3': 0.006903725559823215}
bench 14 : {'n_obs': 1000, 'nfeat': 30, 'max_depth': 10, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.09190410490952093, 'time_ort': 0.0057887961791658945, 'time_mlprodict': 0.025684633281674574, 'time_mlprodict2': 0.028402211737226356, 'time_mlprodict3': 0.017606272457421503}
bench 15 : {'n_obs': 10000, 'nfeat': 30, 'max_depth': 10, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.13303388337953947, 'time_ort': 0.042347337381215766, 'time_mlprodict': 0.20046561888011638, 'time_mlprodict2': 0.22357457838370465, 'time_mlprodict3': 0.04729508174932562}
bench 16 : {'n_obs': 1, 'nfeat': 30, 'max_depth': 12, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.04843824328522065, 'time_ort': 0.0001065022356453396, 'time_mlprodict': 0.0063315499075023195, 'time_mlprodict2': 0.00020421790291688273, 'time_mlprodict3': 0.012155406432048906}
bench 17 : {'n_obs': 10, 'nfeat': 30, 'max_depth': 12, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.051387542102020234, 'time_ort': 0.0010617911524605007, 'time_mlprodict': 0.001187512301839888, 'time_mlprodict2': 0.001325335947331041, 'time_mlprodict3': 0.012480162153951823}
bench 18 : {'n_obs': 100, 'nfeat': 30, 'max_depth': 12, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.06676212206172447, 'time_ort': 0.0018776941346004606, 'time_mlprodict': 0.008858928534512719, 'time_mlprodict2': 0.009235445402252178, 'time_mlprodict3': 0.013766374338107805}
bench 19 : {'n_obs': 1000, 'nfeat': 30, 'max_depth': 12, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.0972531319130212, 'time_ort': 0.008801296184008772, 'time_mlprodict': 0.030396176819604905, 'time_mlprodict2': 0.033091345547952435, 'time_mlprodict3': 0.02090568490199406}
bench 20 : {'n_obs': 10000, 'nfeat': 30, 'max_depth': 12, 'n_estimators': 100, 'method': 'predict', 'n_jobs': 8, 'time_skl': 0.14401230756526015, 'time_ort': 0.06827092500004385, 'time_mlprodict': 0.24750996299553663, 'time_mlprodict2': 0.2744046502879688, 'time_mlprodict3': 0.06602232657106859}
Total time = 221.675 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 0.048438 0.051388 0.066762 0.097253 0.144012
time_ort 0.000107 0.001062 0.001878 0.008801 0.068271
time_mlprodict 0.006332 0.001188 0.008859 0.030396 0.24751
time_mlprodict2 0.000204 0.001325 0.009235 0.033091 0.274405
time_mlprodict3 0.012155 0.01248 0.013766 0.020906 0.066022
speedup_ort 454.809638 48.397034 35.555377 11.049865 2.109424
speedup_mlprodict 7.650298 43.273271 7.53614 3.199519 0.581844
speedup_mlprodict2 237.189015 38.77322 7.228901 2.93893 0.524817
speedup_mlprodict3 3.984914 4.117538 4.849652 4.651995 2.181267
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 53.542 seconds)