zhangge6 / onnx-modifier Goto Github PK

View Code? Open in Web Editor NEW

1.2K 11.0 152.0 8.6 MB

A tool to modify ONNX models in a visualization fashion, based on Netron and Flask.

License: MIT License

Python 3.21% JavaScript 89.37% CSS 2.78% HTML 4.61% Dockerfile 0.03%

editor onnx edit visualize visualization

onnx-modifier's People

Contributors

Stargazers

Watchers

Forkers

zhangfeibao felixzhang7 liuchang0523 triple-mu leoyouli trt2022 jqzhai harleyhk wishinger-li 666dzy666 fengwang wangshuai-github xiaooquanwu dxhpc llpspark ishine zorrock baifanysu a3leon tedthemistokleous deep-learning-engineering wenhuiwang93 yolosyb xiguadong shoz-f junwenzhang chaizhongming warpuv jiangyongyu1 jiezju toby018 solomonooo lynex jd1694 skylake5200 lenan22 llmhao rabbitrunrun xjc01288 ganyz richjjj xiaotaochen namedqianli liaorongfan chengven027 yang-25 hao-star laborer123 liuqinglong110 ywfwyht yxpandjay qianxunzdl123 sdbma leayz-888 liuxubit jizhishutong anonymouszj sarubito huipengzhang chunchih linbaiwpi ai-foundry-face-recognition-project soralaro tinkingloeng pengcuo xiongrui6 ykxggg pudongdong madmrfox thucth-qt mjlovesz haheh bigbiggu martinlilili xiaocongcsu hanzz2007 bearx aphano iamfaith fengyijzh wujinzhong xuanjiawang ryan111111 chengwei920412 wangzixiang99 qingqi123 g7b9 shoulqc xionglewis wjj1993 robotwithcv pzy2000 s-hepo llf10811020205 somaet lzymmm123 normonisping gitkingly smartadpole apollohuang1

onnx-modifier's Issues

Script failed to load - Windows

Hello @ZhangGe6 ,

I appreciate your amazing and helpful work.

I am having an issue when opening the model in windows app. I am getting the following error when trying to open a basic mobilenetv2-7.onnx:

The script 'http://... /static/onnx.js' failed to load.

Thank you in advance for your help !

生成的模型无效

可以加节点, 但是生成的模型是无效的, 填入的节点的值似乎没有起到作用. 导出的模型用netron看新增的节点是没有值的(正常的节点能看到值), 用onnxruntime使用导出的模型推理会报错

sometime error when download a model

no issue anymore. Thanks

使用app_desktop.py打包windows时，报错FlaskUI没有maximized及idle_interval参数

Add Reshape but no shape initializer?

I Add a Reshape, but can not set a initializer for shape?
otherwise, if edit orgin reshape shape, after edited the output shape not change?

Cannot add reshaper

When try to add reshaper, don't know how to add static shape for shape input
Feel not correct

Add a Cast node and modify to attribute but keep roll back to NaN

I create a Cast node. to attribute is undefined initially.
I change to float32, float, Float, Float32, FLOAT, ... etc.
No matter what I put, it show NaN later

I change to 1 and it shows float32 later
I thought I make it. but when I load new model with netron, it say unknown type "1"

Really don't know how to add Cast properly

support modify input dimension？

场景描述：将原来onnx模型batch_size改为1（即输入尺寸由[batch_size, 3, 224, 224]改为[1, 3,224,224]），nerton可视化结果符合预期，但在解析modified onnx模型时通过onnx::ValueInfoProto获取输入尺寸,如tensorInfo.shape().dim(i).dim_value()，结果和batch_size相同，均为[0,3,224,224], 模型为开源mobilenet onnx

可视化
原始模型

修改batch_size=1后模型

无法成功插入slice节点，希望增加修改输入通道数的功能

作者你好，我的情况是这样的，我的模型是在灰度图上训练的，所以输入的通道数只有1，现在我要把输入的通道数变成3，这样才能享受TNN的加速，但我又不想重新制作数据集和训练，所以我想到的是：第一步，修改输入的通道数从1变成3，即[1,1,192,192]变成[1,3,192,192]；第二步，在输入后面插入一个slice算子，只选取三个通道中的第一个通道。这样我就不用重新训练模型了。
现在我跳过第一步，先尝试第二步插入slice 算子，但遇到了以下问题：
我已经修改了slice的输入和输出

也修改了第一个卷积的输入

可slice算子还是无法成功地与第一个卷积连接。
我使用的onnx模型是：model.zip

请问怎么传维度

我在使用reshape的时候，发现无法定义shape的大小，请问这点是什么处理的呢？谢谢

在多输出模型中删除多余输出节点报错：AttributeError: module 'onnx' has no attribute 'shape_inference'

问题

原始onnx模型存在多余输出节点，原始模型输出如下图1所示，希望将输出名称为Yaw，220， 221之外的所有节点都删除。故将其余输出节点在gemm部分删除，删除成功后输出节点如下图2所示，但点击download后报错如图3所示。

原始输出节点-图1

删除后输出节点-图2

报错现象-图3

具体报错log

127.0.0.1 - - [29/Jan/2023 16:31:52] "POST /open_model HTTP/1.1" 200 -
[2023-01-29 16:38:03,331] ERROR in app: Exception on /download [POST]
Traceback (most recent call last):
  File "/Users/cxt/miniconda3/envs/pytorch1.9/lib/python3.6/site-packages/flask/app.py", line 2292, in wsgi_app
    response = self.full_dispatch_request()
  File "/Users/cxt/miniconda3/envs/pytorch1.9/lib/python3.6/site-packages/flask/app.py", line 1815, in full_dispatch_request
    rv = self.handle_user_exception(e)
  File "/Users/cxt/miniconda3/envs/pytorch1.9/lib/python3.6/site-packages/flask/app.py", line 1718, in handle_user_exception
    reraise(exc_type, exc_value, tb)
  File "/Users/cxt/miniconda3/envs/pytorch1.9/lib/python3.6/site-packages/flask/_compat.py", line 35, in reraise
    raise value
  File "/Users/cxt/miniconda3/envs/pytorch1.9/lib/python3.6/site-packages/flask/app.py", line 1813, in full_dispatch_request
    rv = self.dispatch_request()
  File "/Users/cxt/miniconda3/envs/pytorch1.9/lib/python3.6/site-packages/flask/app.py", line 1799, in dispatch_request
    return self.view_functions[rule.endpoint](**req.view_args)
  File "app.py", line 27, in modify_and_download_model
    onnx_modifier.modify(modify_info)
  File "/Users/cxt/Reposities/onnx-modifier/onnx_modifier.py", line 302, in modify
    self.post_process()
  File "/Users/cxt/Reposities/onnx-modifier/onnx_modifier.py", line 278, in post_process
    shape_inference()
  File "/Users/cxt/Reposities/onnx-modifier/onnx_modifier.py", line 266, in shape_inference
    inferred_shape_info = onnx.shape_inference.infer_shapes(self.model_proto)
AttributeError: module 'onnx' has no attribute 'shape_inference'

new output after Conv layer

Is it possible to add new "out_put layer", for instance, output after a convolution layer.
Sometimes, we might want to check feature maps from the convolution layer as for image recognition.

I'd like to get such output. I found such way for Pytorch and Keras/Tensorflow.
I'd like to do the same thing using onnx model in onnxRT by cutting(?) and modifing the onnx model.

I can not find tutorial of it. If there exists such exsamples, please teach me.

如何增加INPUT呢？

比如我现在有一个ADD节点，现在有两个输入，我想变成三个输入，我改如何增加呢

Add a Dockerfile?

I think it maybe a good choice to run app.py inside a docker.

精度问题

增加的节点上填入的值与导出的值不一致, 是类型转换导致的精度截断?

Any other examples to add nodes, like add a multi-dimension constant?

like input is [1,3,480,640] and the constant [1,3,1,1] (100,100,100)
and we want to broadcast add them ?

Thank you!

数据type仍存在问题

删除一个reshape节点后，修改了transpose的属性，但是最后输出的type存在问题。

删除子分支后，分支上的节点变成了inputs。

1，原始模型删除子分支，如下图：

2，保存后打开修剪过的模型，可以看到inputs变多了，如下图：

3，我是通过如下代码对模型的inputs进行了删除：

import onnx                         

inputpath = "./descriptor.onnx"

# 1，修改模型
model = onnx.load(inputpath)
inputs = model.graph.input
inputs_num = 1  # 原始网络的输入会排在列表的最前面，所以调用pop函数删除冗余输入
for i in range(len(inputs)-inputs_num):
    inputs.pop()
print(inputs)

# 2，检查结构并保存模型（覆盖保存）
onnx.checker.check_model(model)
onnx.save(model, inputpath)

如何增加节点呢？

initializer无法修改

我按照您的办法修改initializer，修改之后download，但新的onnx中仍未修改。

请问可以提供大陆网盘的可执行文件吗

谢谢

ConstantofShape去除此单个节点

不能直接删除ConstantofShape单个节点，想请教如果要删除这个节点这种需要如何操作呐

reconnect feature is unspported

for example:
a -> b
b -> c

I modified the output of node a to the input of node c, but it will not reconnect between a and c.

Graph modified but output is not giving

Hi @ZhangGe6
First of all thanks for this wonderful tool.
Actually i modified my graph using this tool and when i am viewing in netron it is proper but the node which i added is giving no output means i can't see what shape it output or what type it outputs it just showing the output name as text. can you please help me to fix this?

Thanks in Advance:)

support feature "mark a tensor as model output"

您好，
想问下可以支持选择一个node从的输入或输出tensor 并把他标记成模型的输出吗？

添加sub节点，导出报错

模型的均值是[0.5,0.5,0.5]，我尝试在onnx的input0后面添加sub节点来实现减均值的操作。但一直没有成功。
下图是正确的onnx的参数：

我把上图中的正确参数：参数和type都复制过来，一直提示错误。

请教一下：该如何处理呢？【版本是今天下载的新版本】

macOS 12.5.1 端口有冲突

设备：Macbook Air M1
运行报错：

Serving Flask app 'app'
Debug mode: on
Address already in use
Port 5000 is in use by another program. Either identify and stop that program, or start the server with a different port.
On macOS, try disabling the 'AirPlay Receiver' service from System Preferences -> Sharing.

修改模型输入名时报错

作者你好，我在用最新版的时候，修改了模型输入名之后点击download按钮报错如下:

期待您的回复！

open model failed

Failed to open the model, has anyone encountered this problem?

选中节点后点击add output没有反应

如题，想在某个节点增加输出但是点击后没有用，没有显示输出节点

Unable to download model after changes

@ZhangGe6 This tool is exactly what I was looking for but unfortunately, after I make my changes on the development server I get the error that on line 27 of app.py, that onnx_modifier is not defined.

Traceback (most recent call last):
  File "/home/travis/p38_venv/lib/python3.8/site-packages/flask/app.py", line 2091, in __call__
    return self.wsgi_app(environ, start_response)
  File "/home/travis/p38_venv/lib/python3.8/site-packages/flask/app.py", line 2076, in wsgi_app
    response = self.handle_exception(e)
  File "/home/travis/p38_venv/lib/python3.8/site-packages/flask/app.py", line 2073, in wsgi_app
    response = self.full_dispatch_request()
  File "/home/travis/p38_venv/lib/python3.8/site-packages/flask/app.py", line 1519, in full_dispatch_request
    rv = self.handle_user_exception(e)
  File "/home/travis/p38_venv/lib/python3.8/site-packages/flask/app.py", line 1517, in full_dispatch_request
    rv = self.dispatch_request()
  File "/home/travis/p38_venv/lib/python3.8/site-packages/flask/app.py", line 1503, in dispatch_request
    return self.ensure_sync(self.view_functions[rule.endpoint])(**req.view_args)
  File "/home/travis/memryx_dev/onnx-modifier/app.py", line 27, in modify_and_download_model
    onnx_modifier.reload()   # allow downloading for multiple times
NameError: name 'onnx_modifier' is not defined

Adding output doesn't work

Hi,

Thank you for creating this tool. It's absolutely great!

Problem

Adding an output as shown here: https://github.com/ZhangGe6/onnx-modifier#add-new-model-outputs doesn't work.

Steps to reproduce the issue:

Load in squeezenet1.0-12.onnx
Click on any node
Click on "Add output"
Nothing happens

Kind regards,
Arne

[Bug Report] aotumatically add a tensor type to model

Hi, I met a bug-like problem here.

Suppose I open a quantized onnx model in onnx-modifier and download it directly. In that case, it will add an explicit tensor type to some ops, such as QuantizeLinear, even with nothing modified.

The figure below shows the explicit type added by onnx-modifier automatically and there is no type information in the original model at the same tensor.

ref onnx: link
onnx-modifier will add type 'uint8' to the outputs of all QuantizeLinear ops in this onnx model.

We hope that onnx-modifier will keep everything that isn't modified the same as before. Thanks!

修改模型中的某个节点后，后续模型的节点需要全部重写修改

将pad算子进行修改，修改为conv，修改后如下图：

修改后模型不能正常识别后续维度大小，也就是分支上不显示尺寸，需要将重新删除算子再进行添加

can't download modified model

加载模型后点download按钮就会提示错误

无法在输出节点前增加新节点

作者你好，我想在输出节点前增加一个argmax节点，该节点位于Add和heatmap之间，最后还要把模型输出名从heatmap重命名成indexmap，如下图所示：

但是我在操作过程中遇到问题，就是Add节点的输出张量名我无法将其从indexmap修改成heatmap（改成其他名字也不可以），好像Add的输出张量名跟模型的输出节点名绑定死了，输出节点名是什么，Add的输出张量名就是什么，见下面视频的后面部分
https://user-images.githubusercontent.com/43233772/220515357-088b3043-770a-4009-86f3-ff40b977f8e7.mp4
模型：
model.zip

仅修改onnx 输入和输出节点名称之后，用trtexec/onnx2trt转换模型找不到输出节点

Sometimes when you save it it doesn't save the outputs

I am using the Windows 64-bit version.
Steps: open mnist-1.onnx file. Click download. Then open modified-mnist-1.onnx This is what I see:
screenshot

Feature request: changing output and input shapes

It would be nice to be able to change the output shape and data type from the UI.
The same goes for the input and intermediate tensors :)

How to prohibit auto-translation to chinese!

Hi,there
I've encountered a problem that once I open the app on windows, everything would be ugly translated into chinese. How could I prevent this from happening?