Hierarchical expansion of large graphs

[jenniferzhu]

I failed to load 2.1 MB .meta file generated from char-rnn model (https://github.com/sherjilozair/char-rnn-tensorflow) with python 2.7 and tensorflow 1.0.

It is loading forever... I am wondering what is wrong with these files?

[lutzroeder]

The graph has more than 5000 nodes and the graph layout is choking. Tried running GraphViz and it's running forever as well. Those graphs will need an expand and explore UI that avoids rendering the entire graph. TensorFlow support is pretty experimental at this point.

[jenniferzhu]

Thanks @lutzroeder
I hope there would be some ways to workaround this issue, since I really like Netron's image of models.

[erwincoumans]

Same issue, meta file is here in case you want to test it:
https://github.com/bulletphysics/bullet3/blob/master/examples/pybullet/gym/pybullet_data/policies/ppo/minitaur_reactive_env/model.ckpt-14000000.meta

[kobemini]

@lutzroeder You mentioned "Those graphs will need an expand and explore UI that avoids rendering the entire graph" , are there any improvements planned for release versions in the future?

[jojivk73]

For large graphs, is it possible to add an option to show a given node + fanin and fanouts upto specified level.
Example netron <g.pbtxt> -node <node_name> -level 3

[JasonWayne]

@jojivk73 there is no such args like node or level in my netron cmd line...

[yairb-gm]

Hi, I'm encountering this issue as well for models I was able to open with previous versions, of onnx files

[lutzroeder]

@yairb-gm can you please share a sample file?

[CarlPoirier]

Hi @lutzroeder,

I've seen there was EfficientDet-D0 mentioned in the other issue about this. It does work now, but Netron still barfs on EfficientDet-D2. Here is an example onnx file. I'm not very fluent in javascript, is there anything I can do quickly to get this fixed?

Thanks,

Carl

[lutzroeder]

pytorch/pytorch#75100

[ZhangGe6]

Hi folks. I developed netron-hierarchy, based on netron codebase and enables visualize models in a hierarchical way. This may be worth a try, especially for LLM models. Hope it helps!

[lutzroeder]

Duplicate of #1370