Rebuild the Model explorer backend - To get the full weights values

[tsantosh1098]

I am trying to modify the ConvertFlatbufferDirectlyToJson() part of the direct_flatbuffer_to_json_graph_convert.cc file.

To re-build the builtin adaptors I followed these steps - https://github.com/google-ai-edge/model-explorer/tree/main/src/builtin-adapter. But after generating the .so file now I am not able to load the extension module ".builtin_tflite_mlir_adapter".

! Failed to load extension module ".builtin_tflite_mlir_adapter":
/home/armnn/Documents/mcw/model_explorer/lib/python3.12/site-packages/ai_edge_model_explorer_adapter/_pywrap_convert_wrapper.so: undefined symbol:
_ZN10tensorflow11CSRMatMulOpIN5Eigen16ThreadPoolDeviceEfEC2EPNS_20OpKernelConstructionE

When I check the path of the .so file, its present in the provided directory.

Currently, I am building this on ubuntu 24 machine - with Bazel 6.5.0, on Intel(R) Core(TM) i7-8665U CPU @ 1.90GHz

[tsantosh1098]

so i built for ai_edge_model_explorer_adapter 0.1.0

I have ai_edge_model_explorer_adapter 0.1.0 which is installed using internet downloaded .whl file
The another ai_edge_model_explorer_adapter 0.1.0 is generated by the local built .whl file
So after i inspected the symbols in both shared libraries here is the difference.
Note: The first one is working, second one (local built is not working)

Is this problem bcz of the glic version ?

[pkgoogle]

Hi @tsantosh1098, can you share the modifications to

ConvertFlatbufferDirectlyToJson() part of the direct_flatbuffer_to_json_graph_convert.cc

So that we may match your changes and test this out? Thanks for your help.

[pkgoogle]

Hi @tsantosh1098, are you sure you need your changes? Have you tried adjusting the maximum element count for constant tensor values:

If you enter a -1 here, it will allow you to see unlimited tensor values. Let me know if you have any more questions.

[tsantosh1098]

Hi @pkgoogle, Thanks for the information.

I was able to achieve to get the full weights by setting the config.const_element_count_limit = -1 in the builtin adapter files.
I though there was a need to do some changes in the backend, but backend already had a support to extract full length weights.

Thanks once again

[pkgoogle]

Hi @tsantosh1098, just to be clear the UI solution also works right? (We'll need to add it to the wiki so I wanted to verify :) ). Also if you have no more open items regarding this issue, please feel free to close as completed, thanks!

[github-actions]

Marking this issue as stale since it has been open for 7 days with no activity. This issue will be closed if no further activity occurs.

[github-actions]

This issue was closed because it has been inactive for 14 days. Please post a new issue if you need further assistance. Thanks!