Why call learn_one after predict_one?

[cdeterman]

Looking through more of the API documentation I am confused by pure_inference_mode. It appears to suggest the learn_one is called whenever predict_one is called in the default setup. This seems strange to me as all the examples such as the Bike Sharing example here the predict_one is always followed by learn_one.

Am I missing something here? Does the documentation for the pure_inference_mode suggest that learn_one isn't necessary in the default mode?

[MaxHalford]

It's a bit more complicated than that. This only applies to pipelines. In a pipeline, the unsupervised parts of the pipeline (say, the scaler before the linear model) are updated when predict_one is called. You don't always want that to happen. For instance, you might just want to make a prediction, without updating the scaler. This is what pure_inference_mode enables.

Does that help? Basically, I think the confusion you have is that you expect learn_one to call learn_one on each part of the pipeline, and likewise for predict_one. But it's not exactly what happens under the hood.