onnxifier 2.1.1

An Open Neural Network Exchange (ONNX) Optimization and Transformation Tool.

ONNXifier

English | 中文

A simple tool to convert any IR format to ONNX file.

Framework	Status
OpenVINO	✅
ONNXRuntime	✅
TensorRT	🚧
TensorRT-LLM	🚧

• ✅: well supported
• 🪛: partially supported
• 🚧: developing

Usage

1. Install from PyPI

pip install onnxifier

2. Convert IR using CLI

onnxify model.xml

usage: onnxify input_model.onnx [output_model.onnx]

onnxify command-line api

options:
  -h, --help            show this help message and exit
  --install-completion [{bash,pwsh}]
                        install shell completion for the specified shell and
                        exit.
  -a [ACTIVATE ...], --activate [ACTIVATE ...]
                        select passes to be activated, activate L1, L2 and L3
                        passes if not set.
  -r [REMOVE ...], --remove [REMOVE ...]
                        specify passes to be removed from activated passes.
  -n, --no-passes       do not run any optimizing passes, just convert the
                        model
  --print [PRINT]       print the name of all optimizing passes
  --format {protobuf,textproto,json,onnxtxt}
                        onnx file format
  -s, --infer-shapes    infer model shapes
  -c CONFIG_FILE, --config-file CONFIG_FILE
                        specify a json-format config file for passes
  -u, --uncheck         no checking output model
  --check               check optimized model with random inputs
  -d, --dry-run         only run passes without saving the output model
  --checker-backend {onnx,openvino,onnxruntime}
                        backend for accuracy checking, defaults to onnxruntime
  -v OPSET_VERSION, --opset-version OPSET_VERSION
                        target opset version, defaults to 20
  -vv [{DEBUG,INFO,WARNING,ERROR,CRITICAL}], --log-level [{DEBUG,INFO,WARNING,ERROR,CRITICAL}]
                        specify the level of log messages to be printed,
                        defaults to INFO
  -R, --recursive       recursively optimize nested functions
  --nodes [NODES ...]   specify a set of node names to apply passes only on
                        these nodes

To print pass information:

onnxify --print all
onnxify --print fuse_swish
onnxify --print l1

Shell Completion

onnxify provides tab-completion for Bash and PowerShell to help you quickly select pass names and options.

After installing onnxifier from PyPI, run the built-in installer once for your shell:

Bash

onnxify --install-completion bash
source ~/.bashrc

PowerShell

onnxify --install-completion pwsh
. $PROFILE

The installer is idempotent — running it again will not duplicate entries in your profile.

Once enabled, you can use Tab to complete pass names after -a / -r / --print, for example:

# Complete a single pass
onnxify model.onnx -a ins<TAB>
# → inspect_sparsity_ratio inspect_weights_distribution insert_conv_before_act_shave

# Complete multiple space-separated passes
onnxify model.onnx -a infer_shape fold_const<TAB>
# → fold_constant

# Complete comma-separated passes
onnxify model.onnx -a fuse_gelu,ins<TAB>
# → fuse_gelu,inspect_sparsity_ratio ...

# Complete --print arguments
onnxify --print l<TAB>
# → l1 l2 l3

TODO

• [OV] Add Loop support.
• [OV] Add NMS support.
• [OV] Add If support.
• [ONNX] Support to optimize If.

Contribute

1. pyright type checking

pip install -U pyright
pyright onnxifier

2. mypy type checking

pip install -U mypy
mypy onnxifier --disable-error-code=import-untyped --disable-error=override --disable-error=call-overload

3. pre-commit checking

pip install -U pre-commit
pre-commit run --all-files