How to export models?

For some applications the models need to be exported from the Monolith platform. Although Monolith offers REST APIs that enable to call models from outside the platform, there might be situations for which APIs are not suitable, and the actual model is needed. For example, a calibration model which needs to be embedded in a firmware with no internet connection.

In general, currently the models themselves (i.e. the code or compiled code) can’t be exported from the platform. There are two options to export models implicitly or indirectly.

Polynomial model coefficients

The Polynomial Regression model allows to expose the coefficients of the fitted model. These coefficients can be used to rebuild the same model on any other system outside of the Monolith platform.

Therefore, if you need to export a model make sure to check the polynomial model and if it is suitable for your problem.

Look-up table

If another model is used, direct model export is not possible. Instead you can generate a look-up table and use that in your firmware. To generate a look-up table go through the following steps:

• Generate a dataset with a grid of points spanning the entire design space. The density of the grid is a trade-off between accuracy and memory required to store and load the table.
• Upload the dataset to the Monolith platform and import it into a notebook with the Tabular importer.
• Use Dataset Prediction with that dataset to generate a model prediction for each point.
• Export the resulting dataset which is the look-up table.