google-ai-edge/model-explorer

ONNX Support

ogencoglu opened this issue · 10 comments

ogencoglu commented

Thanks for the tool. Any plans to support ONNX?

yermandy commented

ONNX would be helpful

aboubezari commented

My team also mainly uses ONNX to visualize models. Support for this would be great!

yijie-yang commented

Yes, we're planning to support ONNX in this June! Stay tuned!

  • 👍1
justinchuby commented

I implemented an adapter: https://github.com/justinchuby/model-explorer-onnx/

You can use it with

pip install --upgrade model-explorer-onnx
model-explorer --extensions=model_explorer_onnx

image

  • 👍4
justinchuby commented

Happy to contribute.

maderix commented

Thanks for this PR !

I see the onnx nodes are traversed with torch tracing, due to which not all connected nodes are displayed on the canvas(even though connectivity can be seen via the inputs and outputs. Is there a different way to display connectivity for all nodes at once?

Edit: I tried manually parsing the onnx graph and exporting it to json. Then creating a json adapter for that schema (just included "nodes" and "edges" ) . For some reason, I can see the graph but it's reversed (from outputs to inputs) :-/

justinchuby commented

@maderix could you share a snapshot to show what you mean? The display should not have anything to do with torch tracing. If possible sharing the model would be helpful too

maderix commented

@justinchuby , sorry I'm an idiot. Had to just flatten the hierarchy to see the missing connection to nodes. Layers were probably messing my graph.
Thanks for the great addition to this repo . Cheers! :)

  • 🚀3
pkgoogle commented

Hi @ogencoglu, does the above adapter satisfy your use case? Thanks for your help.

ogencoglu commented

I will try out, thanks.