Saving and loading trained models

[jdjelic]

I am sorry if this already exists, but I was not able to figure it out.
Namely, if I try to fit a gcm model with auto causal structure assignment of AssignmentQuality.BETTER (or better), the fitting of the model takes a lot of time (measured in hours if not longer).
If I would like to use my learned model in production (or to test some repeated actions, by using the model), then it would be necessary to train the model ones and save it / load it later when needed. (like we can do with all major deep learning frameworks, such as PyTorch).

[bloebp]

Hey,

At the moment, if the model has been fitted with the GOOD and BETTER, you can simply pickle the fitted causal model (saved to disk) and load it. For instance,

import pickle

...

my_scm = StructuralCausalModel(graph)
gcm.auto.assign_causal_mechanisms(my_scm, data)
gcm.git(my_scm, data)

# Save fitted GCM object to disk
with open("./my-gcm.pickle", "wb") as f:
    pickle.dump(my_scm, f)

...

# Load fitted GCM object from disk
with open("./my-gcm.pickle", "rb") as f:
    my_scm = pickle.load(f)

Let me know if this work, I can also add this to the documentation.

[jdjelic]

Ok, that sounds good! Thanks a lot for your quick reply. I'll give it a try!

[dmly]

This is what I do

[jdjelic]

I have additional question though. If I try using the AssignmentQuality.BEST mode, during the fit() I get the following error:
PicklingError: Can't pickle : it's not the same object as sklearn.metrics._classification.accuracy_score

[bloebp]

The BEST option only works with autogluon, which currently might not work with pickle. However, the error seems a bit unrelated. Does it work with GOOD and BETTER? If not, do you have a code snippet?

[jdjelic]

Yes, I installed autogluon and it should call autogluon automatically inside the fit(), right? But for some reason, there is an attempt to pickle as well. With GOOD it does work, but I am not satisfied with the quality of the SCM produced via GOOD, so I tried using BETTER, where fit() was running for more than 7 hours, so I interrupted it and tried BEST, but BEST failed in the way I described. The github code comments state that training should be fast in case of BETTER, but what I experienced doesn't look so :) For comparison, when I try GOOD, the fit()-ing takes only a couple fo seconds. (my pandas df has around 500K rows)

[bloebp]

How many nodes does the graph have?

[jdjelic]

Well, quite a lot! 132.
Unfortunately, I cannot reduce the number of nodes as there are many small confounders that are more or less equally contributing. Does it look hopeless to even try?

[bloebp]

Ok that might explain the slow run time :) A graph with 132 nodes could generally be problematic runtime wise. What is your exact use case? Maybe you don't need to fit the whole graph.

[jdjelic]

Well, I think I have to fit the whole graph, as I want to simulate some interventions. But seemingly I either let it run for a longer time or try to reduce the number of edges I need to fit....