Hi @fredxu-8451,

Thank you for contacting us!

You are right in assuming that the media_tensors.media_scaled refers to media tensors, after being scaled by population and median value, but before transformations like adstock or hill functions. The Adstock and Hill transformations are handled internally, and Meridian allows you to access the assessed values of slope of the Hill/ Adstock transformer etc. from the trained model, by channel, in the inference_data attribute of the meridian model object. You may check out the answer to this previous discussion #639 for detailed steps on accessing the same.

The mmm.adstock_hill_media() method is vectorized to apply the transformation for every sample in your posterior, returning a tensor that represents the full posterior distribution of the transformed media, equivalent to LightweightMMM’s media_transformed.

Here's a sample code snippet:

mmm = # a trained Meridian model

posterior = mmm.inference_data.posterior

media_transformed = mmm.adstock_hill_media(
  media=mmm.media_tensors.media_scaled,
  alpha=posterior.alpha_m.values,
  ec=posterior.ec_m.values,
  slope=posterior.slope_m.values,
  decay_functions=mmm.adstock_decay_spec.media
)

This way you can compute media_transformed for the full posterior distribution. If you want to inspect the transformation for a single posterior draw, you can access the parameters (like alpha, ec, slope) directly from the posterior distribution as mentioned above, and apply the transformations manually. However, while approaching the manual calculation, please make sure the definition of decay_functions matches with what you passed into the model spec.

We'll take your suggestion into consideration regarding a widget to show the progress of sampling.
Our team is constantly working on improving Meridian's functionality and you may check the unreleased features on the CHANGELOG page in GitHub.

Feel free to reach out if you have any questions or suggestions regarding Meridian.

Thank you,

Google Meridian Support Team