After the latest update, Netron remains open taking up memory and CPU after closing the program. I must close it through task manager each time. I am on Windows 10

• Netron app and version: web app 5.5.9?
• OS and browser version: Manjaro GNOME on firefox 97.0.1

Steps to Reproduce:

1. use torch.broadcast_tensors
2. export with torch.trace(...).save()
3. open in netron.app

I have also gotten a Unsupported function 'torch.broadcast_tensors', but have been unable to reproduce it due to this current error. Most likely, the fix for the following repro will cover two bugs.

Please attach or link model files to reproduce the issue if necessary.

Repro:

import torch

class Test(torch.nn.Module):
    def forward(self, a, b):
        a, b = torch.broadcast_tensors(a, b)
        assert a.shape == b.shape == (3, 5)
        return a + b

torch.jit.trace(
    Test(),
    (torch.ones(3, 1), torch.ones(1, 5)),
).save("foobar.pt")

Zipped foobar.pt: foobar.zip

• [x] 1. Opening rm_lstm4f.xml results in TypeError (#192)
• [x] 2. dot files are not opened any more - need to fix it (#192)
• [x] 3. add preflight check for invalid xml and dot content
• [x] 6. Add test files to ./test/models.json (#195) (#211)
• [x] 9. Add support for the version 3 of IR (#196)
• [x] 10. Category color support (#203)
• [x] 11. -metadata.json for coloring, documentation and attribute default filtering (#203).
• [x] 5. Filter attribute defaults based on -metadata.json to show fewer attributes in the graph
• [ ] 7. Show weight tensors
• [x] 8. Graph inputs and outputs should be exposed as Graph.inputs and Graph.outputs
• [x] 12. Move to DOMParser
• [x] 13. Remove dot support

• Netron app and version: 4.4.8 App and Browser
• OS and browser version: Windows 10 + Chrome Version 84.0.4147.135

Steps to Reproduce:

EfficientDet-d0.zip

Please attach or link model files to reproduce the issue if necessary.

Hi there,

First off, just wanted to say thanks for creating such a great tool - Netron is very useful.

I'm having an issue that likely stems from Tensorflow, rather than from Netron, but thought you might have some insights. In my flow, I use TF 1.15 to go from .ckpt --> frozen .pb --> .tflite. Normally it works reasonably smoothly, but a recent run shows an issue with the .tflite file: it is created without errors, it runs, but it performs poorly. Opening it with Netron shows that the activation functions (relu6 in this case) have been removed for every layer. Opening the equivalent .pb file in Netron shows the relu6 functions are present.

Have you seen any cases in which Netron struggled with a TF Lite model (perhaps it can open, but isn't displaying correctly)? Also, how did you figure out the format for .tflite files (perhaps knowing this would allow me to debug it more deeply)?

Thanks in advance.

here's patch to support armnn format. (experimental)

armnn-schema.js is compiled from ArmnnSchema.fbs included in armNN serailizer.

see also:

armnn: https://github.com/ARM-software/armnn

As mensioned in #363, I will check items in below:

• [x] Add sample files to test/models.json and run node test/test.js armnn
• [x] Add tools/armnn script and sync, schema to automate regenerating armnn-schema.js
• [x] Add tools/armnn script to run as part of ./Makefile
• [x] Run make lint

Hi, there is some questions about node's name which in pt model saved by TorchScript. I use netron to view my pt model exported by torch.jit.save()，but the node's name doesn't match with it's real name resolved by TorchScript interface. It looks like the names in pt are arranged numerically from smallest to largest，but this is clearly not the case when they are parsed from TorchScript's interface. I wonder how this kind of situation can be solved, thanks a lot !! Looking forward to your reply.

torch.fx is a library in PyTorch 1.8 that allows python-python model transformations. It works by symbolically tracing the PyTorch model into a graph (fx.GraphModule), which can be transformed and finally exported back to code, or used as a nn.Module directly. Currently there is no mechanism to import the graph IR into netron. An indirect path is to export to ONNX to visualize, which is not as useful if debugging transformations that potentially break ONNX exportability. It seems valuable to visualize the traced graph directly in netron.

I have a lot of basic model files saved in TorchScript and they were able to be opened weeks ago. However I cannot many of them after update Netron to v3.9.1. Many common functions are not supported not, e.g. torch.constant_pad_nd, torch.bmm, torch.avg_pool3d.

• Netron app and version: 4.4.4
• OS and browser version: Windows 10 64bit

Steps to Reproduce:

1. Open OpenVINO IR XML file in netron

Please attach or link model files to reproduce the issue if necessary.

I cannot share the proprietary model that shows dozens of disconnected nodes, but the one linked below does show disconnected subgraphs after conversion to OpenVINO IR. Note that the IR generated using the --generate_deprecated_IR_V7 option displays correctly.

https://github.com/ARM-software/ML-KWS-for-MCU/blob/master/Pretrained_models/Basic_LSTM/Basic_LSTM_S.pb

Convert using:

python 'C:\Program Files (x86)\IntelSWTools\openvino\deployment_tools\model_optimizer\mo.py' --input_model .\Basic_LSTM_S.pb --input=Reshape:0 --input_shape=[1,490] --output=Output-Layer/add

This results in the following disconnected graph display:

For recoverable estimator persistence scikit-learn recommends to use joblib (instead of pickle). Sidenote: It is possible to export trained models into ONNX or PMML but the estimators are not recoverable. For more info refer to here.

Export PyTorch model to FlatBuffers file:

import torch
import torchvision
model = torchvision.models.resnet34(weights=torchvision.models.ResNet34_Weights.DEFAULT)
torch.jit.save_jit_module_to_flatbuffer(torch.jit.script(model), 'resnet34.ff')

Sample files: scriptmodule.ff.zip squeezenet1_1_traced.ff.zip

fix some bugs of megengine C++ model (.mge) visualization:

1. show the shape of the middle tensor;
2. fix scope matching model identifier (mgv2) due to possible leading information;

please help review, thanks~

import torch
import torchvision
import torch.utils.tensorboard
model = torchvision.models.detection.fasterrcnn_resnet50_fpn()
script = torch.jit.script(model)
script.save('fasterrcnn_resnet50_fpn.pt')
with torch.utils.tensorboard.SummaryWriter('log') as writer:
    writer.add_graph(script, ())

fasterrcnn_resnet50_fpn.pt.zip

• OS and browser version: Ubuntu20 + Chrome 97.0.4692.99
• Netron version 5.6.6

Steps to Reproduce:

1. Take attached file and change extension to .pbtxt (this was changed due to GitHub restrictions). This model is build by Protobuf GraphDef structure that enables func structure.
2. Load with Web Netron
3. Press on Placeholder node to view attributes.

Current behavior: OutTensorInfo parameter is of type func and cannot be viewed. If pressing OutTensorInfo link, it doesn't show the data. Expected behavior: OutTensorInfo parameter can be shown and collapses/de-collapses after clicking "+" button.

Please attach or link model files to reproduce the issue if necessary. myModel.txt