OnnxSlim

You can install OnnxSlim using several methods depending on your needs:

Using Prebuilt Version

Install the latest stable version from PyPI:

pip install onnxslim

From Source

Install directly from the main branch on GitHub:

pip install git+https://github.com/inisis/OnnxSlim@main

From Local Repository

Clone the repository and install it locally:

git clone https://github.com/inisis/OnnxSlim && cd OnnxSlim/
pip install .

You can customize the output specifications of an ONNX model by specifying which output nodes to keep using the --outputs flag. This allows you to filter the model to only include the specific outputs required for your application, reducing unnecessary data overhead.

onnxslim yolov5m.onnx slim.onnx --outputs 591 739 443

You can use the Ultralytics CLI to export YOLO models to ONNX format. To apply OnnxSlim optimizations during the export, set the simplify flag to True. If set to False, the model will be exported without OnnxSlim streamlining.

With OnnxSlim (Recommended):

!yolo export format=onnx model=yolo11n.pt simplify=True

Without OnnxSlim:

!yolo export format=onnx model=yolo11n.pt simplify=False

When using the Ultralytics package, onnxslim is automatically installed during the model export process if you enable the simplification feature. You only need to install ultralytics to get started.

!pip install ultralytics

The onnxslim CLI tool is used to slim down and optimize ONNX models. It supports model modification (inputs/outputs/shapes), optimization settings (fusion patterns, size thresholds), and model checking (verifying output consistency via ONNX Runtime).

To use the CLI, you typically provide an input model and an optional output path. The tool will perform a series of optimizations and can optionally verify the results against the original model to ensure correctness.

# Example CLI usage (conceptual based on argument parser structure)
# onnxslim <input_model> <output_model> [options]
# onnxslim input.onnx output.onnx --dtype float16 --optimize

You can use the --inspect flag with the onnxslim CLI to perform a deep dive into an ONNX model's structure. This command provides detailed insights into:

• Input and output details
• Operator information
• Opset version
• Other model metadata and intricacies

onnxslim --inspect UNetModel-fp16.onnx

To see how CSE works in practice, you can run the provided demo. This involves two steps: first, generating an ONNX model that contains redundant subexpressions, and second, using onnxslim to optimize it.

1. Generate the demo model: Run the demo script to create the initial ONNX model (ln_cse.onnx).

2. Slim the model: Use the onnxslim CLI to process the model and produce an optimized version (slim.onnx).

# 1. Generate onnx model for demo
python cse_demo.py

# 2. Apply onnxslim to eliminate common subexpressions
onnxslim ln_cse.onnx slim.onnx

You can use onnxslim to adjust the dimensions of input tensors in an ONNX model. This is useful for ensuring compatibility or optimizing performance for specific neural network architectures. Use the --input_shapes flag to specify the new shapes for your input tensors using the format name:shape (e.g., name:batch,channels,height,width).

# Syntax: onnxslim <input_model> <output_model> --input_shapes <tensor_name>:<shape>
# Example: Changing the input shape of UNetModel-fp16.onnx
onnxslim UNetModel-fp16.onnx slim.onnx --input_shapes cc:1,1,768

To integrate OnnxSlim into your Python code, use the onnxslim.slim() function. This function takes an ONNX model object and returns a slimmed version of it.

import onnx
import onnxslim

# Load the original model
model = onnx.load("model.onnx")

# Slim the model
slimmed_model = onnxslim.slim(model)

# Save the slimmed model if successful
if slimmed_model:
    onnx.save(slimmed_model, "slimmed_model.onnx")