Generating a ModelProto#

This example demonstrates the use of onnxscript to define an ONNX model. onnxscript behaves like a compiler. It converts a script into an ONNX model.

First, we define the implementation of a square-loss function in onnxscript.

import numpy as np
import onnx
from onnxruntime import InferenceSession

from onnxscript import FLOAT, script
from onnxscript import opset15 as op


@script()
def square_loss(X: FLOAT["N", 1], Y: FLOAT["N", 1]) -> FLOAT[1, 1]:  # noqa: F821
    diff = X - Y
    return op.ReduceSum(diff * diff, keepdims=1)

We can convert it to a model (an ONNX ModelProto) as follows:

model = square_loss.to_model_proto()

Let’s see what the generated model looks like.

print(onnx.printer.to_text(model))
<
   ir_version: 8,
   opset_import: ["" : 15]
>
square_loss (float[N,1] X, float[N,1] Y) => (float[1,1] return_val) {
   diff = Sub (X, Y)
   tmp = Mul (diff, diff)
   return_val = ReduceSum <keepdims: int = 1> (tmp)
}

We can run shape-inference and type-check the model using the standard ONNX API.

model = onnx.shape_inference.infer_shapes(model)
onnx.checker.check_model(model)

And finally, we can use onnxruntime to compute the outputs based on this model, using the standard onnxruntime API.

sess = InferenceSession(model.SerializeToString(), providers=("CPUExecutionProvider",))

X = np.array([[0, 1, 2]], dtype=np.float32).T
Y = np.array([[0.1, 1.2, 2.3]], dtype=np.float32).T

got = sess.run(None, {"X": X, "Y": Y})
expected = ((X - Y) ** 2).sum()

print(expected, got)
0.13999999 [array([[0.13999999]], dtype=float32)]

Total running time of the script: (0 minutes 0.021 seconds)

Gallery generated by Sphinx-Gallery