[tools] model-explorer circle adaptor

[seanshpark]

google has provided model graph viewer https://github.com/google-ai-edge/model-explorer . our circle model viewing can be provided with provided adapter extension https://github.com/google-ai-edge/model-explorer/wiki/6.-Develop-Adapter-Extension as separate python package through pypi.

current status of model viewing with netron

• netron is excellent for viewing circle models
• recent huge LLM models have problems with loading time and OOM with large number or nodes (like 10,000 or higher)

I'd like to begin an adapter for circle model, maybe in tools folder :)

[seanshpark]

About model explorer: https://ai.google.dev/edge/model-explorer

[seanshpark]

items to work with

• [x] build (prototype) of circle adapter
• [ ] testing
• [x] CI checkers
• [ ] performance measure and enhance
• [x] produce python package
  • test done with personal project (by seanshpark)
• [ ] .github/workflow to upload to pypi
  • test done with personal project (by seanshpark)
• [ ] ...

please edit this comment for items development and breakdown

steps

• [x] 1: end-to-end build up with single Op (Add or Conv2D) model
• [x] 2: more Ops: add Ops and Ops
• [ ] 3: features(?) and testing(?)
• [ ] 4: dogfooding
• [ ] 5: enhancements
• [ ] ...

[seanshpark]

where to reference?

• src/server/package/src/model_explorer/builtin_tflite_flatbuffer_adapter.py
• ConvertFlatbufferDirectlyToJson method and callers

to publish from github Actions to pypi.org,

• https://docs.pypi.org/trusted-publishers/
• https://docs.github.com/en/actions/security-for-github-actions/security-hardening-your-deployments/about-security-hardening-with-openid-connect
• https://docs.github.com/en/actions/managing-workflow-runs-and-deployments/managing-deployments/managing-environments-for-deployment
• https://openid.net/developers/how-connect-works/

[batcheu]

I've tested my initial version of circle adapter and it seems to be working ;)

• https://github.com/batcheu/ONE/tree/draft/me_adapter/tools/model_explorer_adapter

Screenshot of example circle model

It can also show a huge model (>1GB) either but it's very slow. (seems slower than netron)

[seanshpark]

experiments doing/done

• [x] produce an example-package
  • discover issues and solve them while building an example-package from scratch
  • [x] simple package with unit test
  • https://github.com/seanshpark/example_package_sean
  • [x] Workflow to publish to pypi.org
  • https://github.com/seanshpark/example_package_sean/blob/main/.github/workflows/pub_pypi.yml
  • https://pypi.org/project/example-package-sean/
• [x] produce an example-package, source inside /tools/example-package folder
  • [x] simple package with unit test
  • https://github.com/seanshpark/example_package_subfolder/tree/main/compiler/tools/example_package_subfolder
  • [x] Workflow to publish to pypi.org
  • https://github.com/seanshpark/example_package_subfolder/blob/main/.github/workflows/pub_pypi.yml
  • https://pypi.org/project/example-package-subfolder/

[seanshpark]

Recently, we've added __init__.py file in circle_adapter folder, but seems guide (https://github.com/google-ai-edge/model-explorer/wiki/6.-Develop-Adapter-Extension) doesn't seem to mention this file...

--> I think we can remove this file like other Adapters.