scikit-learn benchmarks#

scikit-learn’s team has started to develop a benchmark located here: scikit-learn_benchmarks I replicate here the steps I used to run and publish it from a local machine.


I followed the steps

git clone
cd asv_benchmarks

Run a benchmark#

I then ran a first benchmark with my current installation of Python.

asv run -b LinearRegression --no-pull

The tests do not store any result with option --option=<python_path>.

Publish a benchmark#

I then published it on a local directory.

asv publish -o html

Server to display the content#

I created a key

openssl req -x509 -newkey rsa:4096 -keyout key.pem -out cert.pem -days 365

And then a small application:

from starlette.routing import Router, Mount
from starlette.staticfiles import StaticFiles

app = Router(routes=[
    Mount('/', app=StaticFiles(directory='html'), name="html"),

And a server:

uvicorn webapp:app ---host <host> -port 8877 --ssl-keyfile=./key.pem --ssl-certfile=./cert.pem

Other benchmarks#

The first benchmark extends the official scikit-learn benchmarks available at scikit-learn_benchmarks. The results can be seen at this Scikit-Learn/ONNX benchmark with AirSpeedVelocity.

The second benchmark is produced using an automated way implemented in mlprodict. The sources are available at asv-skl2onnx and displayed at Prediction with scikit-learn and ONNX benchmark. A subset of these models is available at Prediction with scikit-learn and ONNX benchmark (SVM + Trees).

The last benchmark is a standalone benchmark only comparing onnxruntime and scikit-learn. The sources are available at scikit-onnx-benchmark and displayed at onnxruntime vs scikit-learn for comparison.

I also created two mini benchmark to get a sense of what the previous ones look like: mlprodict model of benchmark, mlprodict model applied to linear models.