Custom pre-process for new model version

[manuleo]

Hello!

We are using clearml-serving to deploy different versions of our model, but with the new versions being deployed we also need to change the pre/post-processing steps because the features we are using are changing while optimizing.

For what we understood so far, the pre/post-processing should be defined only once for all the model versions served through an endpoint, but this does not adapt to our use case. If we just change the pre/post-processing the old versions of our model will not work, and we cannot compare and test.

Is there any plan to implement bindings between a specific pre/post-processing version and a model version? If not, what could be a workaround about it? We thought about implementing a custom model with the processing embedded inside, but will this affect in any way the metrics we get out of the box or any other functionality we get when we define our model to be of a specific type? (In our case, a sklearn one).

Thanks for the answers!!

[ainoam]

@manuleo You'll note cleamrl-serving model add -- endpoint <> --version <> --preprocess <> specifies extra processing per endpoint version deployed - Does this help?

will this affect in any way the metrics we get out of the box or any other functionality we get when we define our model to be of a specific type?

For completeness, implementing a custom wrapper should not affect the builtin monitoring or routing clearml-serving provides.