Let's assume we've got a trainable function "f(df)" which uses an encoder inside of it.
We have two instances of it, trained using different folds, hence they have different encoders.
When we call first instance on df_test, we will cache it as "f_{hash(df_test)}_df", and when we call the second one, we'll overwrite this file, as naming considers only function name and hash of df, but not encoders used inside of the function.
Let's assume we've got a trainable function "f(df)" which uses an encoder inside of it.
We have two instances of it, trained using different folds, hence they have different encoders.
When we call first instance on df_test, we will cache it as "f_{hash(df_test)}_df", and when we call the second one, we'll overwrite this file, as naming considers only function name and hash of df, but not encoders used inside of the function.