Data sets and models

Classifier models

SleepWakeClassifier

 SleepWakeClassifier (model=None, data_processor=None,
                      scaler_pipeline_name:str='scaler',
                      model_pipeline_name:str='model')

Abstract class for sleep/wake classifiers.

source

train_pipeline

 train_pipeline (classifier:__main__.SleepWakeClassifier,
                 examples_X:List[numpy.ndarray]=[],
                 examples_y:List[numpy.ndarray]=[],
                 pairs_Xy:List[Tuple[numpy.ndarray,numpy.ndarray]]=[],
                 **train_kwargs)

*Assumes data is already preprocessed using get_needed_X_y and ready to be passed to the classifier.

Returns the loss history of the model.*

	Type	Default	Details
classifier	SleepWakeClassifier		The classifier object.
examples_X	List	[]	List of input examples. If non-empty, then `examples_y` must also be provided and must have the same length.
examples_y	List	[]	List of target labels. If non-empty, then `examples_X` must also be provided and must have the same length.
pairs_Xy	List	[]	List of input-target pairs. If non-empty, then `examples_X` and `examples_y` must not be provided.
train_kwargs
Returns	List		The loss history of the model.

source

SleepWakeClassifier.train

 SleepWakeClassifier.train (examples_X:List[numpy.ndarray]=[],
                            examples_y:List[numpy.ndarray]=[], pairs_Xy:Li
                            st[Tuple[numpy.ndarray,numpy.ndarray]]=[],
                            **training_kwargs)

*Assumes data is already preprocessed using get_needed_X_y and ready to be passed to the model.

Returns the loss history of the model.*

SGD Linear Classifier

source

LinearModel

 LinearModel (value, names=None, module=None, qualname=None, type=None,
              start=1)

Defines the loss used in sklearn’s SGDClassifier which defines the linear model used for classification.

source

SGDLinearClassifier

 SGDLinearClassifier
                      (data_processor:pisces.data_sets.DataProcessor|None=
                      None, linear_model:__main__.LinearModel=<LinearModel
                      .LOGISTIC_REGRESSION: 'log_loss'>, **kwargs)

Uses Sk-Learn’s SGDCLassifier to train a model. Possible models are logistic regression, perceptron, and SVM. The SGD aspect allows for online learning, or custom training regimes through the partial_fit method. The model is trained with a balanced class weight, and uses L1 regularization. The input data is scaled with a StandardScaler before being passed to the model.

Random Forest

source

RandomForest

 RandomForest (data_processor:pisces.data_sets.DataProcessor|None=None,
               class_weight:str='balanced', **kwargs)

Interface for sklearn’s RandomForestClassifier

Mads Olsen et all classifier

We have downloaded the saved model weights from a research repository from Mads Olsen’s group, and converted those into a saved Keras model to remove the need to re-define all of the layers. This conversion process is shown in ../analyses/convert_mads_olsen_model_to_keras.ipynb.

Thus, we have a TensorFlow model that we can run inference on, and we could train it if we wanted to.

For simplicity, we are just going to run inference. One twist of our method is that the classifier is expecting two high-resolution spectrograms for inputs: 1. 3-axis Accelerometer data 2. PPG (photoplethysmogram) data

Based on visually inspecting examples from the paper, we are going to hack together an input by flipping the accelerometer data along the frequencies axis. The paper images seem to show a similarity between high-frequency accelerometer data and low-frequency PPG data. Surprisingly, this seems to work well.

source

MOResUNetPretrained

 MOResUNetPretrained
                      (data_processor:pisces.data_sets.DataProcessor|None=
                      None, model:keras.src.models.model.Model|None=None,
                      lazy_model_loading:bool=True,
                      initial_lr:float=1e-05, validation_split:float=0.1,
                      epochs:int=10, batch_size:int=1, **kwargs)

Abstract class for sleep/wake classifiers.

Training tools

source

run_splits

 run_splits (split_maker:__main__.SplitMaker,
             data_processor:pisces.data_sets.DataProcessor,
             swc_class:Type[__main__.SleepWakeClassifier],
             epochs:int|None, exclude:List[str]=[],
             linear_model:__main__.LinearModel|None=None)

source

run_split

 run_split (train_indices,
            preprocessed_data_set:List[Tuple[numpy.ndarray,numpy.ndarray]]
            , swc:__main__.SleepWakeClassifier, epochs:int)

source

LeaveOneOutSplitter

 LeaveOneOutSplitter ()

Initialize self. See help(type(self)) for accurate signature.

source

SplitMaker

 SplitMaker ()

Initialize self. See help(type(self)) for accurate signature.