Save and Load a shapelet model

[zillur01]

Hi, I trained a shapelet model. how can i save and load it for inference? I used model.to_json method but the loaded model not working.

[charavelg]

Hello. Can you share the error you get and the code used to create / save / load the model ?

[charavelg]

The serialization workflow seems KO when using keras optimizer rather than a str and/or non default n_shapelets_per_size args.
Such problems have been reported (see #387  and #403).

I have tried to fix both json and pickle serialization / deserialization in #623, if you want to take a look.

For the time being, I think it's better to explicitely not support hdf5 serialization for learning shapelets rather than extensively rework this part for this estimator only.

[zillur01]

I was using this:

# Load a dataset
X_train = TimeSeriesScalerMinMax().fit_transform(X_train)
shapelet_sizes = {300:4}

# Initialize the LearningShapelets model
shp_clf = LearningShapelets(n_shapelets_per_size=shapelet_sizes,
                            optimizer='adam',
                            weight_regularizer=0.01,
                            max_iter=10,
                            verbose=1,
                            random_state=999)

# Fit the model
shp_clf.fit(X_train, y_train)
shp_clf .to_json('shapelet_model.json')
shp_clf = LearningShapelets.from_json('shapelet_model.json')

and when I run shp_clf .predict_proba(window) , this error shows up

numpy.core._exceptions._UFuncNoLoopError: ufunc 'less' did not contain a loop with signature matching types (<class 'numpy.dtypes.Int64DType'>, <class 'numpy.dtypes.StrDType'>) -> None

[zillur01]

I see. so I tried this clf._to_dict(output='shapelet_model.json'). does not look like it save anything? could you please share the save and load instructions?

[charavelg]

I got the same error when using non default n_shapelets_per_size . This should be fixed in the main branch, you can test it with python -m pip install "tslearn[all_features] @ https://github.com/tslearn-team/tslearn/archive/main.zip" on top of your existing environment.

Your save and load instructions are correct

shp_clf.to_json('shapelet_model.json')
shp_clf = LearningShapelets.from_json('shapelet_model.json')

should work . the _to_dict method is just intern machinery.

By the way, the size of your shapelets seems huge (shapelet_sizes = {300:4} means 4 shapelets of 300 samples), what's the shape of your dataset? You can have a look here for an example (not relevant for the save load problem though).

Hope this helps!

[zillur01]

Thanks. I will try that. About the shapelet size, I was going to ask you. My understanding could be incorrect. So I have lots of time series data, each of ~1000x3 size. I noticed there are usually 4 classes where each class can be max 300x3 size. That means each time series may contain multiple classes.