mlprodict¶
Links: github, documentation, mlprodict, blog
mlprodict explores couple of ways to compute predictions faster than the library used to build the machine learned model, mostly scikit-learn which is optimized for training, which is equivalent to batch predictions. One way is to use ONNX. onnxruntime provides an efficient way to compute predictions. mlprodict implements a python/numpy runtime for ONNX which does not have any dependency on scikit-learn.
<<<
import numpy
from sklearn.linear_model import LinearRegression
from sklearn.datasets import load_iris
from mlprodict.onnxrt import OnnxInference, measure_relative_difference
from mlprodict.tools import get_ir_version_from_onnx
iris = load_iris()
X = iris.data[:, :2]
y = iris.target
lr = LinearRegression()
lr.fit(X, y)
# Predictions with scikit-learn.
expected = lr.predict(X[:5])
print(expected)
# Conversion into ONNX.
from mlprodict.onnx_conv import to_onnx
model_onnx = to_onnx(lr, X.astype(numpy.float32))
print("ONNX:", str(model_onnx)[:200] + "\n...")
# Predictions with onnxruntime
model_onnx.ir_version = get_ir_version_from_onnx()
oinf = OnnxInference(model_onnx, runtime='onnxruntime1')
ypred = oinf.run({'X': X[:5].astype(numpy.float32)})
print("ONNX output:", ypred)
# Measuring the maximum difference.
print("max abs diff:", measure_relative_difference(
expected, ypred['variable']))
>>>
[0.172 0.343 0.069 0.059 0.034]
ONNX: ir_version: 6
producer_name: "skl2onnx"
producer_version: "1.6.9996"
domain: "ai.onnx"
model_version: 0
doc_string: ""
graph {
node {
input: "X"
output: "variable"
name: "LinearRegressor
...
ONNX output: {'variable': array([[0.172],
[0.343],
[0.069],
[0.059],
[0.034]], dtype=float32)}
max abs diff: 6.303014714402957e-06
These predictions are obtained with the following ONNX graph.
That concludes the example with ONNX. A similar way was introduced before switching to epkg:ONNX. It is far less advanced but aims at producing a C file replicating the predictions.
<<<
from sklearn.linear_model import LogisticRegression
from sklearn.datasets import load_iris
iris = load_iris()
X = iris.data[:, :2]
y = iris.target
y[y == 2] = 1
lr = LogisticRegression()
lr.fit(X, y)
# Conversion into a graph.
from mlprodict.grammar_sklearn import sklearn2graph
gr = sklearn2graph(lr, output_names=['Prediction', 'Score'])
# Conversion into C
ccode = gr.export(lang='c')
# We print after a little bit of cleaning (remove all comments)
print("\n".join(_ for _ in ccode['code'].split("\n") if "//" not in _))
>>>
int LogisticRegression (float* pred, float* Features)
{
float pred0c0c00c0[2] = {(float)3.3882975578308105, (float)-3.164527654647827};
float* pred0c0c00c1 = Features;
float pred0c0c00;
adot_float_float(&pred0c0c00, pred0c0c00c0, pred0c0c00c1, 2);
float pred0c0c01 = (float)-8.323304176330566;
float pred0c0c0 = pred0c0c00 + pred0c0c01;
float pred0c0;
sign_float(&pred0c0, pred0c0c0);
float pred0[2];
concat_float_float(pred0, pred0c0, pred0c0c0);
memcpy(pred, pred0, 2*sizeof(float));
return 0;
}
Notebook ONNX visualization shows how to visualize an ONNX pipeline.
