ONNX Command-Line Toolbox

Overview

ONNX Command Line Toolbox

Build and Test CodeQL Sanity Coverage

  • Aims to improve your experience of investigating ONNX models.
  • Use it like onnx infershape /path/to/model.onnx. (See the usage section for more.)

Installation

Recommand to install via GitHub repo for the latest functionality.

pip install git+https://github.com/jackwish/onnxcli.git

Two alternative ways are:

  1. Install via pypi package pip install onnxcli
  2. Download and add the code tree to your $PYTHONPATH. This is for development purpose since the command line is different.
    git clone https://github.com/jackwish/onnxcli.git
    export PYTHONPATH=$(pwd)/onnxcli:${PYTHONPATH}
    python onnxcli/cli/dispatcher.py <more args>
    

The onnx draw requires dot command (graphviz) to be avaiable on your machine - which can be installed by command as below on Ubuntu/Debian.

sudo apt install -y graphviz

Usage

Once installed, the onnx and onnxcli commands are avaiable on your machine. You can play with commands such as onnx infershape /path/to/model.onnx. The general format is onnx <sub command> <dedicated arguments ...>. The sub commands are as sections below.

Check the online help with onnx --help and onnx <subcmd> --help for latest usage.

infershape

onnx infershape performs shape inference of the ONNX model. It's an CLI wrapper of onnx.shape_inference. You will find it useful to generate shape information for the models that are extracted by onnx extract.

extract

onnx extract extracts the sub model that is determined by the names of the input and output tensor of the subgraph from the original model. It's a CLI wrapper of onnx.utils.extract_model (which I authorized in the ONNX repo).

inspect

onnx inspect gives you quick view of the information of the given model. It's inspired by the tf-onnx tool.

When working on deep learning, you may like to take a look at what's inside the model. Netron is powerful but doesn't provide fine-grain view.

With onnx inspect, you no longer need to scroll the Netron window to look for nodes or tensors. Instead, you can dump the node attributes and tensor values with a single command.

Click here to see a node example

$ onnx inspect ./assets/tests/conv.float32.onnx --node --indices 0 --detail

Inpect of model ./assets/tests/conv.float32.onnx Graph name: 9 Graph inputs: 1 Graph outputs: 1 Nodes in total: 1 ValueInfo in total: 2 Initializers in total: 2 Sparse Initializers in total: 0 Quantization in total: 0

Node information: Node "output": type "Conv", inputs "['input', 'Variable/read', 'Conv2D_bias']", outputs "['output']" attributes: [name: "dilations" ints: 1 ints: 1 type: INTS , name: "group" i: 1 type: INT , name: "kernel_shape" ints: 3 ints: 3 type: INTS , name: "pads" ints: 1 ints: 1 ints: 1 ints: 1 type: INTS , name: "strides" ints: 1 ints: 1 type: INTS ]

Click here to see a tensor example

$ onnx inspect ./assets/tests/conv.float32.onnx --tensor --names Conv2D_bias --detail

Inpect of model ./assets/tests/conv.float32.onnx Graph name: 9 Graph inputs: 1 Graph outputs: 1 Nodes in total: 1 ValueInfo in total: 2 Initializers in total: 2 Sparse Initializers in total: 0 Quantization in total: 0

Tensor information: Initializer "Conv2D_bias": type FLOAT, shape [16], float data: [0.4517577290534973, -0.014192663133144379, 0.2946248948574066, -0.9742919206619263, -1.2975586652755737, 0.7223454117774963, 0.7835700511932373, 1.7674627304077148, 1.7242872714996338, 1.1230682134628296, -0.2902531623840332, 0.2627834975719452, 1.0175092220306396, 0.5643373131752014, -0.8244842290878296, 1.2169424295425415]

draw

onnx draw draws the graph in dot, svg, png formats. It gives you quick view of the type and shape of the tensors that are fed to a specific node. You can view the model topology in image viewer of browser without waiting for the model to load, which I found is really helpful for large models.

If you are viewing svg in browser, you can even quick search for the nodes and tensors. Together with onnx inspect, it will be very efficient to understand the issue you are looking into.

The node are in ellipses and tensors are in rectangles where the rounded ones are initializers. The node type of the node and the data type and shape of the tenors are also rendered. Here is a Convolution node example.

conv

Contributing

Welcome to contribute new commands or enhance them. Let's make our life easier together.

The workflow is pretty simple:

  1. Starting with GitHub Codespace or clone locally.
  • make setup to config the dependencies (or pip install -r ./requirements.txt if you prefer).
  1. Create a new subcommand
  • Starting by copying and modifying infershape.
  • Register the command in the dispatcher
  • Create a new command line test
  • make test to build and test.
  • make check and make format to fix any code style issues.
  1. Try out, debug, commit, push, and open pull request.
  • The code has been protected by CI. You need to get a pass before merging.
  • Ask if any questions.

License

Apache License Version 2.0.

Comments
  • Some ONNX models don't list activation tensors in GraphProto.value_info

    Some ONNX models don't list activation tensors in GraphProto.value_info

    They should, but they don't. I am not sure why such models behave like this - they cannot pass the ONNX model checker.

    There should be something wrong with the exporter. I can try to figure out which exporter has such issues.

    For onnxcli, any functionality depending on walking GraphProto.value_info may not show the real model. This is not our defect, but the models'. To workaround, you can firstly run shape inference on the model, and the GraphProto.value_info listing issue will be fixed.

    onnx infershape /path/to/input/model /path/to/output/model
    
    documentation 
    opened by zhenhuaw-me 2
  • Integrate the onnx dumper

    Integrate the onnx dumper

    src: https://github.com/onnx/tensorflow-onnx/blob/master/tools/dump-onnx.py

    most of them need to be renamed.

    • [x] inspect to check the model
    • [x] dump dot has high priotiry
    • [ ] print to std if no file specified
    opened by zhenhuaw-me 0
  • Optimizer reports

    Optimizer reports "Unresolved value references" since v0.3.0

    Via pipeline https://github.com/zhenhuaw-me/onnxcli/actions/runs/3453474851/jobs/5764096907.

    A simple model works no issue till optimizer v0.2.7 (verified locally), but starts to fail with optimizer v0.3.0 (verified locally) and still fail with v0.3.2 (the pipeline).

    It's onnx optimize ./assets/tests/conv.float32.onnx optimized.onnx.

    opened by zhenhuaw-me 2
  • Overwrite weights (initializers) with fixed data or random data

    Overwrite weights (initializers) with fixed data or random data

    Bert series ONNX models are very large (x GB) thus not easy to share the real file. We can improve this process by overwriting the weights (initializers)

    • It can be fixed data (e.g. all 0.1 or other value specified), thus the model can be compressed.
    • After sharing, we can recover with numpy style random numbers.

    This can only be used as a sharing method, the generated model are not useful when evaluate accuracy.

    For better usage:

    • Annotation will be added when writing fixed data, thus when re-random we can detect automatically.
    • The tensors can be specified with names or size.
    • Only works for FP32/FP16.
    • 0 removed.
    enhancement 
    opened by zhenhuaw-me 0
  • [draw] show tensor information on the edges

    [draw] show tensor information on the edges

    We currently draw tensors as boxes and operators as circles.

    image

    The graph will be complex if large model. We draw the tensor information on the edges and keep only operators as nodes.

    enhancement 
    opened by zhenhuaw-me 0
  • [infershape] should be able to set tensor shapes - inputs and others

    [infershape] should be able to set tensor shapes - inputs and others

    infershape is not very useful if the input shapes are symbolics (dynamic shapes). If the user can set input shapes, it's more powerful:

    • If set to static shapes, the shape of the model will be known.
    • Even for symbolics, the user can update the input shapes.

    The setup should be optional, and can extend to all the tensors in the model (excluding shape op related).

    Interface should be something like below.

    onnx infershape path/to/input/model.onnx path/to/output/model.onnx --tensor-shape t1:[d0,d1] t2:[d0,d1,d3]
    
    enhancement 
    opened by zhenhuaw-me 0
  • Extract should be able to skip the input tensor names

    Extract should be able to skip the input tensor names

    We should be able to walk the graph starting with the output tensor names and auto infer the input names if not given.

    It would be interesting to figure out if the user provided input tensor names and output tensor names don't cut a subgraph.

    enhancement 
    opened by zhenhuaw-me 0
Releases(v0.2.1)
  • v0.2.1(Nov 13, 2022)

    What's Changed

    • Ping onnxoptimizer to 0.2.7 due to "Unresolved value references" issue. See more in https://github.com/zhenhuaw-me/onnxcli/issues/28
    • convert: enable onnx to json by @zhenhuaw-me in https://github.com/zhenhuaw-me/onnxcli/pull/10
    • inspect: print input and output tensor too by @zhenhuaw-me in https://github.com/zhenhuaw-me/onnxcli/pull/12
    • inspect: dump input output tensor by @zhenhuaw-me in https://github.com/zhenhuaw-me/onnxcli/pull/14
    • inspect: show dimension name instead of value if has any by @zhenhuaw-me in https://github.com/zhenhuaw-me/onnxcli/pull/17
    • draw: gen tensor info for tensors that only have name by @zhenhuaw-me in https://github.com/zhenhuaw-me/onnxcli/pull/18
    • setup: install the dependent python packages by @zhenhuaw-me in https://github.com/zhenhuaw-me/onnxcli/pull/19
    • Check command by @zhenhuaw-me in https://github.com/zhenhuaw-me/onnxcli/pull/21

    Full Changelog: https://github.com/zhenhuaw-me/onnxcli/compare/v0.2.0...v0.2.1

    Source code(tar.gz)
    Source code(zip)
  • v0.2.0(Jan 8, 2022)

  • v0.1.0(Dec 24, 2021)

Owner
黎明灰烬 (王振华 Zhenhua WANG)
A b[i|y]te of ML.sys|Arch|VM.
黎明灰烬 (王振华 Zhenhua WANG)
[ICCV'21] NEAT: Neural Attention Fields for End-to-End Autonomous Driving

NEAT: Neural Attention Fields for End-to-End Autonomous Driving Paper | Supplementary | Video | Poster | Blog This repository is for the ICCV 2021 pap

254 Jan 02, 2023
FCOSR: A Simple Anchor-free Rotated Detector for Aerial Object Detection

FCOSR: A Simple Anchor-free Rotated Detector for Aerial Object Detection FCOSR: A Simple Anchor-free Rotated Detector for Aerial Object Detection arXi

59 Nov 29, 2022
Efficient Householder transformation in PyTorch

Efficient Householder Transformation in PyTorch This repository implements the Householder transformation algorithm for calculating orthogonal matrice

Anton Obukhov 49 Nov 20, 2022
Tensorflow2.0 🍎🍊 is delicious, just eat it! 😋😋

How to eat TensorFlow2 in 30 days ? 🔥 🔥 Click here for Chinese Version(中文版) 《10天吃掉那只pyspark》 🚀 github项目地址: https://github.com/lyhue1991/eat_pyspark

lyhue1991 9.7k Jan 01, 2023
Improving Calibration for Long-Tailed Recognition (CVPR2021)

MiSLAS Improving Calibration for Long-Tailed Recognition Authors: Zhisheng Zhong, Jiequan Cui, Shu Liu, Jiaya Jia [arXiv] [slide] [BibTeX] Introductio

DV Lab 116 Dec 20, 2022
Learning Skeletal Articulations with Neural Blend Shapes

This repository provides an end-to-end library for automatic character rigging and blend shapes generation as well as a visualization tool. It is based on our work Learning Skeletal Articulations wit

Peizhuo 504 Dec 30, 2022
PyTorch/TorchScript compiler for NVIDIA GPUs using TensorRT

PyTorch/TorchScript compiler for NVIDIA GPUs using TensorRT

NVIDIA Corporation 1.8k Dec 30, 2022
Transfer Learning Remote Sensing

Transfer_Learning_Remote_Sensing Simulation R codes for data generation and visualizations are in the folder simulation. Experiment: California Housin

2 Jun 21, 2022
LV-BERT: Exploiting Layer Variety for BERT (Findings of ACL 2021)

LV-BERT Introduction In this repo, we introduce LV-BERT by exploiting layer variety for BERT. For detailed description and experimental results, pleas

Weihao Yu 14 Aug 24, 2022
An efficient PyTorch implementation of the winning entry of the 2017 VQA Challenge.

Bottom-Up and Top-Down Attention for Visual Question Answering An efficient PyTorch implementation of the winning entry of the 2017 VQA Challenge. The

Hengyuan Hu 731 Jan 03, 2023
RANZCR-CLiP 7th Place Solution

RANZCR-CLiP 7th Place Solution This repository is WIP. (18 Mar 2021) Installation git clone https://github.com/analokmaus/kaggle-ranzcr-clip-public.gi

Hiroshechka Y 21 Oct 22, 2022
Predicting Tweet Sentiment Maching Learning and streamlit

Predicting-Tweet-Sentiment-Maching-Learning-and-streamlit (I prefere using Visual Studio Code ) Open the folder in VS Code Run the first cell in requi

1 Nov 20, 2021
AlgoVision - A Framework for Differentiable Algorithms and Algorithmic Supervision

NeurIPS 2021 Paper "Learning with Algorithmic Supervision via Continuous Relaxations"

Felix Petersen 76 Jan 01, 2023
MEDS: Enhancing Memory Error Detection for Large-Scale Applications

MEDS: Enhancing Memory Error Detection for Large-Scale Applications Prerequisites cmake and clang Build MEDS supporting compiler $ make Build Using Do

Secomp Lab at Purdue University 34 Dec 14, 2022
Torch implementation of SegNet and deconvolutional network

Torch implementation of SegNet and deconvolutional network

Fedor Chervinskii 5 Jul 17, 2020
Create Own QR code with Python

Create-Own-QR-code Create Own QR code with Python SO guys in here, you have to install pyqrcode 2. open CMD and type python -m pip install pyqrcode

JehanKandy 10 Jul 13, 2022
OMNIVORE is a single vision model for many different visual modalities

Omnivore: A Single Model for Many Visual Modalities [paper][website] OMNIVORE is a single vision model for many different visual modalities. It learns

Meta Research 451 Dec 27, 2022
A system for quickly generating training data with weak supervision

Programmatically Build and Manage Training Data Announcement The Snorkel team is now focusing their efforts on Snorkel Flow, an end-to-end AI applicat

Snorkel Team 5.4k Jan 02, 2023
ADSPM: Attribute-Driven Spontaneous Motion in Unpaired Image Translation

ADSPM: Attribute-Driven Spontaneous Motion in Unpaired Image Translation This repository provides a PyTorch implementation of ADSPM. Requirements Pyth

24 Jul 24, 2022
Prososdy Morph: A python library for manipulating pitch and duration in an algorithmic way, for resynthesizing speech.

ProMo (Prosody Morph) Questions? Comments? Feedback? Chat with us on gitter! A library for manipulating pitch and duration in an algorithmic way, for

Tim 71 Jan 02, 2023