How to save a trained model

[darkdark87]

How to save a trained model

[lucasmaystre]

Hi @darkdark87 I just released a new version with a fix (v0.1.5). Hopefully it works now—let me know if you still have issues.

[lucasmaystre]

Hi @darkdark87 , thanks for the kind words!

I've actually never had to save models before. I just tried using pickle now at the end of this colab notebook:

import pickle

with open("test.pkl", "wb") as f:
    pickle.dump(model, f)

with open("test.pkl", "rb") as f:
    m2 = pickle.load(f)

m2.plot_scores(
        items=["LAL", "CHI", "BOS"],
        resolution=10/seconds_in_year,
        figsize=(14.0, 3.0),
        timestamps=True)

and it seems to work fine. Could you share a minimum working example where it fails?

[darkdark87]

Hi @lucasmaystre

I believe the errors lies with the number of teams. If I increase the number of teams by a factor 3 I receive an error when pickling the model.

I created a minimum "working example" in this notebook.

[lucasmaystre]

Thanks, the example was helpful. I think I understand where the problem—there are circular references between Item and Observation, and I understand it trips pickle up after a certain point.

Putting this on my TODO list, but it'll probably have to wait for a few weeks.

In the meantime, if all you're interested in doing with the reloaded model is to make predictions, the following should work reasonably well:

import pickle

# Saving.
fitters = {name: val.fitter for name, val in model.item.items()}
with open("test1.pkl", "wb") as f:
    pickle.dump(fitters, f)

# Re-loading.
with open("test1.pkl", "rb") as f:
    fitters = pickle.load(f)

m2 = ks.BinaryModel()
dummy = ks.kernel.Constant(var=1.0)

for name, fitter in fitters.items():
    m2.add_item(name, kernel=dummy)
    m2.item[name].fitter = fitter

It only saves the parameters of the model (which are in the items' fitter property) though, so you won't be able to retrain it, or inspect the observations it was trained on, etc. However, model.probabilites(), item.predict() should work fine.

I'll leave the issue open until I fix persistency in a cleaner way.

[darkdark87]

Sounds good @lucasmaystre
My use case requires retraining of the persisted model, so I'll wait.

Thank you.