ONNX Simplifier

0.6.2 · active · verified Mon Apr 13

ONNX Simplifier (onnxsim) is a Python library designed to reduce the complexity of ONNX models by inferring the computation graph and performing constant folding. This makes ONNX models more efficient for inference and deployment. It is actively maintained with frequent minor releases, currently at version 0.6.2.

Warnings

gotcha Building `onnxsim` from source (e.g., if pre-built wheels are unavailable for your specific platform/Python version) requires `cmake` and a C++ compiler. Installation may fail without these system-level dependencies.
Fix: Ensure `cmake` and a C++ compiler (like GCC or MSVC) are installed on your system before attempting `pip install onnxsim`.
gotcha Models exceeding a protobuf size limit (typically 2GB) can cause errors during loading or simplification due to limitations in the underlying protobuf library used by ONNX.
Fix: Consider splitting very large models into smaller subgraphs if possible, or investigate advanced techniques for handling large ONNX models outside the typical workflow.
gotcha ONNX models with graphs that are not topologically sorted may result in validation errors during simplification.
Fix: Ensure the input ONNX model adheres to the ONNX specification regarding graph topology. Tools like Netron can help visualize and debug model graphs.
deprecated The `--enable-fuse-bn` command-line argument for fusing batch normalization into convolutional layers is deprecated as this optimization is now enabled by default.
Fix: Remove the `--enable-fuse-bn` flag. If you need to skip this optimization, use `--skip-fuse-bn` instead.
gotcha When simplifying models with dynamic input shapes, you often need to explicitly provide the expected input shapes using the `--input-shape` argument in the CLI or `input_shapes` parameter in the `simplify` function.
Fix: Pass a dictionary mapping input names to their expected (or example) shapes, e.g., `simplify(model, input_shapes={'input_name': [1, 3, 224, 224]})` or `onnxsim input.onnx output.onnx --input-shape input_name:1,3,224,224`.

Install

pip install onnxsim Install latest stable version

Imports

simplify
```
from onnxsim import simplify
```

Quickstart

This quickstart demonstrates how to load an ONNX model, simplify it using `onnxsim.simplify`, and save the optimized model. A dummy PyTorch model is created and exported to ONNX for a complete runnable example.

import onnx
from onnxsim import simplify
import torch
import torch.nn as nn
import os

# Create a dummy ONNX model for demonstration
class SimpleNet(nn.Module):
    def __init__(self):
        super().__init__()
        self.fc = nn.Linear(10, 2)
    def forward(self, x):
        return self.fc(x)

model = SimpleNet()
dummy_input = torch.randn(1, 10)
input_model_path = "dummy_model.onnx"
output_model_path = "dummy_model_simplified.onnx"

torch.onnx.export(model, dummy_input, input_model_path,
                  input_names=['input'], output_names=['output'])

# Load your predefined ONNX model
onnx_model = onnx.load(input_model_path)

# Convert model
model_simp, check = simplify(onnx_model)

assert check, "Simplified ONNX model could not be validated"

# Save the simplified model
onnx.save(model_simp, output_model_path)
print(f"Model simplified and saved to {output_model_path}")

# Clean up dummy models
os.remove(input_model_path)
os.remove(output_model_path)

view raw JSON →