Spox

0.17.1 verified Sat May 09 auth: no python

A framework for constructing ONNX computational graphs programmatically. Current version 0.17.1, supports Python >=3.10, and follows a monthly/quarterly release cadence. Spox provides a high-level API to build, inline, and manipulate ONNX models with type-safe operators and shape inference.

pip install spox

Common errors

error AttributeError: module 'spox' has no attribute 'make_model' ↓

cause The function is named make_onnx_model, not make_model.

fix

Use from spox import make_onnx_model or correct the function name.

error ValueError: Failed to infer shape for op MatMul ↓

cause Missing or mismatched shape information for an input or output tensor.

fix

Ensure all argument tensors have correct shapes; use spox.Tensor with dimension symbols or None for unknown.

error spox.exceptions.SpoxValueError: Operator 'GroupNormalization' not found in opset 'ai.onnx.v18' ↓

cause group_normalization was removed from v18, v19, v20 in spox 0.14.0.

fix

Use later opset version (v21+) or alternate implementation.

Warnings

breaking Removal of group_normalization constructor: spox.opset.ai.onnx.v18.group_normalization and its re-exports in v19, v20 removed in 0.14.0 due to upstream deprecation. ↓

fix Use group_normalization from a later opset version (v21+) or implement manually.

deprecated Importing spox._future.operator_overloading triggers DeprecationWarning since 0.13.0. Consider using ndonnx as an alternative. ↓

fix Replace usage with ndonnx.

gotcha Python version requirement: Spox 0.17.0+ requires Python >=3.10. Older versions may support down to 3.8. ↓

fix Ensure Python version >=3.10, or pin spox<0.17.0 for 3.9 compat.

Imports

spox
wrong
```
import spox as _
```
correct
```
import spox
```
No common wrong import; but avoid importing private modules like spox._future directly.
make_onnx_model
wrong
```
from spox import make_model
```
correct
```
from spox import build, make_onnx_model
```
The function is named make_onnx_model, not make_model.
onnx opset
```
from spox.opset.ai.onnx import v17
```
Use specific opset versions like v17, v18, etc., not generic 'onnx'.

Quickstart

Construct an ONNX model for Y = A*X + B using Spox.

import spox
from spox import build, make_onnx_model
from spox.opset.ai.onnx import v17 as op

# Build a simple model: Y = A * X + B
def build_linear_model():
    A = spox.argument(spox.Tensor(spox.Float, (3, 3)))
    X = spox.argument(spox.Tensor(spox.Float, (3, None)))
    B = spox.argument(spox.Tensor(spox.Float, (1, None)))
    Y = op.add(op.matmul(A, X), B)
    return build(inputs={'A': A, 'X': X, 'B': B}, outputs={'Y': Y})

model = build_linear_model()
print(make_onnx_model(model).SerializeToString()[:100])