ONNX Intermediate Representation (IR)

0.2.0 · active · verified Sat Apr 11

onnx-ir provides an efficient in-memory representation for ONNX graphs, allowing for programmatic creation, manipulation, and optimization of ONNX models in Python. It is currently at version 0.2.0 and has a frequent release cadence, often seeing multiple patch releases per month, indicating active development.

Warnings

breaking In `v0.1.9`, the `ir.Input` class was deprecated in favor of `ir.val`. Additionally, attribute signatures for nodes were simplified, requiring tuples for repeating attributes. Code using the old `ir.Input` or the previous attribute signature format will break.
Fix: Replace `ir.Input` with `ir.val`. Update node attribute definitions to use tuples for repeating attributes as specified in the updated API documentation.
gotcha When using the `Value.replace_all_uses_with` method (introduced in `v0.1.12`) with the `replace_graph_outputs` option, users are responsible for manually assigning the original output name to the replacement value if they wish to preserve the graph's signature. Failing to do so can lead to a modified graph signature.
Fix: After calling `Value.replace_all_uses_with` with `replace_graph_outputs=True`, ensure the replacement `Value` has its `.name` attribute set to the original output's name if graph signature preservation is desired.
gotcha The `v0.2.0` release introduced `sympy` as a new dependency for symbolic shape inferencing. This will increase the total install size and might slightly extend the installation time for new environments.
Fix: Ensure your environment is prepared for the additional `sympy` dependency. No code changes are required unless you were explicitly trying to avoid `sympy`.

Install

pip install onnx-ir Install stable version

Imports

Graph
```
from onnx_ir import Graph
```
Node
```
from onnx_ir import Node
```
Value
```
from onnx_ir import Value
```
Type
```
from onnx_ir import Type
```
Shape
```
from onnx_ir import Shape
```

TensorElementDataType

from onnx_ir import TensorElementDataType

Quickstart

This quickstart demonstrates how to create a simple ONNX graph representing an 'Add' operation using `onnx-ir`'s core components: `Value`, `Type`, `Shape`, `Node`, and `Graph`.

import onnx_ir as ir
import numpy as np

# Create input values with specific types and shapes
input_a = ir.Value("input_a", ir.Type(ir.Shape([2, 2]), ir.TensorElementDataType.FLOAT))
input_b = ir.Value("input_b", ir.Type(ir.Shape([2, 2]), ir.TensorElementDataType.FLOAT))

# Define the output value for the Add node
output_c = ir.Value("output_c") 

# Create an 'Add' node with inputs and outputs
add_node = ir.Node("Add", inputs=[input_a, input_b], outputs=[output_c])

# Assemble a graph from inputs, outputs, and nodes
graph = ir.Graph(
    [input_a, input_b], # Graph inputs
    [output_c],         # Graph outputs
    [add_node],         # Nodes in the graph
    "simple_add_graph"  # Name of the graph
)

print(f"Created graph: {graph.name}")
print(f"Graph has {len(graph.nodes)} node(s).")
print(f"Input 'input_a' shape: {input_a.type.shape.dims}")

view raw JSON →