Slow import of AutoSklearn2Classifier

[totalhack]

Describe the bug

AutoSklearn2Classifier takes over a minute to import.

To Reproduce

>>> import time
>>> start=time.time(); from autosklearn.experimental.askl2 import AutoSklearn2Classifier; print(time.time()-start)
63.7589054107666

Expected behavior

It should be as fast as AutoSklearnClassifier.

Environment and installation:

Please give details about your installation:

root@0d672cced72d:/app# uname -a
Linux 0d672cced72d 5.4.39-linuxkit #1 SMP Fri May 8 23:03:06 UTC 2020 x86_64 GNU/Linux

root@0d672cced72d:/app# pip freeze | grep sklearn
auto-sklearn==0.14.2

Python version 3.8.7

[eddiebergman]

Hi @totalhack,

That seems odd, I will try it out today and see what I can find.

[eddiebergman]

So the issue is in how we actually perform the benefits of autosklearn2. We fit a model for predicting resampling strategies the first time the module is loaded. Future loading of the model will skip this fitting and predicting. These predictions are stored in XDG_CACHE_HOME or your home ~ as a backup. The solution is probably to export these predictions directly with the library. I will add this as an issue to work on.

Thank you for pointing this out!

[eddiebergman]

Feel free to continue the discussion in #1364, closing this as it's documented there.