Problems replicating COMPASS web demo using TensorFlow Serving on Docker about what-if-tool

[jameswex]

Thanks for reaching out! It is possible for you to share a jupyter or colab notebook that trains and saves the model, and creates the tf.record file (or share the tf.record file)?

[aranAbm]

Sure! I can share both, I used the WIT_compas jupyter notebook (https://github.com/PAIR-code/what-if-tool/blob/master/WIT_COMPAS.ipynb) and add some code.
You can find the jupyter notebook and the tfrecord file here https://github.com/aranAbm/aam

Thank you!

[jameswex]

Thanks so much for the quick response.

With your notebook, I was able to train the model, save it as an estimator, and see the results in WIT inside of a notebook with the following cell, without any errors in WIT.

from witwidget.notebook.visualization import WitConfigBuilder
from witwidget.notebook.visualization import WitWidget
config_builder = WitConfigBuilder(examples[0:num_datapoints]).set_inference_address("localhost:8500").set_model_name("saved_model")
WitWidget(config_builder, height=tool_height_in_px)

You should be able to use it in TensorBoard as well, by setting the inference address to localhost:8500 and model name to "saved_model", and point the examples path to the tfrecords file on disk.

What settings are you suing in WIT?

[aranAbm]

Well, first I run the docker command I mentioned above and then I open TensorBoard using a jupyter notebook, go to WIT and these are the settings I am using:

• localhost:8500 as inference address
• saved_model as Model name
• I've read that the tfrecord file should be placed at the folder where the saved_model is saved, so
  the paht to examples will be the path to the tfrecord file which in my case is:
  C:\Users\arancha.abad\What if\versiones\1586269839\COMPAS_TFR.tfrecord

Here is an screenshot with the set up:

And I don't know why but suddenly it is working!

Thank you so much for your time!