Open Neural Network Exchange (ONNX) - Weekly Builds

Warnings

• breaking The `model hub integration` feature was removed in ONNX v1.21.0. If your workflow relied on this integration, you will need to update your code.
  Fix: Remove any dependencies on the defunct model hub integration. Consult ONNX documentation for alternative methods of accessing or managing models.
• gotcha The `ml_dtypes` package became a direct dependency for ONNX versions >= 1.19.0. Users upgrading to these versions or using tools like `onnxruntime` with older `onnx` installations might encounter `ModuleNotFoundError` if `ml_dtypes` is not explicitly installed. This is particularly relevant when working with extended data types like FLOAT8.
  Fix: Ensure `ml_dtypes` is installed in your environment: `pip install ml_dtypes>=0.5.0`.
• gotcha There's a common confusion between the `onnx` package (which defines the model format and provides utilities to build/manipulate ONNX graphs) and `onnxruntime` (which is the inference engine used to execute ONNX models efficiently). The `onnx-weekly` package only provides the `onnx` library.
  Fix: Understand the distinct roles: `onnx` for model definition/manipulation, `onnxruntime` for model execution. Install `onnxruntime` separately (`pip install onnxruntime`) for inference capabilities.
• gotcha ONNX models are versioned by IR version and operator set (Opset) versions. Breaking changes to the IR format or operator semantics require version increments. Ensure that the ONNX Runtime or target inference environment you are using supports the specific Opset version of your ONNX model to avoid compatibility issues.
  Fix: Always check the Opset version of your exported ONNX model (`model_def.opset_import`) and confirm it is supported by your chosen ONNX Runtime version. Use `onnx.version_converter` if necessary to convert models to different opset versions.
• gotcha ONNX models are serialized using Google's Protocol Buffers, which imposes a 2GB size limit on individual model files. Very large models may fail to serialize or load correctly.
  Fix: For models exceeding 2GB, consider splitting the model into smaller subgraphs, using external data fields (which store large tensors separately from the main protobuf file), or exploring alternative serialization methods if available.

Install

• pip install onnx-weekly Install latest weekly build

Imports

• onnx

  import onnx

• helper

  from onnx import helper

• checker

  from onnx import checker

• TensorProto

  from onnx import TensorProto

• parser

  import onnx.parser

  Used for parsing ONNX Script models.
• shape_inference

  import onnx.shape_inference

  Provides utilities for shape and type inference in ONNX graphs.

Quickstart

This quickstart demonstrates how to programmatically construct a simple ONNX model (a linear regression: Y = X * A + B) using the `onnx.helper` module, and then validate it using `onnx.checker`. This illustrates the core functionality of defining and manipulating ONNX graphs.

import onnx
from onnx import helper, checker, TensorProto

# Create a simple ONNX graph for Y = X * A + B
# Define inputs and outputs
X = helper.make_tensor_value_info('X', TensorProto.FLOAT, [None, 2])
A = helper.make_tensor_value_info('A', TensorProto.FLOAT, [2, 3])
B = helper.make_tensor_value_info('B', TensorProto.FLOAT, [3])
Y = helper.make_tensor_value_info('Y', TensorProto.FLOAT, [None, 3])

# Create nodes (operators)
# MatMul operation: C = X * A
node_matmul = helper.make_node(
    'MatMul',
    inputs=['X', 'A'],
    outputs=['C'],
)

# Add operation: Y = C + B
node_add = helper.make_node(
    'Add',
    inputs=['C', 'B'],
    outputs=['Y'],
)

# Create the graph
graph_def = helper.make_graph(
    [node_matmul, node_add], # Nodes in the graph
    'simple-linear-regression', # Graph name
    [X, A, B], # Graph inputs
    [Y], # Graph outputs
)

# Create the model
model_def = helper.make_model(graph_def, producer_name='onnx-example')

# Check the model for validity
try:
    checker.check_model(model_def)
    print("Model is valid!")
except checker.ValidationError as e:
    print(f"Model is invalid: {e}")

# Optionally, save the model
# onnx.save(model_def, "simple_model.onnx")