TRILITE¶

class TRILITE(alpha: float = 0.01, weight_ref_min: float = 0.6, percentage_mask_length: float = 0.2, use_mixing_up: bool = True, use_masking: bool = True, z_normalize_pos_neg: bool = True, backbone_network: BaseDeepLearningNetwork = None, latent_space_dim: int = 128, latent_space_activation: str = 'linear', random_state: int | np.random.RandomState | None = None, verbose: bool = False, optimizer: Optimizer | None = None, file_path: str = './', save_best_model: bool = False, save_last_model: bool = False, save_init_model: bool = False, best_file_name: str = 'best_model', last_file_name: str = 'last_model', init_file_name: str = 'init_model', callbacks: Callback | list[Callback] | None = None, batch_size: int = 64, use_mini_batch_size: bool = False, n_epochs: int = 2000)[source]¶

Bases: BaseCollectionTransformer

TRIplet Loss In TimE (TRILITE).

TRILITE [1] is a self-supervised model that learns a latent space through the triplet loss mechanism by reducing the loss between close samples and increasing it between far samples. TRILITE generates the triplets using two techniques, mixing up and masking. For each reference series (ref), a positive representation of ref is generated by mixing it up with two other randomly chosen time series from the dataset then masking a part of it. The weights of the mixing up procedure are randomly chosen for the two randomly selected series in a way that the ref still has the highest weight. The same procedure is used to generated the negative representation however by using another ref.

Parameters:

alphafloat, default = 1e-2: The value that controls the space of the triplet loss, the smaller the value the more difficult the problem becomes, the higher the value the more easy the problem becomes, a balance should be found.
weight_ref_minfloat, default = 0.6: The weight of the reference series used for the triplet generation.
percentage_mask_lengthint, default = 0.2: The percentage of time series length to calculate the length of the masking used for the triplet generation. Default is 20%.
use_mixing_upbool, default = True: Whether or not to use mixing up during the triplet generation phase.
use_maskingbool, default = True: Whether or not to use masking during the triplet generation phase.
z_normalize_pos_negbool, default = True: Whether or not to z_normalize (mean 0 and std 1) pos and neg samples after generating the triplet.
backbone_networkaeon Network, default = None: The backbone network used for the SSL model, it can be any network from the aeon.networks module on condition for it’s structure to be configured as “encoder”, see _config attribute. For TRILITE, the default network used is FCNNetwork.
latent_space_dimint, default = 128: The size of the latent space, applied using a fully connected layer at the end of the network’s output.
latent_space_activationstr, default = “linear”: The activation to control the range of values of the latent space.
random_stateint, RandomState instance or None, default=None: If int, random_state is the seed used by the random number generator; If RandomState instance, random_state is the random number generator; If None, the random number generator is the RandomState instance used by np.random. Seeded random number generation can only be guaranteed on CPU processing, GPU processing will be non-deterministic.
verboseboolean, default = False: Whether to output extra information.
optimizerkeras.optimizer, default = tf.keras.optimizers.Adam(): The keras optimizer used for training.
file_pathstr, default = “./”: File path to save best model.
save_best_modelbool, default = False: Whether or not to save the best model, if the modelcheckpoint callback is used by default, this condition, if True, will prevent the automatic deletion of the best saved model from file and the user can choose the file name.
save_last_modelbool, default = False: Whether or not to save the last model, last epoch trained, using the base class method save_last_model_to_file.
save_init_modelbool, default = False: Whether to save the initialization of the model.
best_file_namestr, default = “best_model”: The name of the file of the best model, if save_best_model is set to False, this parameter is discarded.
last_file_namestr, default = “last_model”: The name of the file of the last model, if save_last_model is set to False, this parameter is discarded.
init_file_namestr, default = “init_model”: The name of the file of the init model, if save_init_model is set to False, this parameter is discarded.
callbackskeras callback or list of callbacks,: default = None The default list of callbacks are set to ModelCheckpoint and ReduceLROnPlateau.
batch_sizeint, default = 64: The number of samples per gradient update.
use_mini_batch_sizebool, default = False: Whether or not to use the mini batch size formula.
n_epochsint, default = 2000: The number of epochs to train the model.

Notes

Capabilities ¶
Missing Values	No
Multithreading	No
Inverse Transform	No
Univariate	Yes
Multivariate	Yes
Unequal Length	No

Adapted from the implementation from Ismail-Fawaz et. al https://github.com/MSD-IRIMAS/TRILITE

References

[1]

Ismail-Fawaz, Ali, Maxime Devanne, Jonathan Weber,

and Germain Forestier. “Enhancing time series classification with self-supervised learning.” In International Conference on Agents and Artificial Intelligence (ICAART), pp. 40-47. SCITEPRESS-Science and Technology Publications, 2023.

Examples

>>> from aeon.transformations.collection.self_supervised import TRILITE
>>> from aeon.datasets import load_unit_test
>>> X_train, y_train = load_unit_test(split="train")
>>> ssl = TRILITE(latent_space_dim=2, n_epochs=5)
>>> ssl.fit(X_train)
TRILITE(...)
>>> X_train_transformed = ssl.transform(X_train)

Methods

`build_model`(input_shape)	Construct a compiled, un-trained, keras model that is ready for training.
`clone`([random_state])	Obtain a clone of the object with the same hyperparameters.
`fit`(X[, y])	Fit transformer to X, optionally using y if supervised.
`fit_transform`(X[, y])	Fit to data, then transform it.
`get_class_tag`(tag_name[, raise_error, ...])	Get tag value from estimator class (only class tags).
`get_class_tags`()	Get class tags from estimator class and all its parent classes.
`get_fitted_params`([deep])	Get fitted parameters.
`get_params`([deep])	Get parameters for this estimator.
`get_tag`(tag_name[, raise_error, ...])	Get tag value from estimator class.
`get_tags`()	Get tags from estimator.
`load_model`(model_path)	Load a pre-trained keras model instead of fitting.
`reset`([keep])	Reset the object to a clean post-init state.
`save_last_model_to_file`([file_path])	Save the last epoch of the trained deep learning model.
`set_params`(**params)	Set the parameters of this estimator.
`set_tags`(**tag_dict)	Set dynamic tags to given values.
`transform`(X[, y])	Transform X and return a transformed version.

build_model(input_shape)[source]¶

Construct a compiled, un-trained, keras model that is ready for training.

In aeon, time series are stored in numpy arrays of shape (d,m), where d is the number of dimensions, m is the series length. Keras/tensorflow assume data is in shape (m,d). This method also assumes (m,d). Transpose should happen in fit.

Parameters:

input_shapetuple[int, int]: The shape of the data fed into the input layer, should be (m, d).

Returns:

outputa compiled Keras Model

clone(random_state=None)[source]¶

Obtain a clone of the object with the same hyperparameters.

A clone is a different object without shared references, in post-init state. This function is equivalent to returning sklearn.clone of self. Equal in value to type(self)(**self.get_params(deep=False)).

Parameters:

random_stateint, RandomState instance, or None, default=None: Sets the random state of the clone. If None, the random state is not set. If int, random_state is the seed used by the random number generator. If RandomState instance, random_state is the random number generator.

Returns:

estimatorobject: Instance of type(self), clone of self (see above)

fit(X, y=None)[source]¶

Fit transformer to X, optionally using y if supervised.

Writes to self: - is_fitted : flag is set to True. - model attributes (ending in “_”) : dependent on estimator

Parameters:

Xnp.ndarray or list

Data to fit transform to, of valid collection type. Input data, any number of channels, equal length series of shape ( n_cases, n_channels, n_timepoints) or list of numpy arrays (number of channels, series length) of shape [n_cases], 2D np.array (n_channels, n_timepoints_i), where n_timepoints_i is length of series i. Other types are allowed and converted into one of the above.

Different estimators have different capabilities to handle different types of input. If self.get_tag("capability:multivariate") is False, they cannot handle multivariate series. If self.get_tag( "capability:unequal_length") is False, they cannot handle unequal length input. In both situations, a ValueError is raised if X has a characteristic that the estimator does not have the capability to handle.

ynp.ndarray, default=None

1D np.array of float or str, of shape (n_cases) - class labels (ground truth) for fitting indices corresponding to instance indices in X. If None, no labels are used in fitting.

Returns:

selfa fitted instance of the estimator

fit_transform(X, y=None)[source]¶

Fit to data, then transform it.

Fits the transformer to X and y and returns a transformed version of X.

State change:: Changes state to “fitted”.

Writes to self: _is_fitted : flag is set to True. model attributes (ending in “_”) : dependent on estimator.

Parameters:

Xnp.ndarray or list

ynp.ndarray, default=None

1D np.array of float or str, of shape (n_cases) - class labels (ground truth) for fitting indices corresponding to instance indices in X. If None, no labels are used in fitting.

Returns:

transformed version of X

classmethod get_class_tag(tag_name, raise_error=True, tag_value_default=None)[source]¶

Get tag value from estimator class (only class tags).

Parameters:

tag_namestr: Name of tag value.
raise_errorbool, default=True: Whether a ValueError is raised when the tag is not found.
tag_value_defaultany type, default=None: Default/fallback value if tag is not found and error is not raised.

Returns:

tag_value: Value of the tag_name tag in cls. If not found, returns an error if raise_error is True, otherwise it returns tag_value_default.

Raises:

ValueError: if raise_error is True and tag_name is not in self.get_tags().keys()

Examples

>>> from aeon.classification import DummyClassifier
>>> DummyClassifier.get_class_tag("capability:multivariate")
True

classmethod get_class_tags()[source]¶

Get class tags from estimator class and all its parent classes.

Returns:

collected_tagsdict: Dictionary of tag name and tag value pairs. Collected from _tags class attribute via nested inheritance. These are not overridden by dynamic tags set by set_tags or class __init__ calls.

get_fitted_params(deep=True)[source]¶

Get fitted parameters.

State required:: Requires state to be “fitted”.

Parameters:

deepbool, default=True: If True, will return the fitted parameters for this estimator and contained subobjects that are estimators.

Returns:

fitted_paramsdict: Fitted parameter names mapped to their values.

get_params(deep=True)¶

Get parameters for this estimator.

Parameters:

deepbool, default=True: If True, will return the parameters for this estimator and contained subobjects that are estimators.

Returns:

paramsdict: Parameter names mapped to their values.

get_tag(tag_name, raise_error=True, tag_value_default=None)[source]¶

Get tag value from estimator class.

Includes dynamic and overridden tags.

Parameters:

tag_namestr: Name of tag to be retrieved.
raise_errorbool, default=True: Whether a ValueError is raised when the tag is not found.
tag_value_defaultany type, default=None: Default/fallback value if tag is not found and error is not raised.

Returns:

tag_value: Value of the tag_name tag in self. If not found, returns an error if raise_error is True, otherwise it returns tag_value_default.

Raises:

ValueError: if raise_error is True and tag_name is not in self.get_tags().keys()

Examples

>>> from aeon.classification import DummyClassifier
>>> d = DummyClassifier()
>>> d.get_tag("capability:multivariate")
True

get_tags()[source]¶

Get tags from estimator.

Includes dynamic and overridden tags.

Returns:

collected_tagsdict: Dictionary of tag name and tag value pairs. Collected from _tags class attribute via nested inheritance and then any overridden and new tags from __init__ or set_tags.

load_model(model_path)[source]¶

Load a pre-trained keras model instead of fitting.

When calling this function, all functionalities can be used such as predict, predict_proba etc. with the loaded model.

Parameters:

model_pathstr (path including model name and extension): The directory where the model will be saved including the model name with a “.keras” extension. Example: model_path=”path/to/file/best_model.keras”

Returns:

None

reset(keep=None)[source]¶

Reset the object to a clean post-init state.

After a self.reset() call, self is equal or similar in value to type(self)(**self.get_params(deep=False)), assuming no other attributes were kept using keep.

Detailed behaviour:

removes any object attributes, except:: hyper-parameters (arguments of __init__) object attributes containing double-underscores, i.e., the string “__”

runs __init__ with current values of hyperparameters (result of get_params)

Not affected by the reset are:

object attributes containing double-underscores class and object methods, class attributes any attributes specified in the keep argument

Parameters:

keepNone, str, or list of str, default=None: If None, all attributes are removed except hyperparameters. If str, only the attribute with this name is kept. If list of str, only the attributes with these names are kept.

Returns:

selfobject: Reference to self.

Raises:

TypeError: If ‘keep’ is not a string or a list of strings.

save_last_model_to_file(file_path='./')[source]¶

Save the last epoch of the trained deep learning model.

Parameters:

file_pathstr, default = “./”: The directory where the model will be saved

Returns:

None

set_params(**params)¶

Set the parameters of this estimator.

The method works on simple estimators as well as on nested objects (such as Pipeline). The latter have parameters of the form <component>__<parameter> so that it’s possible to update each component of a nested object.

Parameters:

**paramsdict: Estimator parameters.

Returns:

selfestimator instance: Estimator instance.

set_tags(**tag_dict)[source]¶

Set dynamic tags to given values.

Parameters:

**tag_dictdict: Dictionary of tag name and tag value pairs.

Returns:

selfobject: Reference to self.

transform(X, y=None)[source]¶

Transform X and return a transformed version.

State required:: Requires state to be “fitted”.

Accesses in self: _is_fitted : must be True fitted model attributes (ending in “_”) : must be set, accessed by _transform

Parameters:

Xnp.ndarray or list

ynp.ndarray, default=None

1D np.array of float or str, of shape (n_cases) - class labels (ground truth) for fitting indices corresponding to instance indices in X. If None, no labels are used in fitting.

Returns:

transformed version of X