Open source of domain to topic model.

[ydennisy]

Hi!

I have been checking out chrome://topics-internals/ thank you for this tool it is very useful!

I was wondering if you could go one step further and open up the domain to topic model itself?

I can see the model path and version here:

Model version: 2206021246
Model file path: /Users/*******/Library/Application Support/Google/Chrome/Profile 2/optimization_guide_prediction_model_downloads/64f1ed1f-3fb6-4a8f-aaa0-fe2c58dc1723/model.tflite

Would you consider open sourcing the model, or at least providing a guide, or data set on which it was trained?

Any information on this would be much appreciated!

[ydennisy]

tagging @jkarlin

[jkarlin]

Hi Dennis. The model can be loaded via TensorFlow Lite. I know a number of people have been asking how they can run it manually and so I gave it a spin myself. Here is what I did:

I basically did what's listed here.

git clone https://github.com/tensorflow/tflite-support.git
cd tflite-support
sudo apt-get install bazel-5.1.1 
bazel run -c opt tensorflow_lite_support/examples/task/text/desktop:bert_nl_classifier_demo -- --model_path=<path to your tflite file as shown in chrome://topics-internals> --text="example com"

Replace any '-', '_', '.', and '+' chars in your input text with whitespace ' '.

Note that the first run is slow as bazel has to download a bunch of third-party libraries and build them all.

The output is a list of topic ids and their score. You can map the ids to strings with the taxonomy file.

In regards to the training data, I don't believe there are any current plans to release that data.

[jkarlin]

A colleague developed a python-based colab notebook for folks to try running the model manually with as well.

Edit: changed the link

[ydennisy]

Hey @jkarlin that is great! Thanks very much for this!