TSPulseClassifier

class TSPulseClassifier(model_path='ibm-granite/granite-timeseries-tspulse-r1', revision='tspulse-block-dualhead-512-p16-r1', config=None, context_length=512, scaling=True, batch_size=32, epochs=1, learning_rate=0.001, freeze_backbone=True, train_val_split=0.0, device='auto', seed=42)[source]

Time series classifier wrapping IBM TSPulse via granite-tsfm.

Loads a pretrained TSPulse backbone from Hugging Face, freezes most backbone weights, and fine-tunes the classification head and patch-embedding layers on the training data.

Implementation adapted from [1].

Parameters:

model_pathstr, default=”ibm-granite/granite-timeseries-tspulse-r1”

Hugging Face model id or local path.

revisionstr, default=”tspulse-block-dualhead-512-p16-r1”

Model revision on the Hugging Face Hub.

configdict, optional, default=None

Extra keyword arguments forwarded to tsfm_public.models.tspulse.TSPulseForClassification.from_pretrained.

Any key you supply overrides the built-in default for that key, except for the following two keys which are always inferred from the training data:

num_input_channels: set to the number of channels in X
num_targets: set to the number of distinct labels in y

If config is None, then following default overrides are applied:

head_gated_attention_activation="softmax": gated attention activation in the classification head
channel_virtual_expand_scale=2: virtual expansion factor for channel mixing in the decoder/head
mask_ratio=0.3: fraction of patches masked during training (use 0 to disable)
head_reduce_d_model=1: reduction factor for model dimension in the head
disable_mask_in_classification_eval=True: disables masking at evaluation/prediction time
fft_time_consistent_masking=True: uses masked time-series for FFT during training
decoder_mode="mix_channel": decoder channel mixing mode (alternative: "common_channel")
head_aggregation_dim="patch": aggregation dimension used by the head
head_aggregation=None: use model-default aggregation module
loss="cross_entropy": loss for classification fine-tuning
ignore_mismatched_sizes=True: allows loading when head shapes differ from the checkpoint

context_lengthint, default=512

Series length passed to the preprocessor and dataset.

scalingbool, default=True

Whether to scale input channels in the preprocessor.

batch_sizeint, default=32

Training batch size.

epochsint, default=1

Number of fine-tuning epochs.

learning_ratefloat, optional, default=1e-3

Optimizer learning rate for fine-tuning.

freeze_backbonebool, default=True

If True, freeze backbone weights except patch-embedding layers.

train_val_splitfloat, default=0.0

Fraction of training data held out for validation during fine-tuning. 0.0 uses all training data.

devicestr, default=”auto”

PyTorch device ("auto", "cpu", "cuda", "mps").

seedint, default=42

Random seed for training.

Attributes:

is_fitted: Whether fit has been called.

References

[1]

https://github.com/ibm-granite/granite-tsfm/blob/main/notebooks/hfdemo/tspulse_classification.ipynb

[2]

https://arxiv.org/abs/2505.13033

Examples

>>> from sktime.classification.foundation_models.tspulse import TSPulseClassifier
>>> from sktime.datasets import load_unit_test
>>> X_train, y_train = load_unit_test(split="train", return_type="nested_univ")
>>> X_test, _ = load_unit_test(split="test", return_type="nested_univ")
>>> clf = TSPulseClassifier(epochs=1, batch_size=8)
>>> clf.fit(X_train, y_train)
>>> y_pred = clf.predict(X_test)

Methods

`check_is_fitted`([method_name])	Check if the estimator has been fitted.
`clone`()	Obtain a clone of the object with same hyper-parameters and config.
`clone_tags`(estimator[, tag_names])	Clone tags from another object as dynamic override.
`create_test_instance`([parameter_set])	Construct an instance of the class, using first test parameter set.
`create_test_instances_and_names`([parameter_set])	Create list of all test instances and a list of names for them.
`fit`(X, y)	Fit time series classifier to training data.
`fit_predict`(X, y[, cv, change_state])	Fit and predict labels for sequences in X.
`fit_predict_proba`(X, y[, cv, change_state])	Fit and predict labels probabilities for sequences in X.
`get_class_tag`(tag_name[, tag_value_default])	Get class tag value from class, with tag level inheritance from parents.
`get_class_tags`()	Get class tags from class, with tag level inheritance from parent classes.
`get_config`()	Get config flags for self.
`get_fitted_params`([deep])	Get fitted parameters.
`get_param_defaults`()	Get object's parameter defaults.
`get_param_names`([sort])	Get object's parameter names.
`get_params`([deep])	Get a dict of parameters values for this object.
`get_tag`(tag_name[, tag_value_default, ...])	Get tag value from instance, with tag level inheritance and overrides.
`get_tags`()	Get tags from instance, with tag level inheritance and overrides.
`get_test_params`([parameter_set])	Return testing parameter settings for the estimator.
`is_composite`()	Check if the object is composed of other BaseObjects.
`load_from_path`(serial)	Load object from file location.
`load_from_serial`(serial)	Load object from serialized memory container.
`predict`(X)	Predicts labels for sequences in X.
`predict_proba`(X)	Predicts labels probabilities for sequences in X.
`reset`()	Reset the object to a clean post-init state.
`save`([path, serialization_format])	Save serialized self to bytes-like object or to (.zip) file.
`score`(X, y)	Scores predicted labels against ground truth labels on X.
`set_config`(**config_dict)	Set config flags to given values.
`set_params`(**params)	Set the parameters of this object.
`set_random_state`([random_state, deep, ...])	Set random_state pseudo-random seed parameters for self.
`set_tags`(**tag_dict)	Set instance level tag overrides to given values.