LearningShapelets: impossible to pickle the model

[alevangel]

https://github.com/tslearn-team/tslearn/blob/42a56cce63d8263982d616fc2bef9009ccbedab4/tslearn/shapelets/shapelets.py#L212-L886

I'm having issues trying to save my fitted model, getting this error: AttributeError: Can't pickle local object 'make_gradient_clipnorm_fn.<locals>.<lambda>'

How to handle it?

To reproduce:

from tslearn.shapelets import LearningShapelets

clf = LearningShapelets(n_shapelets_per_size={20:1, 30:1},
                                max_iter=1000,
                                optimizer=Adam(learning_rate=0.1),
                                verbose=0)
clf.fit(X_train, y_train)
clf.to_pickle("path/to/file.pkl")

Versions: python 3.6 tslearn 0.5.2 tensorflow 2.6.2

[GillesVandewiele]

Does the fix described in #387 resolve the issue?

[alevangel]

Does the fix described in #387 resolve the issue?

@GillesVandewiele Nope, because this tryies to solve the 'loading' pointing out that "from_" is a method class, but I don't see nothing about saving the model different from what I did. :(

[GillesVandewiele]

It could perhaps be related to an older version of tensorflow, or something related to the optimizer being used: https://github.com/NREL/phygnn/issues/16

One way to circumvent the issues is to just store the learned shapelets and load them in again when doing inference, these contain all the valuable information of a shapelet model.

[alevangel]

One way to circumvent the issues is to just store the learned shapelets and load them in again when doing inference, these contain all the valuable information of a shapelet model.

@GillesVandewiele Yes, this is absolutely the way of hacking this issue.

For the version of tensorflow I'm using is 2.6.2 (updated the original post)

[rtavenar]

@alevangel : could you post a code snippet that does that? This would be helpful:

• for other users to see how to handle such errors
• in the future for tslearn devs (maybe you?) to figure out how to implement a fix

[alevangel]

I undesrtood that the problem is for the Adam optimizer, that have problem to be serialized. I'm trying to understand how to fix it.

@rtavenar sure, I'll share it here as soon as I have it.

[alevangel]

Just a quick fix, since the problem is the Adam optimizer, and after the fit i won't need it anymore I just removed the optimizer.

dummy = LearningShapelets(*args)  # with the standard optimizer
my_clf = LearningShapelets(*args)  # with the Adma optimizer
my_clf.fit(data)
my_clf.optimizer = dummy.optimizer
my_clf.to_pickle(path)