https://github.com/mlomb/onnx2code

Convert ONNX models to plain C++ code (without dependencies)
https://github.com/mlomb/onnx2code

assembly cpp inference machine-learning onnx python

Last synced: 3 months ago
JSON representation

Convert ONNX models to plain C++ code (without dependencies)

• Host: GitHub
• URL: https://github.com/mlomb/onnx2code
• Owner: mlomb
• Created: 2022-09-16T06:22:50.000Z (over 2 years ago)
• Default Branch: main
• Last Pushed: 2023-03-27T23:27:37.000Z (almost 2 years ago)
• Last Synced: 2024-10-06T10:41:00.542Z (4 months ago)
• Topics: assembly, cpp, inference, machine-learning, onnx, python
• Language: Python
• Homepage:
• Size: 3.34 MB
• Stars: 17
• Watchers: 5
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# onnx2code

Generate plain C++ code for inference of ONNX models without dependencies

This project was made as an alternative to a final exam for the assignment "Computer Organization II". You can read the writeup in [docs/TP Final onnx2code.pdf](docs/TP%20Final%20onnx2code.pdf) (in Spanish).

## Model support

The following models have been tested and work as expected.

| Model | Size |

|---|---|

| [mnist](https://github.com/onnx/models/tree/main/vision/classification/mnist) | 26 KB |

| [Super_Resolution](https://github.com/onnx/models/tree/main/vision/super_resolution/sub_pixel_cnn_2016) | 240 KB |

| [squeezenet1.1](https://github.com/onnx/models/tree/main/vision/classification/squeezenet) | 9 MB |

| [emotion_ferplus](https://github.com/onnx/models/tree/main/vision/body_analysis/emotion_ferplus) | 34 MB |

| [inception-v2](https://github.com/onnx/models/tree/main/vision/classification/inception_and_googlenet/inception_v2) | 44 MB |

| [resnet50-caffe2-v1](https://github.com/onnx/models/tree/main/vision/classification/resnet) | 98 MB |

| [VGG 16 and VGG 16-bn](https://github.com/onnx/models/tree/main/vision/classification/vgg) | 527 MB |

| [VGG 19 and VGG 19-bn](https://github.com/onnx/models/tree/main/vision/classification/vgg) | 548 MB |

| [VGG 19-caffe2](https://github.com/onnx/models/tree/main/vision/classification/vgg) | 561 MB |

* Minimum ONNX opset version: **7**

* Quantized models are not supported

## Operator support

Only `float` data type is supported.

| Operator | Attribute support |

|---|---|

| [Add](https://github.com/onnx/onnx/blob/main/docs/Operators.md#Add), [Div](https://github.com/onnx/onnx/blob/main/docs/Operators.md#Div), [Mul](https://github.com/onnx/onnx/blob/main/docs/Operators.md#Mul), [Sub](https://github.com/onnx/onnx/blob/main/docs/Operators.md#Sub) | ✅ with broadcasting |

| [Concat](https://github.com/onnx/onnx/blob/main/docs/Operators.md#Concat) | ✅ with multiple inputs
✅ axis |

| [Conv](https://github.com/onnx/onnx/blob/main/docs/Operators.md#Conv) | ✅ bias
✅ stride
✅ padding (and `auto_pad`)
❌ dilations
❌ depthwise (group != 1) |

| [Sum](https://github.com/onnx/onnx/blob/main/docs/Operators.md#Sum) | ✅ with multiple inputs
❌ with broadcasting |

| [Relu](https://github.com/onnx/onnx/blob/main/docs/Operators.md#Relu), [Tanh](https://github.com/onnx/onnx/blob/main/docs/Operators.md#Tanh), [Sigmoid](https://github.com/onnx/onnx/blob/main/docs/Operators.md#Sigmoid),  [Clip](https://github.com/onnx/onnx/blob/main/docs/Operators.md#Clip) | ✅ |

| [Gemm](https://github.com/onnx/onnx/blob/main/docs/Operators.md#Gemm) | ✅ with bias
❌ transpose A
✅ tranpose B
❌ alpha != 1
❌ beta != 1 |

| [Identity](https://github.com/onnx/onnx/blob/main/docs/Operators.md#Identity) | ✅ |

| [MaxPool](https://github.com/onnx/onnx/blob/main/docs/Operators.md#MaxPool), [AveragePool](https://github.com/onnx/onnx/blob/main/docs/Operators.md#AveragePool) | ✅ stride
✅  padding (and `auto_pad`)
❌ dilations
❌ storage_order != 0
❌ count_include_pad != 0 |

| [Softmax](https://github.com/onnx/onnx/blob/main/docs/Operators.md#Softmax) | ✅ stride
✅ axis |

| [Transpose](https://github.com/onnx/onnx/blob/main/docs/Operators.md#Transpose) | ✅ perm |

## Setting up with Docker

We provide a ready to use [Docker image](https://hub.docker.com/r/mlomb/onnx2code):

```sh

docker run --rm -it -v $pwd/mnist.onnx:/app/input.onnx:ro -v $pwd/output:/app/output:rw mlomb/onnx2code:latest --variations=im2col,loop-tiling --checks=3

```

The command above will generate C++ code for the `mnist.onnx` model in the `output` folder.

## Setting up locally

### Prerequisites

* gcc (required if checking models)

* Python 3.10

* [pipenv](https://pypi.org/project/pipenv/)

Clone and install dependencies with `pipenv install`.

### Run

To generate code from an ONNX model, run the following command inside a pipenv shell:

```sh

python -m onnx2code --variation=im2col,loop-tiling mnist.onnx output_folder --checks=3

```