ONNX GraphSurgeon

Table of Contents

Introduction
Installation
Examples
Understanding The Basics
Advanced
- Working With Models With External Data

Introduction

ONNX GraphSurgeon is a Python library that allows you to create and modify ONNX models.

If you prefer a GUI, try Nsight DL Designer, which integrates ONNX GraphSurgeon and provides various extra features.

Installation

Using Prebuilt Wheels

python3 -m pip install onnx_graphsurgeon --extra-index-url https://pypi.ngc.nvidia.com

Building From Source

Using Make Targets

make install

Building Manually

Build a wheel:

make build

Install the wheel manually from outside the repository:

python3 -m pip install onnx-graphsurgeon/dist/onnx_graphsurgeon-*-py2.py3-none-any.whl

Examples

The examples directory contains several examples of common use-cases of ONNX GraphSurgeon.

The visualizations provided were generated using Netron.

Understanding The Basics

ONNX GraphSurgeon is composed of three major components: Importers, the IR, and Exporters.

Importers

Importers are used to import a graph into the ONNX GraphSurgeon IR. The importer interface is defined in base_importer.py.

ONNX GraphSurgeon also provides high-level importer APIs for ease of use:

graph = gs.import_onnx(onnx.load("model.onnx"))

IR

The Intermediate Representation (IR) is where all modifications to the graph are made. It can also be used to create new graphs from scratch. The IR involves three components: Tensors, Nodes, and Graphs.

Nearly all of the member variables of each component can be freely modified. For details on the various attributes of these classes, you can view the help output using help(<class_or_instance>) in an interactive shell, or using print(help(<class_or_instance>)) in a script, where <class_or_instance> is an ONNX GraphSurgeon type, or an instance of that type.

Tensor

Tensors are divided into two subclasses: Variable and Constant.

A Constant is a tensor whose values are known upfront, and can be retrieved as a NumPy array and modified. Note: The values property of a Constant is loaded on-demand. If the property is not accessed, the values will not be loaded as a NumPy array.
A Variable is a tensor whose values are unknown until inference-time, but may contain information about data type and shape.

The inputs and outputs of Tensors are always Nodes.

An example constant tensor from ResNet50:

>>> print(tensor)
Constant (gpu_0/res_conv1_bn_s_0)
[0.85369843 1.1515082  0.9152944  0.9577646  1.0663182  0.55629414
 1.2009839  1.1912311  2.2619808  0.62263143 1.1149117  1.4921428
 0.89566356 1.0358194  1.431092   1.5360111  1.25086    0.8706703
 1.2564877  0.8524589  0.9436758  0.7507614  0.8945271  0.93587166
 1.8422242  3.0609846  1.3124607  1.2158023  1.3937513  0.7857263
 0.8928106  1.3042281  1.0153942  0.89356416 1.0052011  1.2964457
 1.1117343  1.0669073  0.91343874 0.92906713 1.0465593  1.1261675
 1.4551278  1.8252873  1.9678202  1.1031747  2.3236883  0.8831993
 1.1133649  1.1654979  1.2705412  2.5578163  0.9504889  1.0441847
 1.0620039  0.92997414 1.2119316  1.3101407  0.7091761  0.99814713
 1.3404484  0.96389204 1.3435135  0.9236031 ]

An example variable tensor from ResNet50:

>>> print(tensor)
Variable (gpu_0/data_0): (shape=[1, 3, 224, 224], dtype=float32)

Node

A Node defines an operation in the graph. A node may specify attributes; attribute values can be any Python primitive types, as well as ONNX GraphSurgeon Graphs or Tensors

The inputs and outputs of Nodes are always Tensors

An example ReLU node from ResNet50:

>>> print(node)
 (Relu)
    Inputs: [Tensor (gpu_0/res_conv1_bn_1)]
    Outputs: [Tensor (gpu_0/res_conv1_bn_2)]

In this case, the node has no attributes. Otherwise, attributes are displayed as an OrderedDict.

A Note On Modifying Inputs And Outputs

The inputs/outputs members of nodes and tensors have special logic that will update the inputs/outputs of all affected nodes/tensors when you make a change. This means, for example, that you do not need to update the inputs of a Node when you make a change to the outputs of its input tensor.

Consider the following node:

>>> print(node)
 (Relu).
    Inputs: [Tensor (gpu_0/res_conv1_bn_1)]
    Outputs: [Tensor (gpu_0/res_conv1_bn_2)]

The input tensor can be accessed like so:

>>> tensor = node.inputs[0]
>>> print(tensor)
Tensor (gpu_0/res_conv1_bn_1)
>>> print(tensor.outputs)
[ (Relu).
	Inputs: [Tensor (gpu_0/res_conv1_bn_1)]
	Outputs: [Tensor (gpu_0/res_conv1_bn_2)]

If we remove the node from the outputs of the tensor, this is reflected in the node inputs as well:

>>> del tensor.outputs[0]
>>> print(tensor.outputs)
[]
>>> print(node)
 (Relu).
    Inputs: []
    Outputs: [Tensor (gpu_0/res_conv1_bn_2)]

Graph

A Graph contains zero or more Nodes and input/output Tensors.

Intermediate tensors are not explicitly tracked, but are instead retrieved from the nodes contained within the graph.

The Graph class exposes several functions. A small subset is listed here:

cleanup(): Removes unused nodes and tensors in the graph
toposort(): Topologically sorts the graph.
tensors(): Returns a Dict[str, Tensor] mapping tensor names to tensors, by walking over all the tensors in the graph. This is an O(N) operation, and so may be slow for large graphs.

To see the full Graph API, you can see help(onnx_graphsurgeon.Graph) in an interactive Python shell.

Exporters

Exporters are used to export the ONNX GraphSurgeon IR to ONNX or other types of graphs. The exporter interface is defined in base_exporter.py.

ONNX GraphSurgeon also provides high-level exporter APIs for ease of use:

onnx.save(gs.export_onnx(graph), "model.onnx")

Advanced

Working With Models With External Data

Using models with externally stored data with ONNX-GraphSurgeon is almost the same as working with ONNX models without external data. Refer to the official ONNX documentation for details on how to load such models. To import the model into ONNX-GraphSurgeon, you can use the import_onnx function as normal.

During export, you just need to take one additional step:

Export the model from ONNX-GraphSurgeon as normal:
```
model = gs.export_onnx(graph)
```
Update the model so that it writes its data to the external location. If the location is not specified, it defaults to the same directory as the ONNX model:
```
from onnx.external_data_helper import convert_model_to_external_data

convert_model_to_external_data(model, location="model.data")
```
Then you can save the model as usual:
```
onnx.save(model, "model.onnx")
```