Store arrays in one onnx graph

Once a model is converted it can be useful to store an array as a constant in the graph an retrieve it through an output. This allows the user to store training parameters or other informations like a vocabulary. Last sections shows how to remove an output or to promote an intermediate result to an output.

Train and convert a model

We download one model from the ONNX Zoo but the model could be trained and produced by another converter library.

import pprint
import numpy
from onnx import load
from onnxruntime import InferenceSession
from sklearn.datasets import load_iris
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split
from skl2onnx import to_onnx
from skl2onnx.helpers.onnx_helper import (
    add_output_initializer,
    select_model_inputs_outputs,
)


data = load_iris()
X, y = data.data.astype(numpy.float32), data.target
X_train, X_test, y_train, y_test = train_test_split(X, y)
model = LogisticRegression(penalty="elasticnet", C=2.0, solver="saga", l1_ratio=0.5)
model.fit(X_train, y_train)

onx = to_onnx(model, X_train[:1], target_opset=12, options={"zipmap": False})

/home/xadupre/vv/this312/lib/python3.12/site-packages/sklearn/linear_model/_sag.py:348: ConvergenceWarning: The max_iter was reached which means the coef_ did not converge
  warnings.warn(

Add training parameter

new_onx = add_output_initializer(
    onx, ["C", "l1_ratio"], [numpy.array([model.C]), numpy.array([model.l1_ratio])]
)

Inference

sess = InferenceSession(new_onx.SerializeToString(), providers=["CPUExecutionProvider"])
print("output names:", [o.name for o in sess.get_outputs()])
res = sess.run(None, {"X": X_test[:2]})
print("outputs")
pprint.pprint(res)

output names: ['label', 'probabilities', 'C', 'l1_ratio']
outputs
[array([0, 2], dtype=int64),
 array([[9.9722755e-01, 2.7724325e-03, 3.5656236e-10],
       [1.3919589e-04, 7.1695939e-02, 9.2816484e-01]], dtype=float32),
 array([2.]),
 array([0.5])]

The major draw back of this solution is increase the prediction time as onnxruntime copies the constants for every prediction. It is possible either to store those constant in a separate ONNX graph or to removes them.

Select outputs

Next function removes unneeded outputs from a model, not only the constants. Next model only keeps the probabilities.

simple_onx = select_model_inputs_outputs(new_onx, ["probabilities"])

sess = InferenceSession(
    simple_onx.SerializeToString(), providers=["CPUExecutionProvider"]
)
print("output names:", [o.name for o in sess.get_outputs()])
res = sess.run(None, {"X": X_test[:2]})
print("outputs")
pprint.pprint(res)

# Function *select_model_inputs_outputs* add also promote an intermediate
# result to an output.
#

output names: ['probabilities']
outputs
[array([[9.9722755e-01, 2.7724325e-03, 3.5656236e-10],
       [1.3919589e-04, 7.1695939e-02, 9.2816484e-01]], dtype=float32)]

This example only uses ONNX graph in memory and never saves or loads a model. This can be done by using the following snippets of code.

Save a model

with open("simplified_model.onnx", "wb") as f:
    f.write(simple_onx.SerializeToString())

Load a model

model = load("simplified_model.onnx")

sess = InferenceSession(model.SerializeToString(), providers=["CPUExecutionProvider"])
print("output names:", [o.name for o in sess.get_outputs()])
res = sess.run(None, {"X": X_test[:2]})
print("outputs")
pprint.pprint(res)

output names: ['probabilities']
outputs
[array([[9.9722755e-01, 2.7724325e-03, 3.5656236e-10],
       [1.3919589e-04, 7.1695939e-02, 9.2816484e-01]], dtype=float32)]