rapidtide 3.0.11__py3-none-any.whl → 3.1__py3-none-any.whl

rapidtide/dlfiltertorch.py

@@ -0,0 +1,4843 @@
+#!/usr/bin/env python
+# -*- coding: utf-8 -*-
+#
+#   Copyright 2016-2025 Blaise Frederick
+#
+#   Licensed under the Apache License, Version 2.0 (the "License");
+#   you may not use this file except in compliance with the License.
+#   You may obtain a copy of the License at
+#
+#       http://www.apache.org/licenses/LICENSE-2.0
+#
+#   Unless required by applicable law or agreed to in writing, software
+#   distributed under the License is distributed on an "AS IS" BASIS,
+#   WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#   See the License for the specific language governing permissions and
+#   limitations under the License.
+#
+#
+import glob
+import logging
+import os
+import sys
+import warnings
+
+import matplotlib as mpl
+import matplotlib.pyplot as plt
+import numpy as np
+import tqdm
+
+with warnings.catch_warnings():
+    warnings.simplefilter("ignore")
+    try:
+        import pyfftw
+    except ImportError:
+        pyfftwpresent = False
+    else:
+        pyfftwpresent = True
+
+from scipy import fftpack
+from statsmodels.robust.scale import mad
+
+if pyfftwpresent:
+    fftpack = pyfftw.interfaces.scipy_fftpack
+    pyfftw.interfaces.cache.enable()
+
+import torch
+import torch.nn as nn
+import torch.optim as optim
+from torch.utils.data import DataLoader, TensorDataset
+
+import rapidtide.io as tide_io
+
+LGR = logging.getLogger("GENERAL")
+LGR.debug("setting backend to Agg")
+mpl.use("Agg")
+
+# Disable GPU if desired
+if torch.cuda.is_available():
+    device = torch.device("cuda")
+    LGR.info(f"Using CUDA device: {torch.cuda.get_device_name(0)}")
+elif torch.backends.mps.is_available():
+    device = torch.device("mps")
+    LGR.info(f"Using MPS device")
+else:
+    device = torch.device("cpu")
+    LGR.info("Using CPU")
+
+LGR.debug(f"pytorch version: >>>{torch.__version__}<<<")
+
+
+class DeepLearningFilter:
+    """Base class for deep learning filter"""
+
+    thesuffix = "sliceres"
+    thedatadir = "/Users/frederic/Documents/MR_data/physioconn/timecourses"
+    inputfrag = "abc"
+    targetfrag = "xyz"
+    namesuffix = None
+    modelroot = "."
+    excludethresh = 4.0
+    modelname = None
+    intermediatemodelpath = None
+    usebadpts = False
+    activation = "tanh"
+    dofft = False
+    readlim = None
+    countlim = None
+    lossfilename = None
+    train_x = None
+    train_y = None
+    val_x = None
+    val_y = None
+    model = None
+    modelpath = None
+    inputsize = None
+    infodict = {}
+
+    def __init__(
+        self,
+        window_size: int = 128,
+        num_layers: int = 5,
+        dropout_rate: float = 0.3,
+        num_pretrain_epochs: int = 0,
+        num_epochs: int = 1,
+        activation: str = "relu",
+        modelroot: str = ".",
+        dofft: bool = False,
+        excludethresh: float = 4.0,
+        usebadpts: bool = False,
+        thesuffix: str = "25.0Hz",
+        modelpath: str = ".",
+        thedatadir: str = "/Users/frederic/Documents/MR_data/physioconn/timecourses",
+        inputfrag: str = "abc",
+        targetfrag: str = "xyz",
+        corrthresh: float = 0.5,
+        excludebysubject: bool = True,
+        startskip: int = 200,
+        endskip: int = 200,
+        step: int = 1,
+        namesuffix: str | None = None,
+        readlim: int | None = None,
+        readskip: int | None = None,
+        countlim: int | None = None,
+        **kwargs,
+    ) -> None:
+        """
+        Initialize the DeepLearningFilter with specified parameters.
+
+        This constructor sets up the configuration for a deep learning model used
+        for filtering physiological timecourses. It initializes various hyperparameters,
+        paths, and flags that control the behavior of the model and data processing.
+
+        Parameters
+        ----------
+        window_size : int, optional
+            Size of the sliding window used for processing time series data. Default is 128.
+        num_layers : int, optional
+            Number of layers in the neural network model. Default is 5.
+        dropout_rate : float, optional
+            Dropout rate for regularization during training. Default is 0.3.
+        num_pretrain_epochs : int, optional
+            Number of pre-training epochs. Default is 0.
+        num_epochs : int, optional
+            Number of training epochs. Default is 1.
+        activation : str, optional
+            Activation function to use in the model. Default is "relu".
+        modelroot : str, optional
+            Root directory for model storage. Default is ".".
+        dofft : bool, optional
+            Whether to apply FFT transformation to input data. Default is False.
+        excludethresh : float, optional
+            Threshold for excluding data points based on correlation. Default is 4.0.
+        usebadpts : bool, optional
+            Whether to include bad points in the input. Default is False.
+        thesuffix : str, optional
+            Suffix to append to filenames. Default is "25.0Hz".
+        modelpath : str, optional
+            Path to save or load the model. Default is ".".
+        thedatadir : str, optional
+            Directory containing the physiological data files. Default is
+            "/Users/frederic/Documents/MR_data/physioconn/timecourses".
+        inputfrag : str, optional
+            Fragment identifier for input data. Default is "abc".
+        targetfrag : str, optional
+            Fragment identifier for target data. Default is "xyz".
+        corrthresh : float, optional
+            Correlation threshold for filtering. Default is 0.5.
+        excludebysubject : bool, optional
+            Whether to exclude data by subject. Default is True.
+        startskip : int, optional
+            Number of samples to skip at the beginning of each timecourse. Default is 200.
+        endskip : int, optional
+            Number of samples to skip at the end of each timecourse. Default is 200.
+        step : int, optional
+            Step size for sliding window. Default is 1.
+        namesuffix : str, optional
+            Suffix to append to model name. Default is None.
+        readlim : int, optional
+            Limit on number of samples to read. Default is None.
+        readskip : int, optional
+            Number of samples to skip when reading data. Default is None.
+        countlim : int, optional
+            Limit on number of timecourses to process. Default is None.
+        **kwargs
+            Additional keyword arguments passed to the parent class.
+
+        Notes
+        -----
+        The `inputsize` is dynamically set based on the `usebadpts` flag:
+        - If `usebadpts` is True, input size is 2.
+        - Otherwise, input size is 1.
+
+        Examples
+        --------
+        >>> filter = DeepLearningFilter(
+        ...     window_size=256,
+        ...     num_layers=6,
+        ...     dropout_rate=0.2,
+        ...     modelroot="/models",
+        ...     dofft=True
+        ... )
+        """
+        self.window_size = window_size
+        self.dropout_rate = dropout_rate
+        self.num_pretrain_epochs = num_pretrain_epochs
+        self.num_epochs = num_epochs
+        self.usebadpts = usebadpts
+        self.num_layers = num_layers
+        if self.usebadpts:
+            self.inputsize = 2
+        else:
+            self.inputsize = 1
+        self.activation = activation
+        self.modelroot = modelroot
+        self.dofft = dofft
+        self.thesuffix = thesuffix
+        self.thedatadir = thedatadir
+        self.modelpath = modelpath
+        LGR.info(f"modeldir from DeepLearningFilter: {self.modelpath}")
+        self.corrthresh = corrthresh
+        self.excludethresh = excludethresh
+        self.readlim = readlim
+        self.readskip = readskip
+        self.countlim = countlim
+        self.model = None
+        self.initialized = False
+        self.trained = False
+        self.usetensorboard = False
+        self.inputfrag = inputfrag
+        self.targetfrag = targetfrag
+        self.namesuffix = namesuffix
+        self.startskip = startskip
+        self.endskip = endskip
+        self.step = step
+        self.excludebysubject = excludebysubject
+        self.device = device
+
+        # populate infodict
+        self.infodict["window_size"] = self.window_size
+        self.infodict["usebadpts"] = self.usebadpts
+        self.infodict["dofft"] = self.dofft
+        self.infodict["corrthresh"] = self.corrthresh
+        self.infodict["excludethresh"] = self.excludethresh
+        self.infodict["num_pretrain_epochs"] = self.num_pretrain_epochs
+        self.infodict["num_epochs"] = self.num_epochs
+        self.infodict["modelname"] = self.modelname
+        self.infodict["dropout_rate"] = self.dropout_rate
+        self.infodict["startskip"] = self.startskip
+        self.infodict["endskip"] = self.endskip
+        self.infodict["step"] = self.step
+        self.infodict["train_arch"] = sys.platform
+
+    def loaddata(self) -> None:
+        """
+        Load and preprocess data for training and validation.
+
+        This method initializes the data loading process by calling the `prep` function
+        with a set of parameters derived from the instance attributes. It handles both
+        FFT and non-FFT modes of data preprocessing. The loaded data is stored in
+        instance variables for use in subsequent training steps.
+
+        Parameters
+        ----------
+        self : object
+            The instance of the class containing the following attributes:
+            - initialized : bool
+                Indicates whether the model has been initialized.
+            - dofft : bool
+                Whether to apply FFT transformation to the data.
+            - window_size : int
+                Size of the sliding window used for data segmentation.
+            - thesuffix : str
+                Suffix to append to filenames when reading data.
+            - thedatadir : str
+                Directory path where the data files are located.
+            - inputfrag : str
+                Fragment identifier for input data.
+            - targetfrag : str
+                Fragment identifier for target data.
+            - startskip : int
+                Number of samples to skip at the beginning of each file.
+            - endskip : int
+                Number of samples to skip at the end of each file.
+            - corrthresh : float
+                Correlation threshold for filtering data.
+            - step : int
+                Step size for sliding window.
+            - usebadpts : bool
+                Whether to include bad points in the data.
+            - excludethresh : float
+                Threshold for excluding data points.
+            - excludebysubject : bool
+                Whether to exclude data by subject.
+            - readlim : int
+                Limit on the number of samples to read.
+            - readskip : int
+                Number of samples to skip while reading.
+            - countlim : int
+                Limit on the number of data points to process.
+
+        Returns
+        -------
+        None
+            This method does not return any value. It modifies the instance attributes
+            in place.
+
+        Raises
+        ------
+        Exception
+            If the model is not initialized prior to calling this method.
+
+        Notes
+        -----
+        The method assigns the following attributes to the instance after loading:
+        - train_x : array-like
+            Training input data.
+        - train_y : array-like
+            Training target data.
+        - val_x : array-like
+            Validation input data.
+        - val_y : array-like
+            Validation target data.
+        - Ns : int
+            Number of samples.
+        - tclen : int
+            Length of time series.
+        - thebatchsize : int
+            Batch size for training.
+
+        Examples
+        --------
+        >>> model = MyModel()
+        >>> model.initialized = True
+        >>> model.loaddata()
+        >>> print(model.train_x.shape)
+        (1000, 10)
+        """
+        if not self.initialized:
+            raise Exception("model must be initialized prior to loading data")
+
+        if self.dofft:
+            (
+                self.train_x,
+                self.train_y,
+                self.val_x,
+                self.val_y,
+                self.Ns,
+                self.tclen,
+                self.thebatchsize,
+                dummy,
+                dummy,
+            ) = prep(
+                self.window_size,
+                thesuffix=self.thesuffix,
+                thedatadir=self.thedatadir,
+                inputfrag=self.inputfrag,
+                targetfrag=self.targetfrag,
+                startskip=self.startskip,
+                endskip=self.endskip,
+                corrthresh=self.corrthresh,
+                step=self.step,
+                dofft=self.dofft,
+                usebadpts=self.usebadpts,
+                excludethresh=self.excludethresh,
+                excludebysubject=self.excludebysubject,
+                readlim=self.readlim,
+                readskip=self.readskip,
+                countlim=self.countlim,
+            )
+        else:
+            (
+                self.train_x,
+                self.train_y,
+                self.val_x,
+                self.val_y,
+                self.Ns,
+                self.tclen,
+                self.thebatchsize,
+            ) = prep(
+                self.window_size,
+                thesuffix=self.thesuffix,
+                thedatadir=self.thedatadir,
+                inputfrag=self.inputfrag,
+                targetfrag=self.targetfrag,
+                startskip=self.startskip,
+                endskip=self.endskip,
+                corrthresh=self.corrthresh,
+                step=self.step,
+                dofft=self.dofft,
+                usebadpts=self.usebadpts,
+                excludethresh=self.excludethresh,
+                excludebysubject=self.excludebysubject,
+                readlim=self.readlim,
+                readskip=self.readskip,
+                countlim=self.countlim,
+            )
+
+    def predict_model(self, X: np.ndarray) -> np.ndarray:
+        """
+        Make predictions using the trained model.
+
+        Parameters
+        ----------
+        X : np.ndarray
+            Input features for prediction. Shape should be (n_samples, n_features)
+            where n_samples is the number of samples and n_features is the number
+            of features expected by the model.
+
+        Returns
+        -------
+        np.ndarray
+            Model predictions. Shape will depend on the specific model type but
+            typically follows (n_samples,) for regression or (n_samples, n_classes)
+            for classification.
+
+        Notes
+        -----
+        This method sets the model to inference mode by calling with training=False.
+        The predictions are made without computing gradients, making it efficient
+        for inference tasks. Input data is automatically converted to PyTorch tensors
+        and moved to the appropriate device. Special handling is included for
+        tensor dimension permutation to match model expectations.
+
+        Examples
+        --------
+        >>> # Assuming model is already trained
+        >>> X_test = np.array([[1.0, 2.0], [3.0, 4.0]])
+        >>> predictions = model.predict_model(X_test)
+        >>> print(predictions)
+        """
+        self.model.eval()
+        with torch.no_grad():
+            if isinstance(X, np.ndarray):
+                X = torch.from_numpy(X).float().to(self.device)
+            # PyTorch expects (batch, channels, length) but we have (batch, length, channels)
+            X = X.permute(0, 2, 1)
+            output = self.model(X)
+            # Convert back to (batch, length, channels)
+            output = output.permute(0, 2, 1)
+            return output.cpu().numpy()
+
+    def evaluate(self) -> tuple[list, list, float, float]:
+        """
+        Evaluate the model performance on validation data and compute loss metrics.
+
+        This method performs model evaluation by computing prediction errors and
+        saving training/validation loss curves. It calculates both prediction error
+        (difference between predicted and actual values) and raw error (difference
+        between input and actual values). The method also generates and saves a
+        plot of the training and validation loss over epochs.
+
+        Parameters
+        ----------
+        self : object
+            The instance of the class containing the model and data attributes.
+
+        Returns
+        -------
+        tuple[list, list, float, float]
+            A tuple containing:
+            - training_loss : list
+                List of training loss values per epoch
+            - validation_loss : list
+                List of validation loss values per epoch
+            - prediction_error : float
+                Mean squared error between predicted and actual values
+            - raw_error : float
+                Mean squared error between input features and actual values
+
+        Notes
+        -----
+        This method modifies the instance attributes:
+        - self.lossfilename: Path to the saved loss plot
+        - self.pred_error: Computed prediction error
+        - self.raw_error: Computed raw error
+        - self.loss: Training loss history
+        - self.val_loss: Validation loss history
+
+        The method saves:
+        - Loss plot as PNG file
+        - Loss metrics as text file
+
+        Examples
+        --------
+        >>> model = MyModel()
+        >>> train_loss, val_loss, pred_error, raw_error = model.evaluate()
+        >>> print(f"Prediction Error: {pred_error}")
+        Prediction Error: 0.1234
+        """
+        self.lossfilename = os.path.join(self.modelname, "loss.png")
+        LGR.info(f"lossfilename: {self.lossfilename}")
+
+        YPred = self.predict_model(self.val_x)
+
+        error = self.val_y - YPred
+        self.pred_error = np.mean(np.square(error))
+
+        error2 = self.val_x - self.val_y
+        self.raw_error = np.mean(np.square(error2))
+        LGR.info(f"Prediction Error: {self.pred_error}\tRaw Error: {self.raw_error}")
+
+        f = open(os.path.join(self.modelname, "loss.txt"), "w")
+        f.write(
+            self.modelname
+            + ": Prediction Error: "
+            + str(self.pred_error)
+            + " Raw Error: "
+            + str(self.raw_error)
+            + "\n"
+        )
+        f.close()
+
+        epochs = range(len(self.loss))
+
+        self.updatemetadata()
+
+        plt.figure()
+        plt.plot(epochs, self.loss, "bo", label="Training loss")
+        plt.plot(epochs, self.val_loss, "b", label="Validation loss")
+        plt.title("Training and validation loss")
+        plt.legend()
+        plt.savefig(self.lossfilename)
+        plt.close()
+
+        return self.loss, self.val_loss, self.pred_error, self.raw_error
+
+    def initmetadata(self) -> None:
+        """
+        Initialize and store metadata information for the model.
+
+        This function creates a dictionary containing various model configuration parameters
+        and writes them to a JSON file for future reference and reproducibility.
+
+        Parameters
+        ----------
+        self : object
+            The instance of the class containing the metadata attributes. Expected to have
+            the following attributes:
+            - `nettype`: Type of neural network
+            - `window_size`: Size of the window used for processing
+            - `usebadpts`: Flag indicating whether bad points are handled
+            - `dofft`: Flag indicating whether FFT is used
+            - `excludethresh`: Threshold for exclusion
+            - `num_epochs`: Number of training epochs
+            - `num_layers`: Number of layers in the model
+            - `dropout_rate`: Dropout rate for regularization
+            - `modelname`: Name of the model
+
+        Returns
+        -------
+        None
+            This function does not return any value but writes metadata to a JSON file.
+
+        Notes
+        -----
+        The metadata includes:
+        - Window size for processing
+        - Bad point handling flag
+        - FFT usage flag
+        - Exclusion threshold
+        - Number of epochs and layers
+        - Dropout rate
+        - Operating system platform
+        - Model name
+
+        The metadata is saved to ``{modelname}/model_meta.json`` where ``modelname``
+        is the model's name attribute.
+
+        Examples
+        --------
+        >>> model = MyModel()
+        >>> model.initmetadata()
+        >>> # Metadata stored in modelname/model_meta.json
+        """
+
+        self.infodict = {}
+        self.infodict["nettype"] = self.nettype
+        self.infodict["window_size"] = self.window_size
+        self.infodict["usebadpts"] = self.usebadpts
+        self.infodict["dofft"] = self.dofft
+        self.infodict["excludethresh"] = self.excludethresh
+        self.infodict["num_epochs"] = self.num_epochs
+        self.infodict["num_layers"] = self.num_layers
+        self.infodict["dropout_rate"] = self.dropout_rate
+        self.infodict["train_arch"] = sys.platform
+        self.infodict["modelname"] = self.modelname
+        tide_io.writedicttojson(self.infodict, os.path.join(self.modelname, "model_meta.json"))
+
+    def updatemetadata(self) -> None:
+        """
+        Update metadata dictionary with model metrics and save to JSON file.
+
+        This method updates the internal information dictionary with various model
+        performance metrics and writes the complete metadata to a JSON file for
+        model persistence and tracking.
+
+        Parameters
+        ----------
+        self : object
+            The instance of the class containing the metadata and model information.
+            Expected to have the following attributes:
+            - infodict : dict
+                Dictionary containing model metadata.
+            - loss : float
+                Training loss value.
+            - val_loss : float
+                Validation loss value.
+            - raw_error : float
+                Raw error metric.
+            - pred_error : float
+                Prediction error metric.
+            - modelname : str
+                Name/path of the model for file output.
+
+        Returns
+        -------
+        None
+            This method does not return any value but modifies the `infodict` in-place
+            and writes to a JSON file.
+
+        Notes
+        -----
+        The method writes metadata to ``{modelname}/model_meta.json`` where
+        ``modelname`` is the model name attribute of the instance.
+
+        Examples
+        --------
+        >>> model = MyModel()
+        >>> model.updatemetadata()
+        >>> # Creates model_meta.json with loss, val_loss, raw_error, and pred_error
+        """
+        self.infodict["loss"] = self.loss
+        self.infodict["val_loss"] = self.val_loss
+        self.infodict["raw_error"] = self.raw_error
+        self.infodict["prediction_error"] = self.pred_error
+        tide_io.writedicttojson(self.infodict, os.path.join(self.modelname, "model_meta.json"))
+
+    def savemodel(self, altname: str | None = None) -> None:
+        """
+        Save the model to disk with the specified name.
+
+        This method saves the current model to a Keras file format (.keras) in a
+        directory named according to the model name or an alternative name provided.
+
+        Parameters
+        ----------
+        altname : str, optional
+            Alternative name to use for saving the model. If None, uses the
+            model's default name stored in `self.modelname`. Default is None.
+
+        Returns
+        -------
+        None
+            This method does not return any value.
+
+        Notes
+        -----
+        The model is saved in the Keras format (.keras) and stored in a directory
+        with the same name as the model. The method logs the saving operation
+        using the logger instance `LGR`.
+
+        Examples
+        --------
+        >>> # Save model with default name
+        >>> savemodel()
+        >>>
+        >>> # Save model with alternative name
+        >>> savemodel(altname="my_custom_model")
+        """
+        if altname is None:
+            modelsavename = self.modelname
+        else:
+            modelsavename = altname
+        LGR.info(f"saving {modelsavename}")
+        torch.save(
+            {
+                "model_state_dict": self.model.state_dict(),
+                "model_config": (
+                    self.model.get_config() if hasattr(self.model, "get_config") else None
+                ),
+            },
+            os.path.join(modelsavename, "model.pth"),
+        )
+
+    def loadmodel(self, modelname: str, verbose: bool = False) -> None:
+        """
+        Load a trained model from disk and initialize model parameters.
+
+        Load a Keras model from the specified model directory, along with associated
+        metadata and configuration information. The function attempts to load the model
+        in Keras format first, falling back to HDF5 format if the Keras format is not found.
+
+        Parameters
+        ----------
+        modelname : str
+            Name of the model to load, corresponding to a subdirectory in ``self.modelpath``.
+        verbose : bool, optional
+            If True, print model summary and metadata information. Default is False.
+
+        Returns
+        -------
+        None
+            This method modifies the instance attributes in-place and does not return anything.
+
+        Notes
+        -----
+        The function attempts to load the model in the following order:
+        1. Keras format (model.keras)
+        2. HDF5 format (model.h5)
+
+        If neither format is found, the function exits with an error message.
+
+        The loaded model metadata is stored in ``self.infodict``, and model configuration
+        is stored in ``self.config``. Additional attributes like ``window_size`` and
+        ``usebadpts`` are extracted from the metadata and stored as instance attributes.
+
+        Examples
+        --------
+        >>> loader = ModelLoader()
+        >>> loader.loadmodel("my_model", verbose=True)
+        loading my_model
+        Model: "sequential"
+        _________________________________________________________________
+        Layer (type)                Output Shape              Param #
+        =================================================================
+        ...
+        >>> print(loader.window_size)
+        100
+        """
+        # read in the data
+        LGR.info(f"loading {modelname}")
+
+        # load additional information first to reconstruct model
+        self.infodict = tide_io.readdictfromjson(
+            os.path.join(self.modelpath, modelname, "model_meta.json")
+        )
+        if verbose:
+            print(self.infodict)
+        self.window_size = self.infodict["window_size"]
+        self.usebadpts = self.infodict["usebadpts"]
+
+        # Load the model as a dict
+        checkpoint = torch.load(
+            os.path.join(self.modelpath, modelname, "model.pth"), map_location=self.device
+        )
+
+        # Reconstruct the model architecture (must be done by subclass)
+        if self.infodict["nettype"] == "cnn":
+            self.num_filters = checkpoint["model_config"]["num_filters"]
+            self.kernel_size = checkpoint["model_config"]["kernel_size"]
+            self.num_layers = checkpoint["model_config"]["num_layers"]
+            self.dropout_rate = checkpoint["model_config"]["dropout_rate"]
+            self.dilation_rate = checkpoint["model_config"]["dilation_rate"]
+            self.activation = checkpoint["model_config"]["activation"]
+            self.inputsize = checkpoint["model_config"]["inputsize"]
+
+            self.model = CNNModel(
+                self.num_filters,
+                self.kernel_size,
+                self.num_layers,
+                self.dropout_rate,
+                self.dilation_rate,
+                self.activation,
+                self.inputsize,
+            )
+        elif self.infodict["nettype"] == "autoencoder":
+            self.encoding_dim = checkpoint["model_config"]["encoding_dim"]
+            self.num_layers = checkpoint["model_config"]["num_layers"]
+            self.dropout_rate = checkpoint["model_config"]["dropout_rate"]
+            self.activation = checkpoint["model_config"]["activation"]
+            self.inputsize = checkpoint["model_config"]["inputsize"]
+
+            self.model = DenseAutoencoderModel(
+                self.window_size,
+                self.encoding_dim,
+                self.num_layers,
+                self.dropout_rate,
+                self.activation,
+                self.inputsize,
+            )
+        elif self.infodict["nettype"] == "multiscalecnn":
+            self.num_filters = checkpoint["model_config"]["num_filters"]
+            self.kernel_sizes = checkpoint["model_config"]["kernel_sizes"]
+            self.input_lens = checkpoint["model_config"]["input_lens"]
+            self.input_width = checkpoint["model_config"]["input_width"]
+            self.dilation_rate = checkpoint["model_config"]["dilation_rate"]
+
+            self.model = MultiscaleCNNModel(
+                self.num_filters,
+                self.kernel_sizes,
+                self.input_lens,
+                self.input_width,
+                self.dilation_rate,
+            )
+        elif self.infodict["nettype"] == "convautoencoder":
+            self.encoding_dim = checkpoint["model_config"]["encoding_dim"]
+            self.num_filters = checkpoint["model_config"]["num_filters"]
+            self.kernel_size = checkpoint["model_config"]["kernel_size"]
+            self.dropout_rate = checkpoint["model_config"]["dropout_rate"]
+            self.activation = checkpoint["model_config"]["activation"]
+            self.inputsize = checkpoint["model_config"]["inputsize"]
+
+            self.model = ConvAutoencoderModel(
+                self.window_size,
+                self.encoding_dim,
+                self.num_filters,
+                self.kernel_size,
+                self.dropout_rate,
+                self.activation,
+                self.inputsize,
+            )
+        elif self.infodict["nettype"] == "crnn":
+            self.num_filters = checkpoint["model_config"]["num_filters"]
+            self.kernel_size = checkpoint["model_config"]["kernel_size"]
+            self.encoding_dim = checkpoint["model_config"]["encoding_dim"]
+            self.dropout_rate = checkpoint["model_config"]["dropout_rate"]
+            self.activation = checkpoint["model_config"]["activation"]
+            self.inputsize = checkpoint["model_config"]["inputsize"]
+
+            self.model = CRNNModel(
+                self.num_filters,
+                self.kernel_size,
+                self.encoding_dim,
+                self.dropout_rate,
+                self.activation,
+                self.inputsize,
+            )
+        elif self.infodict["nettype"] == "lstm":
+            self.num_units = checkpoint["model_config"]["num_units"]
+            self.num_layers = checkpoint["model_config"]["num_layers"]
+            self.dropout_rate = checkpoint["model_config"]["dropout_rate"]
+            self.inputsize = checkpoint["model_config"]["inputsize"]
+
+            self.model = LSTMModel(
+                self.num_units,
+                self.num_layers,
+                self.dropout_rate,
+                self.window_size,
+                self.inputsize,
+            )
+        elif self.infodict["nettype"] == "hybrid":
+            self.num_filters = checkpoint["model_config"]["num_filters"]
+            self.kernel_size = checkpoint["model_config"]["kernel_size"]
+            self.num_units = checkpoint["model_config"]["num_units"]
+            self.num_layers = checkpoint["model_config"]["num_layers"]
+            self.dropout_rate = checkpoint["model_config"]["dropout_rate"]
+            self.activation = checkpoint["model_config"]["activation"]
+            self.inputsize = checkpoint["model_config"]["inputsize"]
+            self.invert = checkpoint["model_config"]["invert"]
+
+            self.model = HybridModel(
+                self.num_filters,
+                self.kernel_size,
+                self.num_units,
+                self.num_layers,
+                self.dropout_rate,
+                self.activation,
+                self.inputsize,
+                self.window_size,
+                self.invert,
+            )
+        else:
+            print(f"nettype {self.infodict['nettype']} is not supported!")
+            sys.exit()
+
+        self.model.load_state_dict(checkpoint["model_state_dict"])
+        self.model.to(self.device)
+
+    def initialize(self) -> None:
+        """
+        Initialize the model by setting up network architecture and metadata.
+
+        This method performs a series of initialization steps including retrieving
+        the model name, creating the network architecture, displaying model summary,
+        saving the model configuration, initializing metadata, and setting appropriate
+        flags to indicate initialization status.
+
+        Parameters
+        ----------
+        self : object
+            The instance of the model class being initialized.
+
+        Returns
+        -------
+        None
+            This method does not return any value.
+
+        Notes
+        -----
+        This method should be called before any training or prediction operations.
+        The initialization process sets `self.initialized` to True and `self.trained`
+        to False, indicating that the model is ready for training but has not been
+        trained yet.
+
+        Examples
+        --------
+        >>> model = MyModel()
+        >>> model.initialize()
+        >>> print(model.initialized)
+        True
+        >>> print(model.trained)
+        False
+        """
+        self.getname()
+        self.makenet()
+        print(self.model)
+        self.savemodel()
+        self.initmetadata()
+        self.initialized = True
+        self.trained = False
+
+    def train(self) -> None:
+        """
+        Train the model using the provided training and validation datasets.
+
+        This method performs model training with optional pretraining and logging. It supports
+        TensorBoard logging, model checkpointing, early stopping, and NaN termination. The trained
+        model is saved at the end of training.
+
+        Parameters
+        ----------
+        self : object
+            The instance of the class containing the model and training configuration.
+            Expected attributes include:
+            - `model`: PyTorch model to be trained.
+            - `train_x`, `train_y`, `val_x`, `val_y`: Training and validation data as numpy arrays.
+            - `device`: Device to run the training on (e.g., 'cpu' or 'cuda').
+            - `num_pretrain_epochs`: Number of pretraining epochs (default: 0).
+            - `num_epochs`: Number of main training epochs.
+            - `modelname`: Directory name to save model checkpoints and logs.
+            - `usetensorboard`: Boolean flag to enable TensorBoard logging.
+            - `savemodel()`: Method to save the final trained model.
+
+        Returns
+        -------
+        None
+            This function does not return any value.
+
+        Notes
+        -----
+        - If `self.usetensorboard` is True, TensorBoard logging is enabled.
+        - If `self.num_pretrain_epochs` is greater than 0, a pretraining phase is performed
+          before the main training loop.
+        - The model is saved after training using the `savemodel()` method.
+        - Training uses `ModelCheckpoint`, `EarlyStopping`, and `TerminateOnNaN` callbacks
+          to manage training process and prevent overfitting or NaN issues.
+        - Intermediate model checkpoints are saved during training.
+        - The best model (based on validation loss) is retained and restored upon early stopping.
+
+        Examples
+        --------
+        >>> trainer = ModelTrainer(model, train_x, train_y, val_x, val_y)
+        >>> trainer.train()
+        """
+        self.model.train()
+        self.model.to(self.device)
+
+        # Convert numpy arrays to PyTorch tensors and transpose for Conv1d
+        print("converting tensors")
+        train_x_tensor = torch.from_numpy(self.train_x).float().permute(0, 2, 1)
+        train_y_tensor = torch.from_numpy(self.train_y).float().permute(0, 2, 1)
+        val_x_tensor = torch.from_numpy(self.val_x).float().permute(0, 2, 1)
+        val_y_tensor = torch.from_numpy(self.val_y).float().permute(0, 2, 1)
+
+        print("setting data")
+        train_dataset = TensorDataset(train_x_tensor, train_y_tensor)
+        val_dataset = TensorDataset(val_x_tensor, val_y_tensor)
+
+        train_loader = DataLoader(train_dataset, batch_size=1024, shuffle=True)
+        val_loader = DataLoader(val_dataset, batch_size=1024, shuffle=False)
+
+        print("setting criterion")
+        criterion = nn.MSELoss()
+
+        print("setting optimizer")
+        optimizer = optim.RMSprop(self.model.parameters())
+
+        self.loss = []
+        self.val_loss = []
+
+        best_val_loss = float("inf")
+        patience = 10
+        patience_counter = 0
+
+        total_epochs = self.num_pretrain_epochs + self.num_epochs
+
+        for epoch in range(total_epochs):
+            print(f"Epoch {epoch+1}/{total_epochs}")
+            # Training phase
+            self.model.train()
+            train_loss_epoch = 0.0
+            # for batch_x, batch_y in train_loader:
+            for batch_x, batch_y in tqdm.tqdm(
+                train_loader,
+                desc="Batch",
+                unit="batches",
+                disable=False,
+            ):
+                batch_x, batch_y = batch_x.to(self.device), batch_y.to(self.device)
+
+                optimizer.zero_grad()
+                outputs = self.model(batch_x)
+                loss = criterion(outputs, batch_y)
+
+                if torch.isnan(loss):
+                    LGR.error("NaN loss detected, terminating training")
+                    break
+
+                loss.backward()
+                optimizer.step()
+                train_loss_epoch += loss.item()
+
+            train_loss_epoch /= len(train_loader)
+            self.loss.append(train_loss_epoch)
+
+            # Validation phase
+            self.model.eval()
+            val_loss_epoch = 0.0
+            with torch.no_grad():
+                for batch_x, batch_y in val_loader:
+                    batch_x, batch_y = batch_x.to(self.device), batch_y.to(self.device)
+                    outputs = self.model(batch_x)
+                    loss = criterion(outputs, batch_y)
+                    val_loss_epoch += loss.item()
+
+            val_loss_epoch /= len(val_loader)
+            self.val_loss.append(val_loss_epoch)
+
+            LGR.info(
+                f"Epoch {epoch+1}/{total_epochs} - Loss: {train_loss_epoch:.4f} - Val Loss: {val_loss_epoch:.4f}"
+            )
+
+            # Save checkpoint
+            self.intermediatemodelpath = os.path.join(
+                self.modelname, f"model_e{epoch+1:02d}_v{val_loss_epoch:.4f}.pth"
+            )
+            torch.save(
+                {
+                    "epoch": epoch,
+                    "model_state_dict": self.model.state_dict(),
+                    "optimizer_state_dict": optimizer.state_dict(),
+                    "loss": train_loss_epoch,
+                    "val_loss": val_loss_epoch,
+                },
+                self.intermediatemodelpath,
+            )
+
+            # Early stopping
+            if val_loss_epoch < best_val_loss:
+                best_val_loss = val_loss_epoch
+                patience_counter = 0
+                # Save best model
+                torch.save(self.model.state_dict(), os.path.join(self.modelname, "best_model.pth"))
+            else:
+                patience_counter += 1
+                if patience_counter >= patience:
+                    LGR.info(f"Early stopping triggered after {epoch+1} epochs")
+                    # Restore best weights
+                    self.model.load_state_dict(
+                        torch.load(
+                            os.path.join(self.modelname, "best_model.pth"), weights_only=True
+                        )
+                    )
+                    break
+            self.evaluate()
+
+        self.savemodel()
+        self.trained = True
+
+    def apply(self, inputdata: np.ndarray, badpts: np.ndarray | None = None) -> np.ndarray:
+        """
+        Apply a sliding-window prediction model to the input data, optionally incorporating bad points.
+
+        This function performs a sliding-window prediction using a pre-trained model. It scales the input
+        data using the median absolute deviation (MAD), applies the model to overlapping windows of data,
+        and aggregates predictions with a weighted scheme. Optionally, bad points can be included in
+        the prediction process to influence the model's behavior.
+
+        Parameters
+        ----------
+        inputdata : np.ndarray
+            Input data array of shape (N,) to be processed.
+        badpts : np.ndarray | None, optional
+            Array of same shape as `inputdata` indicating bad or invalid points. If None, no bad points
+            are considered. Default is None.
+
+        Returns
+        -------
+        np.ndarray
+            Predicted data array of the same shape as `inputdata`, with predictions aggregated and
+            weighted across overlapping windows.
+
+        Notes
+        -----
+        - The function uses a sliding window of size `self.window_size` to process input data.
+        - Predictions are aggregated by summing over overlapping windows.
+        - A triangular weight scheme is applied to the aggregated predictions to reduce edge effects.
+        - If `self.usebadpts` is True, `badpts` are included as an additional feature in the model input.
+
+        Examples
+        --------
+        >>> model = MyModel(window_size=10, usebadpts=True)
+        >>> input_data = np.random.randn(100)
+        >>> bad_points = np.zeros_like(input_data)
+        >>> result = model.apply(input_data, bad_points)
+        """
+        initscale = mad(inputdata)
+        scaleddata = inputdata / initscale
+        predicteddata = scaleddata * 0.0
+        weightarray = scaleddata * 0.0
+        N_pts = len(scaleddata)
+        if self.usebadpts:
+            if badpts is None:
+                badpts = scaleddata * 0.0
+            X = np.zeros(((N_pts - self.window_size - 1), self.window_size, 2))
+            for i in range(X.shape[0]):
+                X[i, :, 0] = scaleddata[i : i + self.window_size]
+                X[i, :, 1] = badpts[i : i + self.window_size]
+        else:
+            X = np.zeros(((N_pts - self.window_size - 1), self.window_size, 1))
+            for i in range(X.shape[0]):
+                X[i, :, 0] = scaleddata[i : i + self.window_size]
+
+        Y = self.predict_model(X)
+        for i in range(X.shape[0]):
+            predicteddata[i : i + self.window_size] += Y[i, :, 0]
+
+        weightarray[:] = self.window_size
+        weightarray[0 : self.window_size] = np.linspace(
+            1.0, self.window_size, self.window_size, endpoint=False
+        )
+        weightarray[-(self.window_size + 1) : -1] = np.linspace(
+            self.window_size, 1.0, self.window_size, endpoint=False
+        )
+        return initscale * predicteddata / weightarray
+
+
+class CNNModel(nn.Module):
+    def __init__(
+        self,
+        num_filters: int,
+        kernel_size: int,
+        num_layers: int,
+        dropout_rate: float,
+        dilation_rate: int,
+        activation: str,
+        inputsize: int,
+    ) -> None:
+        """
+        Initialize the CNNModel with specified architecture parameters.
+
+        Parameters
+        ----------
+        num_filters : int
+            Number of convolutional filters in each layer.
+        kernel_size : int
+            Size of the convolutional kernel.
+        num_layers : int
+            Total number of layers in the network.
+        dropout_rate : float
+            Dropout rate for regularization.
+        dilation_rate : int
+            Dilation rate for dilated convolutions in intermediate layers.
+        activation : str
+            Activation function to use; options are 'relu' or 'tanh'.
+        inputsize : int
+            Size of the input features.
+
+        Returns
+        -------
+        None
+            This method initializes the model in-place and does not return any value.
+
+        Notes
+        -----
+        The model consists of an input layer, intermediate layers with dilated convolutions,
+        and an output layer. Batch normalization and dropout are applied after each convolutional
+        layer except the output layer. The activation function is applied after each convolutional
+        layer based on the `activation` parameter.
+
+        Examples
+        --------
+        >>> model = CNNModel(
+        ...     num_filters=64,
+        ...     kernel_size=3,
+        ...     num_layers=4,
+        ...     dropout_rate=0.2,
+        ...     dilation_rate=2,
+        ...     activation="relu",
+        ...     inputsize=10
+        ... )
+        """
+        super(CNNModel, self).__init__()
+
+        self.num_filters = num_filters
+        self.kernel_size = kernel_size
+        self.num_layers = num_layers
+        self.dropout_rate = dropout_rate
+        self.dilation_rate = dilation_rate
+        self.activation = activation
+        self.inputsize = inputsize
+
+        self.layers = nn.ModuleList()
+
+        # Input layer
+        self.layers.append(nn.Conv1d(inputsize, num_filters, kernel_size, padding="same"))
+        self.layers.append(nn.BatchNorm1d(num_filters))
+        self.layers.append(nn.Dropout(dropout_rate))
+        if activation == "relu":
+            self.layers.append(nn.ReLU())
+        elif activation == "tanh":
+            self.layers.append(nn.Tanh())
+        else:
+            self.layers.append(nn.ReLU())
+
+        # Intermediate layers
+        for _ in range(num_layers - 2):
+            self.layers.append(
+                nn.Conv1d(
+                    num_filters,
+                    num_filters,
+                    kernel_size,
+                    dilation=dilation_rate,
+                    padding="same",
+                )
+            )
+            self.layers.append(nn.BatchNorm1d(num_filters))
+            self.layers.append(nn.Dropout(dropout_rate))
+            if activation == "relu":
+                self.layers.append(nn.ReLU())
+            elif activation == "tanh":
+                self.layers.append(nn.Tanh())
+            else:
+                self.layers.append(nn.ReLU())
+
+        # Output layer
+        self.layers.append(nn.Conv1d(num_filters, inputsize, kernel_size, padding="same"))
+
+    def forward(self, x):
+        """
+        Forward pass through all layers.
+
+        Applies each layer in the network sequentially to the input tensor.
+
+        Parameters
+        ----------
+        x : torch.Tensor
+            Input tensor to the forward pass. Shape should be compatible with the
+            first layer's expected input dimensions.
+
+        Returns
+        -------
+        torch.Tensor
+            Output tensor after passing through all layers. Shape will depend on
+            the output dimensions of the last layer in the network.
+
+        Notes
+        -----
+        This method applies layers in the order they were added to the network.
+        Each layer's forward method is called sequentially, with the output of
+        one layer becoming the input to the next.
+
+        Examples
+        --------
+        >>> import torch
+        >>> model = MyNetwork()
+        >>> input_tensor = torch.randn(32, 10)
+        >>> output = model.forward(input_tensor)
+        >>> print(output.shape)
+        torch.Size([32, 5])
+        """
+        for layer in self.layers:
+            x = layer(x)
+        return x
+
+    def get_config(self):
+        """
+        Get the configuration parameters of the model.
+
+        Returns
+        -------
+        dict
+            A dictionary containing all configuration parameters with their current values:
+            - num_filters: int, number of filters in the convolutional layers
+            - kernel_size: int, size of the convolutional kernel
+            - num_layers: int, number of layers in the network
+            - dropout_rate: float, dropout rate for regularization
+            - dilation_rate: int, dilation rate for dilated convolution
+            - activation: str, activation function used in layers
+            - inputsize: tuple, input dimensions of the model
+
+        Notes
+        -----
+        This method returns a copy of the current configuration. Modifications to the
+        returned dictionary will not affect the original model configuration.
+
+        Examples
+        --------
+        >>> config = model.get_config()
+        >>> print(config['num_filters'])
+        32
+        """
+        return {
+            "num_filters": self.num_filters,
+            "kernel_size": self.kernel_size,
+            "num_layers": self.num_layers,
+            "dropout_rate": self.dropout_rate,
+            "dilation_rate": self.dilation_rate,
+            "activation": self.activation,
+            "inputsize": self.inputsize,
+        }
+
+
+class CNNDLFilter(DeepLearningFilter):
+    def __init__(
+        self,
+        num_filters: int = 10,
+        kernel_size: int = 5,
+        dilation_rate: int = 1,
+        *args,
+        **kwargs,
+    ) -> None:
+        """
+        Initialize CNN deep learning filter.
+
+        Parameters
+        ----------
+        num_filters : int, optional
+            Number of convolutional filters to use, by default 10
+        kernel_size : int, optional
+            Size of the convolutional kernel, by default 5
+        dilation_rate : int, optional
+            Dilation rate for the convolutional layers, by default 1
+        *args
+            Variable length argument list passed to parent class
+        **kwargs
+            Arbitrary keyword arguments passed to parent class
+
+        Returns
+        -------
+        None
+            This method initializes the instance and does not return any value
+
+        Notes
+        -----
+        This constructor sets up the basic configuration for a CNN filter with
+        specified number of filters, kernel size, and dilation rate. The network
+        type is automatically set to "cnn" and information is stored in infodict
+        for later reference.
+
+        Examples
+        --------
+        >>> filter = CNNDLFilter(num_filters=32, kernel_size=3, dilation_rate=2)
+        >>> print(filter.num_filters)
+        32
+        """
+        self.num_filters = num_filters
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.nettype = "cnn"
+        self.infodict["nettype"] = self.nettype
+        self.infodict["num_filters"] = self.num_filters
+        self.infodict["kernel_size"] = self.kernel_size
+        super(CNNDLFilter, self).__init__(*args, **kwargs)
+
+    def getname(self):
+        """
+        Generate and configure the model name and path based on current parameters.
+
+        This method constructs a descriptive model name string using various instance
+        attributes and creates the corresponding directory path. The generated name
+        includes information about model architecture, hyperparameters, and configuration
+        options. The method also ensures the model directory exists by creating it if
+        necessary.
+
+        Parameters
+        ----------
+        self : object
+            The instance containing model configuration parameters.
+
+        Returns
+        -------
+        None
+            This method does not return a value but modifies instance attributes:
+            - self.modelname: Generated model name string
+            - self.modelpath: Full path to the model directory
+
+        Notes
+        -----
+        The generated model name follows a specific format:
+        "model_cnn_pytorch_wXXX_lYY_fnZZ_flZZ_eXXX_tY_ctZ_sZ_dZ_activation[options]"
+
+        Where:
+        - XXX: window_size (3 digits zero-padded)
+        - YY: num_layers (2 digits zero-padded)
+        - ZZ: num_filters (2 digits zero-padded)
+        - ZZ: kernel_size (2 digits zero-padded)
+        - XXX: num_epochs (3 digits zero-padded)
+        - Y: excludethresh (single digit)
+        - Z: corrthresh (single digit)
+        - Z: step (single digit)
+        - Z: dilation_rate (single digit)
+
+        Options are appended if corresponding boolean flags are True:
+        - _usebadpts: when usebadpts is True
+        - _excludebysubject: when excludebysubject is True
+
+        Examples
+        --------
+        >>> model = MyModel()
+        >>> model.window_size = 128
+        >>> model.num_layers = 3
+        >>> model.num_filters = 16
+        >>> model.kernel_size = 3
+        >>> model.num_epochs = 100
+        >>> model.excludethresh = 0.5
+        >>> model.corrthresh = 0.8
+        >>> model.step = 1
+        >>> model.dilation_rate = 2
+        >>> model.activation = "relu"
+        >>> model.usebadpts = True
+        >>> model.excludebysubject = False
+        >>> model.namesuffix = "test"
+        >>> model.getname()
+        >>> print(model.modelname)
+        'model_cnn_pytorch_w128_l03_fn16_fl03_e100_t0_ct0_s1_d2_relu_usebadpts_test'
+        """
+        self.modelname = "_".join(
+            [
+                "model",
+                "cnn",
+                "pytorch",
+                "w" + str(self.window_size).zfill(3),
+                "l" + str(self.num_layers).zfill(2),
+                "fn" + str(self.num_filters).zfill(2),
+                "fl" + str(self.kernel_size).zfill(2),
+                "e" + str(self.num_epochs).zfill(3),
+                "t" + str(self.excludethresh),
+                "ct" + str(self.corrthresh),
+                "s" + str(self.step),
+                "d" + str(self.dilation_rate),
+                self.activation,
+            ]
+        )
+        if self.usebadpts:
+            self.modelname += "_usebadpts"
+        if self.excludebysubject:
+            self.modelname += "_excludebysubject"
+        if self.namesuffix is not None:
+            self.modelname += "_" + self.namesuffix
+        self.modelpath = os.path.join(self.modelroot, self.modelname)
+
+        try:
+            os.makedirs(self.modelpath)
+        except OSError:
+            pass
+
+    def makenet(self):
+        """
+        Create and configure a CNN model for neural network training.
+
+        This method initializes a CNNModel with the specified parameters and moves
+        it to the designated device (CPU or GPU). The model configuration is
+        determined by the instance attributes set prior to calling this method.
+
+        Parameters
+        ----------
+        self : object
+            The instance containing the following attributes:
+            - num_filters : int
+                Number of filters in each convolutional layer
+            - kernel_size : int or tuple
+                Size of the convolutional kernel
+            - num_layers : int
+                Number of convolutional layers in the network
+            - dropout_rate : float
+                Dropout rate for regularization
+            - dilation_rate : int or tuple
+                Dilation rate for dilated convolutions
+            - activation : str or callable
+                Activation function to use
+            - inputsize : tuple
+                Input dimensions for the model
+            - device : torch.device
+                Device to move the model to (CPU or GPU)
+
+        Returns
+        -------
+        None
+            This method does not return any value. It modifies the instance
+            by setting the `model` attribute to the created CNNModel.
+
+        Notes
+        -----
+        The method assumes that all required attributes are properly initialized
+        before calling. The model is automatically moved to the specified device
+        using the `.to()` method.
+
+        Examples
+        --------
+        >>> # Assuming all required attributes are set
+        >>> makenet()
+        >>> # Model is now available as self.model
+        >>> print(self.model)
+        CNNModel(...)
+        """
+        self.model = CNNModel(
+            self.num_filters,
+            self.kernel_size,
+            self.num_layers,
+            self.dropout_rate,
+            self.dilation_rate,
+            self.activation,
+            self.inputsize,
+        )
+        self.model.to(self.device)
+
+
+class DenseAutoencoderModel(nn.Module):
+    def __init__(
+        self,
+        window_size: int,
+        encoding_dim: int,
+        num_layers: int,
+        dropout_rate: float,
+        activation: str,
+        inputsize: int,
+    ) -> None:
+        """
+        Initialize a dense autoencoder model with configurable architecture.
+
+        This constructor builds a symmetric dense autoencoder with customizable number of layers,
+        encoding dimension, dropout rate, and activation function. The model architecture follows
+        a symmetric encoder-decoder structure, where the bottleneck layer reduces the feature space
+        to `encoding_dim`, and then expands back to the original input size.
+
+        Parameters
+        ----------
+        window_size : int
+            The size of the input window (number of time steps or features per sample).
+        encoding_dim : int
+            The dimension of the latent space (bottleneck layer size).
+        num_layers : int
+            Total number of layers in the autoencoder (including input and output).
+        dropout_rate : float
+            Dropout rate applied to all hidden layers to prevent overfitting.
+        activation : str
+            Activation function to use in hidden layers. Supported values: 'relu', 'tanh'.
+            Defaults to 'relu' if an unsupported value is provided.
+        inputsize : int
+            The number of features per time step in the input data.
+
+        Returns
+        -------
+        None
+            This method initializes the model in-place and does not return any value.
+
+        Notes
+        -----
+        - The model uses batch normalization after each linear layer except the output layer.
+        - The architecture is symmetric: the number of neurons decreases from input to bottleneck
+          and then increases back to the output size.
+        - Logging is performed at various stages to track layer sizes during construction.
+
+        Examples
+        --------
+        >>> model = DenseAutoencoderModel(
+        ...     window_size=10,
+        ...     encoding_dim=5,
+        ...     num_layers=4,
+        ...     dropout_rate=0.2,
+        ...     activation="relu",
+        ...     inputsize=3
+        ... )
+        """
+        super(DenseAutoencoderModel, self).__init__()
+
+        self.window_size = window_size
+        self.encoding_dim = encoding_dim
+        self.num_layers = num_layers
+        self.dropout_rate = dropout_rate
+        self.activation = activation
+        self.inputsize = inputsize
+
+        self.layers = nn.ModuleList()
+
+        # Calculate initial size factor
+        sizefac = 2
+        for i in range(1, num_layers - 1):
+            sizefac = int(sizefac * 2)
+        LGR.info(f"input layer - sizefac: {sizefac}")
+
+        # Input layer
+        self.layers.append(nn.Linear(window_size * inputsize, sizefac * encoding_dim))
+        self.layers.append(nn.BatchNorm1d(sizefac * encoding_dim))
+        self.layers.append(nn.Dropout(dropout_rate))
+        if activation == "relu":
+            self.layers.append(nn.ReLU())
+        elif activation == "tanh":
+            self.layers.append(nn.Tanh())
+        else:
+            self.layers.append(nn.ReLU())
+
+        # Encoding layers
+        for i in range(1, num_layers - 1):
+            sizefac = int(sizefac // 2)
+            LGR.info(f"encoder layer {i + 1}, sizefac: {sizefac}")
+            self.layers.append(nn.Linear(sizefac * 2 * encoding_dim, sizefac * encoding_dim))
+            self.layers.append(nn.BatchNorm1d(sizefac * encoding_dim))
+            self.layers.append(nn.Dropout(dropout_rate))
+            if activation == "relu":
+                self.layers.append(nn.ReLU())
+            elif activation == "tanh":
+                self.layers.append(nn.Tanh())
+            else:
+                self.layers.append(nn.ReLU())
+
+        # Encoding layer (bottleneck)
+        sizefac = int(sizefac // 2)
+        LGR.info(f"encoding layer - sizefac: {sizefac}")
+        self.layers.append(nn.Linear(sizefac * 2 * encoding_dim, encoding_dim))
+        self.layers.append(nn.BatchNorm1d(encoding_dim))
+        self.layers.append(nn.Dropout(dropout_rate))
+        if activation == "relu":
+            self.layers.append(nn.ReLU())
+        elif activation == "tanh":
+            self.layers.append(nn.Tanh())
+        else:
+            self.layers.append(nn.ReLU())
+
+        # Decoding layers
+        for i in range(1, num_layers):
+            sizefac = int(sizefac * 2)
+            LGR.info(f"decoding layer {i}, sizefac: {sizefac}")
+            if i == 1:
+                self.layers.append(nn.Linear(encoding_dim, sizefac * encoding_dim))
+            else:
+                self.layers.append(nn.Linear(sizefac // 2 * encoding_dim, sizefac * encoding_dim))
+            self.layers.append(nn.BatchNorm1d(sizefac * encoding_dim))
+            self.layers.append(nn.Dropout(dropout_rate))
+            if activation == "relu":
+                self.layers.append(nn.ReLU())
+            elif activation == "tanh":
+                self.layers.append(nn.Tanh())
+            else:
+                self.layers.append(nn.ReLU())
+
+        # Output layer
+        self.layers.append(nn.Linear(sizefac * encoding_dim, window_size * inputsize))
+
+    def forward(self, x):
+        """
+        Forward pass through the network layers.
+
+        Applies a series of layers to the input tensor, flattening and reshaping
+        as needed for processing.
+
+        Parameters
+        ----------
+        x : torch.Tensor
+            Input tensor with shape (batch, channels, length) where batch is the
+            number of samples, channels is the number of channels, and length is
+            the sequence length.
+
+        Returns
+        -------
+        torch.Tensor
+            Output tensor with shape (batch, channels, length) where the
+            dimensions match the input shape after processing through the layers.
+
+        Notes
+        -----
+        The function first flattens the input from (batch, channels, length) to
+        (batch, channels*length) to enable processing through linear layers,
+        then reshapes the output back to the original format.
+
+        Examples
+        --------
+        >>> import torch
+        >>> # Assuming self.layers contains appropriate layer definitions
+        >>> x = torch.randn(32, 3, 100)  # batch_size=32, channels=3, length=100
+        >>> output = model.forward(x)
+        >>> output.shape
+        torch.Size([32, 3, 100])
+        """
+        # Flatten input from (batch, channels, length) to (batch, channels*length)
+        batch_size = x.shape[0]
+        x = x.reshape(batch_size, -1)
+
+        for layer in self.layers:
+            x = layer(x)
+
+        # Reshape back to (batch, channels, length)
+        x = x.reshape(batch_size, self.inputsize, self.window_size)
+        return x
+
+    def get_config(self):
+        """
+        Get the configuration parameters of the model.
+
+        Returns
+        -------
+        dict
+            A dictionary containing the model configuration parameters with the following keys:
+            - "window_size" (int): The size of the sliding window used for input sequences
+            - "encoding_dim" (int): The dimensionality of the encoding layer
+            - "num_layers" (int): The number of layers in the model
+            - "dropout_rate" (float): The dropout rate for regularization
+            - "activation" (str): The activation function used in the model
+            - "inputsize" (int): The size of the input features
+
+        Notes
+        -----
+        This method returns a copy of the internal configuration parameters. Modifications
+        to the returned dictionary will not affect the original model configuration.
+
+        Examples
+        --------
+        >>> model = MyModel()
+        >>> config = model.get_config()
+        >>> print(config['window_size'])
+        10
+        """
+        return {
+            "window_size": self.window_size,
+            "encoding_dim": self.encoding_dim,
+            "num_layers": self.num_layers,
+            "dropout_rate": self.dropout_rate,
+            "activation": self.activation,
+            "inputsize": self.inputsize,
+        }
+
+
+class DenseAutoencoderDLFilter(DeepLearningFilter):
+    def __init__(self, encoding_dim: int = 10, *args, **kwargs) -> None:
+        """
+        Initialize DenseAutoencoderDLFilter instance.
+
+        Parameters
+        ----------
+        encoding_dim : int, default=10
+            Dimension of the encoding layer in the autoencoder. This determines
+            the size of the latent representation learned by the model.
+        *args
+            Variable length argument list passed to the parent class constructor.
+        **kwargs
+            Arbitrary keyword arguments passed to the parent class constructor.
+
+        Returns
+        -------
+        None
+            This method initializes the instance and does not return any value.
+
+        Notes
+        -----
+        This constructor sets up the autoencoder architecture by:
+        1. Storing the encoding dimension as an instance attribute
+        2. Setting the network type to "autoencoder"
+        3. Updating the info dictionary with network type and encoding dimension
+        4. Calling the parent class constructor with any additional arguments
+
+        Examples
+        --------
+        >>> filter = DenseAutoencoderDLFilter(encoding_dim=15)
+        >>> print(filter.encoding_dim)
+        15
+        >>> print(filter.nettype)
+        'autoencoder'
+        """
+        self.encoding_dim = encoding_dim
+        self.nettype = "autoencoder"
+        self.infodict["nettype"] = self.nettype
+        self.infodict["encoding_dim"] = self.encoding_dim
+        super(DenseAutoencoderDLFilter, self).__init__(*args, **kwargs)
+
+    def getname(self):
+        """
+        Generate and return the model name and path based on current configuration parameters.
+
+        This method constructs a descriptive model name string using various configuration
+        parameters and creates the corresponding directory path for model storage. The
+        generated name includes information about window size, encoding dimensions, epochs,
+        thresholds, step size, activation function, and additional flags.
+
+        Parameters
+        ----------
+        self : object
+            The instance containing the following attributes:
+            - window_size : int
+                Size of the sliding window
+            - encoding_dim : int
+                Dimension of the encoding layer
+            - num_epochs : int
+                Number of training epochs
+            - excludethresh : float
+                Threshold for excluding data points
+            - corrthresh : float
+                Correlation threshold for filtering
+            - step : int
+                Step size for sliding window
+            - activation : str
+                Activation function name
+            - usebadpts : bool
+                Flag indicating whether to use bad points
+            - excludebysubject : bool
+                Flag indicating whether to exclude by subject
+            - namesuffix : str, optional
+                Additional suffix to append to model name
+            - modelroot : str
+                Root directory for model storage
+
+        Returns
+        -------
+        None
+            This method modifies the instance attributes in-place:
+            - self.modelname : str
+                Generated model name string
+            - self.modelpath : str
+                Full path to the model directory
+
+        Notes
+        -----
+        The generated model name follows a specific format:
+        "model_denseautoencoder_pytorch_wXXX_enXXX_eXXX_tX.XX_ctX.XX_sX_activation[flags]"
+
+        Examples
+        --------
+        >>> model = MyModel()
+        >>> model.window_size = 100
+        >>> model.encoding_dim = 50
+        >>> model.num_epochs = 1000
+        >>> model.excludethresh = 0.5
+        >>> model.corrthresh = 0.8
+        >>> model.step = 10
+        >>> model.activation = 'relu'
+        >>> model.usebadpts = True
+        >>> model.excludebysubject = False
+        >>> model.namesuffix = 'test'
+        >>> model.modelroot = '/path/to/models'
+        >>> model.getname()
+        >>> print(model.modelname)
+        'model_denseautoencoder_pytorch_w100_en050_e1000_t0.5_ct0.8_s10_relu_usebadpts_test'
+        """
+        self.modelname = "_".join(
+            [
+                "model",
+                "denseautoencoder",
+                "pytorch",
+                "w" + str(self.window_size).zfill(3),
+                "en" + str(self.encoding_dim).zfill(3),
+                "e" + str(self.num_epochs).zfill(3),
+                "t" + str(self.excludethresh),
+                "ct" + str(self.corrthresh),
+                "s" + str(self.step),
+                self.activation,
+            ]
+        )
+        if self.usebadpts:
+            self.modelname += "_usebadpts"
+        if self.excludebysubject:
+            self.modelname += "_excludebysubject"
+        if self.namesuffix is not None:
+            self.modelname += "_" + self.namesuffix
+        self.modelpath = os.path.join(self.modelroot, self.modelname)
+
+        try:
+            os.makedirs(self.modelpath)
+        except OSError:
+            pass
+
+    def makenet(self):
+        """
+        Create and configure a dense autoencoder model.
+
+        This method initializes a DenseAutoencoderModel with the specified parameters
+        and moves it to the designated device (CPU or GPU).
+
+        Parameters
+        ----------
+        self : object
+            The instance of the class containing this method. Expected to have the
+            following attributes:
+            - window_size : int
+            - encoding_dim : int
+            - num_layers : int
+            - dropout_rate : float
+            - activation : str or callable
+            - inputsize : int
+            - device : torch.device
+
+        Returns
+        -------
+        None
+            This method does not return any value. It sets the model attribute
+            of the instance to the created DenseAutoencoderModel.
+
+        Notes
+        -----
+        The model is automatically moved to the device specified by self.device.
+        This method should be called after all required parameters have been set
+        on the instance.
+
+        Examples
+        --------
+        >>> # Assuming a class with the makenet method
+        >>> instance = MyClass()
+        >>> instance.window_size = 100
+        >>> instance.encoding_dim = 32
+        >>> instance.num_layers = 3
+        >>> instance.dropout_rate = 0.2
+        >>> instance.activation = 'relu'
+        >>> instance.inputsize = 100
+        >>> instance.device = torch.device('cuda')
+        >>> instance.makenet()
+        >>> # Model is now available as instance.model
+        """
+        self.model = DenseAutoencoderModel(
+            self.window_size,
+            self.encoding_dim,
+            self.num_layers,
+            self.dropout_rate,
+            self.activation,
+            self.inputsize,
+        )
+        self.model.to(self.device)
+
+
+class MultiscaleCNNModel(nn.Module):
+    def __init__(
+        self,
+        num_filters,
+        kernel_sizes,
+        input_lens,
+        input_width,
+        dilation_rate,
+    ):
+        """
+        Initialize the MultiscaleCNNModel.
+
+        This constructor initializes a multiscale CNN model with three parallel branches
+        processing input at different scales. Each branch uses dilated convolutions to
+        capture features at different receptive fields.
+
+        Parameters
+        ----------
+        num_filters : int
+            Number of filters (channels) to use in the convolutional layers.
+        kernel_sizes : list of int
+            List of three kernel sizes for the three branches (small, medium, large scales).
+        input_lens : list of int
+            List of input lengths for each branch, corresponding to the input sequence lengths.
+        input_width : int
+            Width of the input features (number of input channels).
+        dilation_rate : int
+            Dilation rate to use in the dilated convolutional layers.
+
+        Returns
+        -------
+        None
+            This method initializes the model instance and does not return any value.
+
+        Notes
+        -----
+        The model creates three parallel branches with different kernel sizes to capture
+        multi-scale temporal features. Each branch uses dilated convolutions to increase
+        the receptive field without increasing the number of parameters significantly.
+
+        The final dense layer reduces the combined features to a single output value,
+        followed by a sigmoid activation for binary classification.
+
+        Examples
+        --------
+        >>> model = MultiscaleCNNModel(
+        ...     num_filters=64,
+        ...     kernel_sizes=[3, 5, 7],
+        ...     input_lens=[100, 100, 100],
+        ...     input_width=10,
+        ...     dilation_rate=2
+        ... )
+        """
+        super(MultiscaleCNNModel, self).__init__()
+
+        self.num_filters = num_filters
+        self.kernel_sizes = kernel_sizes
+        self.input_lens = input_lens
+        self.input_width = input_width
+        self.dilation_rate = dilation_rate
+
+        # Create three separate branches for different scales
+        self.branch_small = self._make_branch(kernel_sizes[0])
+        self.branch_med = self._make_branch(kernel_sizes[1])
+        self.branch_large = self._make_branch(kernel_sizes[2])
+
+        # Final dense layer
+        self.fc = nn.Linear(150, 1)
+        self.sigmoid = nn.Sigmoid()
+
+    def _make_branch(self, kernel_size):
+        """
+        Create a convolutional branch for the neural network architecture.
+
+        This method constructs a sequential neural network branch consisting of
+        convolutional, pooling, flattening, linear, activation, and dropout layers.
+
+        Parameters
+        ----------
+        kernel_size : int
+            The size of the convolutional kernel to be used in the Conv1d layer.
+
+        Returns
+        -------
+        torch.nn.Sequential
+            A sequential container containing the following layers:
+            - Conv1d: 1D convolutional layer with input_width as input channels
+            - AdaptiveMaxPool1d: 1D adaptive max pooling with output size 1
+            - Flatten: Flattens the tensor for linear layer input
+            - Linear: Linear layer with num_filters input features and 50 output features
+            - Tanh: Hyperbolic tangent activation function
+            - Dropout: 30% dropout regularization
+
+        Notes
+        -----
+        The branch is designed to process 1D input data through a convolutional
+        feature extraction pathway followed by a fully connected classifier head.
+        The padding="same" parameter ensures the output size matches the input size
+        for the convolutional layer.
+
+        Examples
+        --------
+        >>> branch = self._make_branch(kernel_size=3)
+        >>> print(type(branch))
+        <class 'torch.nn.modules.container.Sequential'>
+        """
+        return nn.Sequential(
+            nn.Conv1d(self.input_width, self.num_filters, kernel_size, padding="same"),
+            nn.AdaptiveMaxPool1d(1),
+            nn.Flatten(),
+            nn.Linear(self.num_filters, 50),
+            nn.Tanh(),
+            nn.Dropout(0.3),
+        )
+
+    def forward(self, x_small, x_med, x_large):
+        """
+        Forward pass of the multi-scale feature extraction network.
+
+        This function processes input tensors through three parallel branches with different
+        receptive fields and concatenates the outputs before applying a final fully connected
+        layer with sigmoid activation.
+
+        Parameters
+        ----------
+        x_small : torch.Tensor
+            Input tensor for the small-scale branch with shape (batch_size, channels, height, width)
+        x_med : torch.Tensor
+            Input tensor for the medium-scale branch with shape (batch_size, channels, height, width)
+        x_large : torch.Tensor
+            Input tensor for the large-scale branch with shape (batch_size, channels, height, width)
+
+        Returns
+        -------
+        torch.Tensor
+            Output tensor with shape (batch_size, num_classes) containing sigmoid-activated
+            predictions for each class
+
+        Notes
+        -----
+        The function assumes that `self.branch_small`, `self.branch_med`, `self.branch_large`,
+        `self.fc`, and `self.sigmoid` are properly initialized components of the class.
+
+        Examples
+        --------
+        >>> import torch
+        >>> # Assuming model is initialized
+        >>> x_small = torch.randn(1, 3, 32, 32)
+        >>> x_med = torch.randn(1, 3, 64, 64)
+        >>> x_large = torch.randn(1, 3, 128, 128)
+        >>> output = model.forward(x_small, x_med, x_large)
+        >>> print(output.shape)
+        torch.Size([1, num_classes])
+        """
+        # Process each branch
+        out_small = self.branch_small(x_small)
+        out_med = self.branch_med(x_med)
+        out_large = self.branch_large(x_large)
+
+        # Concatenate outputs
+        merged = torch.cat([out_small, out_med, out_large], dim=1)
+
+        # Final output
+        out = self.fc(merged)
+        out = self.sigmoid(out)
+
+        return out
+
+    def get_config(self):
+        """
+        Get the configuration parameters of the layer.
+
+        Returns
+        -------
+        dict
+            A dictionary containing the layer configuration parameters with the following keys:
+            - "num_filters" (int): Number of filters in the layer
+            - "kernel_sizes" (list of int): Size of the convolutional kernels
+            - "input_lens" (list of int): Lengths of input sequences
+            - "input_width" (int): Width of the input data
+            - "dilation_rate" (int): Dilation rate for dilated convolution
+
+        Notes
+        -----
+        This method returns a copy of the internal configuration parameters
+        that can be used to reconstruct the layer with the same settings.
+
+        Examples
+        --------
+        >>> config = layer.get_config()
+        >>> print(config['num_filters'])
+        32
+        """
+        return {
+            "num_filters": self.num_filters,
+            "kernel_sizes": self.kernel_sizes,
+            "input_lens": self.input_lens,
+            "input_width": self.input_width,
+            "dilation_rate": self.dilation_rate,
+        }
+
+
+class MultiscaleCNNDLFilter(DeepLearningFilter):
+    def __init__(
+        self,
+        num_filters=10,
+        kernel_sizes=[4, 8, 12],
+        input_lens=[64, 128, 192],
+        input_width=1,
+        dilation_rate=1,
+        *args,
+        **kwargs,
+    ):
+        """
+        Initialize the MultiscaleCNNDLFilter layer.
+
+        This constructor initializes a multiscale CNN filter with configurable
+        kernel sizes, input lengths, and dilation rates for multi-scale feature extraction.
+
+        Parameters
+        ----------
+        num_filters : int, optional
+            Number of filters to use in each convolutional layer, default is 10
+        kernel_sizes : list of int, optional
+            List of kernel sizes for different convolutional layers, default is [4, 8, 12]
+        input_lens : list of int, optional
+            List of input sequence lengths for different scales, default is [64, 128, 192]
+        input_width : int, optional
+            Width of the input data, default is 1
+        dilation_rate : int, optional
+            Dilation rate for the convolutional layers, default is 1
+        *args : tuple
+            Variable length argument list passed to parent class
+        **kwargs : dict
+            Arbitrary keyword arguments passed to parent class
+
+        Returns
+        -------
+        None
+            This method initializes the object and does not return any value
+
+        Notes
+        -----
+        The multiscale CNN filter uses multiple convolutional layers with different
+        kernel sizes and dilation rates to capture features at multiple scales.
+        The input data is processed through parallel convolutional branches,
+        each with different kernel sizes and dilation rates.
+
+        Examples
+        --------
+        >>> filter_layer = MultiscaleCNNDLFilter(
+        ...     num_filters=20,
+        ...     kernel_sizes=[3, 6, 9],
+        ...     input_lens=[32, 64, 128],
+        ...     input_width=2,
+        ...     dilation_rate=2
+        ... )
+        """
+        self.num_filters = num_filters
+        self.kernel_sizes = kernel_sizes
+        self.input_lens = input_lens
+        self.input_width = input_width
+        self.dilation_rate = dilation_rate
+        self.nettype = "multiscalecnn"
+        self.infodict["nettype"] = self.nettype
+        self.infodict["num_filters"] = self.num_filters
+        self.infodict["kernel_sizes"] = self.kernel_sizes
+        self.infodict["input_lens"] = self.input_lens
+        self.infodict["input_width"] = self.input_width
+        super(MultiscaleCNNDLFilter, self).__init__(*args, **kwargs)
+
+    def getname(self):
+        """
+        Generate and return the model name and path based on current configuration parameters.
+
+        This method constructs a descriptive model name string by joining various configuration
+        parameters with specific prefixes and zero-padded numeric values. The resulting name
+        is used to create a unique directory path for model storage.
+
+        Parameters
+        ----------
+        self : object
+            The instance containing model configuration parameters.
+
+        Returns
+        -------
+        None
+            This method does not return a value but sets the following attributes:
+            - self.modelname: str, the generated model name
+            - self.modelpath: str, the full path to the model directory
+
+        Notes
+        -----
+        The generated model name includes the following components:
+        - Model type: "model_multiscalecnn_pytorch"
+        - Window size: "w" + zero-padded window size
+        - Number of layers: "l" + zero-padded layer count
+        - Number of filters: "fn" + zero-padded filter count
+        - First kernel size: "fl" + zero-padded kernel size
+        - Number of epochs: "e" + zero-padded epoch count
+        - Exclusion threshold: "t" + threshold value
+        - Correlation threshold: "ct" + threshold value
+        - Step size: "s" + zero-padded step value
+        - Dilation rate: "d" + dilation rate value
+        - Activation function name
+
+        Additional suffixes are appended if:
+        - usebadpts is True: "_usebadpts"
+        - excludebysubject is True: "_excludebysubject"
+        - namesuffix is not None: "_{namesuffix}"
+
+        Examples
+        --------
+        >>> model = MyModel()
+        >>> model.window_size = 128
+        >>> model.num_layers = 5
+        >>> model.num_filters = 32
+        >>> model.kernel_sizes = [3, 5, 7]
+        >>> model.num_epochs = 100
+        >>> model.excludethresh = 0.5
+        >>> model.corrthresh = 0.8
+        >>> model.step = 16
+        >>> model.dilation_rate = 2
+        >>> model.activation = "relu"
+        >>> model.usebadpts = True
+        >>> model.excludebysubject = False
+        >>> model.namesuffix = "exp1"
+        >>> model.getname()
+        >>> print(model.modelname)
+        'model_multiscalecnn_pytorch_w128_l05_fn32_fl03_e100_t0.5_ct0.8_s16_d2_relu_usebadpts_exp1'
+        """
+        self.modelname = "_".join(
+            [
+                "model",
+                "multiscalecnn",
+                "pytorch",
+                "w" + str(self.window_size).zfill(3),
+                "l" + str(self.num_layers).zfill(2),
+                "fn" + str(self.num_filters).zfill(2),
+                "fl" + str(self.kernel_sizes[0]).zfill(2),
+                "e" + str(self.num_epochs).zfill(3),
+                "t" + str(self.excludethresh),
+                "ct" + str(self.corrthresh),
+                "s" + str(self.step),
+                "d" + str(self.dilation_rate),
+                self.activation,
+            ]
+        )
+        if self.usebadpts:
+            self.modelname += "_usebadpts"
+        if self.excludebysubject:
+            self.modelname += "_excludebysubject"
+        if self.namesuffix is not None:
+            self.modelname += "_" + self.namesuffix
+        self.modelpath = os.path.join(self.modelroot, self.modelname)
+
+        try:
+            os.makedirs(self.modelpath)
+        except OSError:
+            pass
+
+    def makenet(self):
+        """
+        Create and initialize a multiscale CNN model for network construction.
+
+        This method initializes a MultiscaleCNNModel with the specified parameters
+        and moves the model to the designated device (CPU or GPU).
+
+        Parameters
+        ----------
+        self : object
+            The instance containing the following attributes:
+            - num_filters : int
+                Number of filters for the CNN layers
+            - kernel_sizes : list of int
+                List of kernel sizes for different scales
+            - input_lens : list of int
+                List of input lengths for different scales
+            - input_width : int
+                Width of the input data
+            - dilation_rate : int
+                Dilation rate for the convolutional layers
+            - device : torch.device
+                Device to move the model to (e.g., 'cuda' or 'cpu')
+
+        Returns
+        -------
+        None
+            This method does not return any value but modifies the instance
+            by setting the `model` attribute to the created MultiscaleCNNModel.
+
+        Notes
+        -----
+        The method assumes that all required attributes are properly initialized
+        in the instance before calling this method. The model is automatically
+        moved to the specified device using the `.to()` method.
+
+        Examples
+        --------
+        >>> # Assuming instance with required attributes is created
+        >>> instance.makenet()
+        >>> print(instance.model)
+        MultiscaleCNNModel(...)
+        """
+        self.model = MultiscaleCNNModel(
+            self.num_filters,
+            self.kernel_sizes,
+            self.input_lens,
+            self.input_width,
+            self.dilation_rate,
+        )
+        self.model.to(self.device)
+
+
+class ConvAutoencoderModel(nn.Module):
+    def __init__(
+        self,
+        window_size,
+        encoding_dim,
+        num_filters,
+        kernel_size,
+        dropout_rate,
+        activation,
+        inputsize,
+    ):
+        """
+        Initialize the ConvAutoencoderModel.
+
+        This class implements a convolutional autoencoder for time series data. The model
+        consists of an encoder and a decoder, with symmetric architecture. The encoder
+        reduces the input dimensionality through convolutional and pooling layers, while
+        the decoder reconstructs the input from the encoded representation.
+
+        Parameters
+        ----------
+        window_size : int
+            The length of the input time series window.
+        encoding_dim : int
+            The dimensionality of the latent space representation.
+        num_filters : int
+            The number of filters in the first convolutional layer.
+        kernel_size : int
+            The size of the convolutional kernels.
+        dropout_rate : float
+            The dropout rate applied after each convolutional layer.
+        activation : str
+            The activation function to use. Supported values are "relu" and "tanh".
+        inputsize : int
+            The number of input channels (e.g., number of features in the time series).
+
+        Returns
+        -------
+        None
+            This method initializes the model in-place and does not return any value.
+
+        Notes
+        -----
+        The model uses a symmetric encoder-decoder architecture. The encoder reduces
+        the input size through 4 max-pooling layers, and the decoder reconstructs
+        the input using upsample and convolutional layers. The final layer uses
+        a convolution with padding to match the input size.
+
+        Examples
+        --------
+        >>> model = ConvAutoencoderModel(
+        ...     window_size=100,
+        ...     encoding_dim=32,
+        ...     num_filters=32,
+        ...     kernel_size=3,
+        ...     dropout_rate=0.2,
+        ...     activation="relu",
+        ...     inputsize=1
+        ... )
+        """
+        super(ConvAutoencoderModel, self).__init__()
+
+        self.window_size = window_size
+        self.encoding_dim = encoding_dim
+        self.num_filters = num_filters
+        self.kernel_size = kernel_size
+        self.dropout_rate = dropout_rate
+        self.activation = activation
+        self.inputsize = inputsize
+
+        # Get activation function
+        if activation == "relu":
+            act_fn = nn.ReLU
+        elif activation == "tanh":
+            act_fn = nn.Tanh
+        else:
+            act_fn = nn.ReLU
+
+        # Initial conv block
+        self.encoder_layers = nn.ModuleList()
+        self.encoder_layers.append(nn.Conv1d(inputsize, num_filters, kernel_size, padding="same"))
+        self.encoder_layers.append(nn.BatchNorm1d(num_filters))
+        self.encoder_layers.append(nn.Dropout(dropout_rate))
+        self.encoder_layers.append(act_fn())
+        self.encoder_layers.append(nn.MaxPool1d(2, padding=1))
+
+        # Encoding path (3 layers)
+        nfilters = num_filters
+        self.filter_list = []
+        for _ in range(3):
+            nfilters *= 2
+            self.filter_list.append(nfilters)
+            self.encoder_layers.append(
+                nn.Conv1d(nfilters // 2, nfilters, kernel_size, padding="same")
+            )
+            self.encoder_layers.append(nn.BatchNorm1d(nfilters))
+            self.encoder_layers.append(nn.Dropout(dropout_rate))
+            self.encoder_layers.append(act_fn())
+            self.encoder_layers.append(nn.MaxPool1d(2, padding=1))
+
+        # Calculate size after pooling
+        self.encoded_size = window_size
+        for _ in range(4):  # 4 pooling layers
+            self.encoded_size = (self.encoded_size + 1) // 2
+
+        # Bottleneck
+        self.flatten = nn.Flatten()
+        self.encode_fc = nn.Linear(nfilters * self.encoded_size, encoding_dim)
+        self.encode_act = act_fn()
+        self.decode_fc = nn.Linear(encoding_dim, nfilters * self.encoded_size)
+        self.decode_act = act_fn()
+        self.unflatten_size = (nfilters, self.encoded_size)
+
+        # Decoding path (mirror)
+        self.decoder_layers = nn.ModuleList()
+        for i, filters in enumerate(reversed(self.filter_list)):
+            self.decoder_layers.append(nn.Upsample(scale_factor=2, mode="nearest"))
+            if i == 0:
+                self.decoder_layers.append(
+                    nn.Conv1d(nfilters, filters, kernel_size, padding="same")
+                )
+            else:
+                self.decoder_layers.append(
+                    nn.Conv1d(self.filter_list[-i], filters, kernel_size, padding="same")
+                )
+            self.decoder_layers.append(nn.BatchNorm1d(filters))
+            self.decoder_layers.append(nn.Dropout(dropout_rate))
+            self.decoder_layers.append(act_fn())
+
+        # Final upsampling
+        self.decoder_layers.append(nn.Upsample(scale_factor=2, mode="nearest"))
+        self.decoder_layers.append(nn.Conv1d(num_filters, inputsize, kernel_size, padding="same"))
+
+    def forward(self, x):
+        """
+        Forward pass of the autoencoder.
+
+        Applies encoding, bottleneck processing, and decoding to the input tensor.
+
+        Parameters
+        ----------
+        x : torch.Tensor
+            Input tensor of shape (batch_size, channels, height, width) where
+            height and width should match the expected window size.
+
+        Returns
+        -------
+        torch.Tensor
+            Reconstructed tensor of shape (batch_size, channels, window_size, width)
+            with the same spatial dimensions as the input.
+
+        Notes
+        -----
+        The forward pass consists of three main stages:
+        1. Encoding: Input is passed through encoder layers
+        2. Bottleneck: Flattening, encoding, and decoding with activation functions
+        3. Decoding: Reconstructed features are passed through decoder layers
+
+        The output is cropped or padded to match the original window size.
+
+        Examples
+        --------
+        >>> import torch
+        >>> model = AutoEncoder()
+        >>> x = torch.randn(1, 3, 64, 64)
+        >>> output = model.forward(x)
+        >>> print(output.shape)
+        torch.Size([1, 3, 64, 64])
+        """
+        # Encoding
+        for layer in self.encoder_layers:
+            x = layer(x)
+
+        # Bottleneck
+        x = self.flatten(x)
+        x = self.encode_fc(x)
+        x = self.encode_act(x)
+        x = self.decode_fc(x)
+        x = self.decode_act(x)
+        x = x.view(x.size(0), *self.unflatten_size)
+
+        # Decoding
+        for layer in self.decoder_layers:
+            x = layer(x)
+
+        # Crop/pad to original window size
+        if x.size(2) > self.window_size:
+            x = x[:, :, : self.window_size]
+        elif x.size(2) < self.window_size:
+            pad_size = self.window_size - x.size(2)
+            x = nn.functional.pad(x, (0, pad_size))
+
+        return x
+
+    def get_config(self):
+        """
+        Get the configuration parameters of the model.
+
+        Returns
+        -------
+        dict
+            A dictionary containing all configuration parameters with their current values:
+            - "window_size" (int): Size of the sliding window
+            - "encoding_dim" (int): Dimension of the encoding layer
+            - "num_filters" (int): Number of filters in the convolutional layers
+            - "kernel_size" (int): Size of the convolutional kernel
+            - "dropout_rate" (float): Dropout rate for regularization
+            - "activation" (str): Activation function to use
+            - "inputsize" (int): Size of the input data
+
+        Notes
+        -----
+        This method returns a copy of the current configuration. Modifications to the
+        returned dictionary will not affect the original model configuration.
+
+        Examples
+        --------
+        >>> config = model.get_config()
+        >>> print(config['window_size'])
+        100
+        """
+        return {
+            "window_size": self.window_size,
+            "encoding_dim": self.encoding_dim,
+            "num_filters": self.num_filters,
+            "kernel_size": self.kernel_size,
+            "dropout_rate": self.dropout_rate,
+            "activation": self.activation,
+            "inputsize": self.inputsize,
+        }
+
+
+class ConvAutoencoderDLFilter(DeepLearningFilter):
+    def __init__(
+        self,
+        encoding_dim: int = 10,
+        num_filters: int = 5,
+        kernel_size: int = 5,
+        dilation_rate: int = 1,
+        *args,
+        **kwargs,
+    ) -> None:
+        """
+        Initialize ConvAutoencoderDLFilter instance.
+
+        Parameters
+        ----------
+        encoding_dim : int, optional
+            Dimension of the encoded representation, by default 10
+        num_filters : int, optional
+            Number of filters in the convolutional layers, by default 5
+        kernel_size : int, optional
+            Size of the convolutional kernel, by default 5
+        dilation_rate : int, optional
+            Dilation rate for the convolutional layers, by default 1
+        *args
+            Variable length argument list
+        **kwargs
+            Arbitrary keyword arguments
+
+        Returns
+        -------
+        None
+            This method does not return any value
+
+        Notes
+        -----
+        This constructor initializes a convolutional autoencoder with dilated filters.
+        The network type is set to "convautoencoder" and various configuration parameters
+        are stored in the infodict for later reference.
+
+        Examples
+        --------
+        >>> autoencoder = ConvAutoencoderDLFilter(
+        ...     encoding_dim=15,
+        ...     num_filters=8,
+        ...     kernel_size=3,
+        ...     dilation_rate=2
+        ... )
+        """
+        self.encoding_dim = encoding_dim
+        self.num_filters = num_filters
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.nettype = "convautoencoder"
+        self.infodict["num_filters"] = self.num_filters
+        self.infodict["kernel_size"] = self.kernel_size
+        self.infodict["nettype"] = self.nettype
+        self.infodict["encoding_dim"] = self.encoding_dim
+        super(ConvAutoencoderDLFilter, self).__init__(*args, **kwargs)
+
+    def getname(self):
+        """
+        Generate and configure the model name and path based on current parameters.
+
+        This method constructs a descriptive model name string using various
+        configuration parameters and creates the corresponding directory path.
+        The generated name includes information about window size, encoding dimensions,
+        filters, kernel size, epochs, thresholds, and other model configuration options.
+
+        Parameters
+        ----------
+        self : object
+            The instance of the class containing model configuration parameters.
+            Expected attributes include:
+            - window_size : int
+            - encoding_dim : int
+            - num_filters : int
+            - kernel_size : int
+            - num_epochs : int
+            - excludethresh : float
+            - corrthresh : float
+            - step : int
+            - activation : str
+            - usebadpts : bool
+            - excludebysubject : bool
+            - namesuffix : str, optional
+            - modelroot : str
+
+        Returns
+        -------
+        None
+            This method does not return a value but modifies the instance attributes:
+            - self.modelname : str
+            - self.modelpath : str
+
+        Notes
+        -----
+        The model name is constructed with the following components:
+        - "model_convautoencoder_pytorch" as base identifier
+        - Window size with 3-digit zero-padded formatting (wXXX)
+        - Encoding dimension with 3-digit zero-padded formatting (enXXX)
+        - Number of filters with 2-digit zero-padded formatting (fnXX)
+        - Kernel size with 2-digit zero-padded formatting (flXX)
+        - Number of epochs with 3-digit zero-padded formatting (eXXX)
+        - Exclusion threshold (tX.XX)
+        - Correlation threshold (ctX.XX)
+        - Step size (sX)
+        - Activation function name
+
+        Additional suffixes are appended based on:
+        - usebadpts flag
+        - excludebysubject flag
+        - namesuffix parameter
+
+        Examples
+        --------
+        >>> model = MyModel()
+        >>> model.window_size = 100
+        >>> model.encoding_dim = 50
+        >>> model.getname()
+        >>> print(model.modelname)
+        'model_convautoencoder_pytorch_w100_en050_fn10_fl05_e001_t0.5_ct0.8_s1_relu'
+        """
+        self.modelname = "_".join(
+            [
+                "model",
+                "convautoencoder",
+                "pytorch",
+                "w" + str(self.window_size).zfill(3),
+                "en" + str(self.encoding_dim).zfill(3),
+                "fn" + str(self.num_filters).zfill(2),
+                "fl" + str(self.kernel_size).zfill(2),
+                "e" + str(self.num_epochs).zfill(3),
+                "t" + str(self.excludethresh),
+                "ct" + str(self.corrthresh),
+                "s" + str(self.step),
+                self.activation,
+            ]
+        )
+        if self.usebadpts:
+            self.modelname += "_usebadpts"
+        if self.excludebysubject:
+            self.modelname += "_excludebysubject"
+        if self.namesuffix is not None:
+            self.modelname += "_" + self.namesuffix
+        self.modelpath = os.path.join(self.modelroot, self.modelname)
+
+        try:
+            os.makedirs(self.modelpath)
+        except OSError:
+            pass
+
+    def makenet(self):
+        """
+        Create and initialize a convolutional autoencoder model.
+
+        This method constructs a ConvAutoencoderModel with the specified parameters
+        and moves it to the designated device (CPU or GPU).
+
+        Parameters
+        ----------
+        self : object
+            The instance of the class containing this method. Expected to have the
+            following attributes:
+            - window_size : int
+                Size of the input window
+            - encoding_dim : int
+                Dimension of the encoded representation
+            - num_filters : int
+                Number of filters in the convolutional layers
+            - kernel_size : int
+                Size of the convolutional kernel
+            - dropout_rate : float
+                Dropout rate for regularization
+            - activation : str or callable
+                Activation function to use
+            - inputsize : tuple
+                Input size dimensions
+            - device : torch.device
+                Device to move the model to (CPU or GPU)
+
+        Returns
+        -------
+        None
+            This method does not return any value. It initializes the model
+            attribute of the class instance.
+
+        Notes
+        -----
+        The method assumes that `ConvAutoencoderModel` is a valid class that accepts
+        the specified parameters. The model is automatically moved to the device
+        specified by `self.device`.
+
+        Examples
+        --------
+        >>> class MyModel:
+        ...     def __init__(self):
+        ...         self.window_size = 100
+        ...         self.encoding_dim = 32
+        ...         self.num_filters = 64
+        ...         self.kernel_size = 3
+        ...         self.dropout_rate = 0.2
+        ...         self.activation = 'relu'
+        ...         self.inputsize = (1, 100)
+        ...         self.device = torch.device('cpu')
+        ...         self.model = None
+        ...
+        ...     def makenet(self):
+        ...         self.model = ConvAutoencoderModel(
+        ...             self.window_size,
+        ...             self.encoding_dim,
+        ...             self.num_filters,
+        ...             self.kernel_size,
+        ...             self.dropout_rate,
+        ...             self.activation,
+        ...             self.inputsize,
+        ...         )
+        ...         self.model.to(self.device)
+        ...
+        >>> model = MyModel()
+        >>> model.makenet()
+        >>> print(model.model)
+        """
+        self.model = ConvAutoencoderModel(
+            self.window_size,
+            self.encoding_dim,
+            self.num_filters,
+            self.kernel_size,
+            self.dropout_rate,
+            self.activation,
+            self.inputsize,
+        )
+        self.model.to(self.device)
+
+
+class CRNNModel(nn.Module):
+    def __init__(
+        self, num_filters, kernel_size, encoding_dim, dropout_rate, activation, inputsize
+    ):
+        """
+        Initialize the CRNNModel.
+
+        This function initializes a Convolutional Recurrent Neural Network (CRNN) model
+        with convolutional front-end, bidirectional LSTM layers, and output mapping.
+        The model processes sequential data through convolutional layers, applies
+        bidirectional LSTM encoding, and maps the output back to the original input size.
+
+        Parameters
+        ----------
+        num_filters : int
+            Number of filters in the convolutional layers
+        kernel_size : int
+            Size of the convolutional kernel
+        encoding_dim : int
+            Dimension of the LSTM encoding (hidden state size)
+        dropout_rate : float
+            Dropout rate for regularization
+        activation : str
+            Activation function to use ('relu' or 'tanh')
+        inputsize : int
+            Size of the input features
+
+        Returns
+        -------
+        None
+            Initializes the CRNNModel instance
+
+        Notes
+        -----
+        The model uses a bidirectional LSTM with batch_first=True.
+        The convolutional layers use 'same' padding to maintain sequence length.
+        Default activation function is ReLU if an invalid activation is provided.
+
+        Examples
+        --------
+        >>> model = CRNNModel(
+        ...     num_filters=32,
+        ...     kernel_size=3,
+        ...     encoding_dim=64,
+        ...     dropout_rate=0.2,
+        ...     activation='relu',
+        ...     inputsize=128
+        ... )
+        """
+        super(CRNNModel, self).__init__()
+
+        self.num_filters = num_filters
+        self.kernel_size = kernel_size
+        self.encoding_dim = encoding_dim
+        self.dropout_rate = dropout_rate
+        self.activation = activation
+        self.inputsize = inputsize
+
+        # Get activation function
+        if activation == "relu":
+            act_fn = nn.ReLU
+        elif activation == "tanh":
+            act_fn = nn.Tanh
+        else:
+            act_fn = nn.ReLU
+
+        # Convolutional front-end
+        self.conv1 = nn.Conv1d(inputsize, num_filters, kernel_size, padding="same")
+        self.bn1 = nn.BatchNorm1d(num_filters)
+        self.dropout1 = nn.Dropout(dropout_rate)
+        self.act1 = act_fn()
+
+        self.conv2 = nn.Conv1d(num_filters, num_filters * 2, kernel_size, padding="same")
+        self.bn2 = nn.BatchNorm1d(num_filters * 2)
+        self.dropout2 = nn.Dropout(dropout_rate)
+        self.act2 = act_fn()
+
+        # Bidirectional LSTM
+        self.lstm = nn.LSTM(num_filters * 2, encoding_dim, batch_first=True, bidirectional=True)
+
+        # Output mapping
+        self.fc_out = nn.Linear(encoding_dim * 2, inputsize)
+
+    def forward(self, x):
+        """
+        Forward pass through the neural network architecture.
+
+        This function processes input data through a convolutional neural network
+        followed by an LSTM layer and a fully connected output layer. The input
+        is first processed through two convolutional blocks, then reshaped for
+        LSTM processing, and finally converted back to the original output format.
+
+        Parameters
+        ----------
+        x : torch.Tensor
+            Input tensor of shape (batch_size, channels, length) containing
+            the input sequence data to be processed.
+
+        Returns
+        -------
+        torch.Tensor
+            Output tensor of shape (batch_size, channels, length) containing
+            the processed sequence data after passing through all layers.
+
+        Notes
+        -----
+        The function performs the following operations in sequence:
+        1. Two convolutional blocks with batch normalization, dropout, and activation
+        2. Permute operation to reshape data for LSTM processing (batch, seq_len, features)
+        3. LSTM layer processing
+        4. Fully connected output layer
+        5. Final permutation to restore original shape (batch, channels, length)
+
+        Examples
+        --------
+        >>> import torch
+        >>> model = YourModelClass()
+        >>> x = torch.randn(32, 1, 100)  # batch_size=32, channels=1, length=100
+        >>> output = model.forward(x)
+        >>> print(output.shape)  # torch.Size([32, 1, 100])
+        """
+        # Conv layers expect (batch, channels, length)
+        x = self.conv1(x)
+        x = self.bn1(x)
+        x = self.dropout1(x)
+        x = self.act1(x)
+
+        x = self.conv2(x)
+        x = self.bn2(x)
+        x = self.dropout2(x)
+        x = self.act2(x)
+
+        # LSTM expects (batch, seq_len, features)
+        x = x.permute(0, 2, 1)
+        x, _ = self.lstm(x)
+
+        # Output layer
+        x = self.fc_out(x)
+
+        # Convert back to (batch, channels, length)
+        x = x.permute(0, 2, 1)
+
+        return x
+
+    def get_config(self):
+        """
+        Get the configuration parameters of the model.
+
+        Returns
+        -------
+        dict
+            A dictionary containing the model configuration parameters with the following keys:
+            - "num_filters" (int): Number of filters in the convolutional layers
+            - "kernel_size" (int): Size of the convolutional kernel
+            - "encoding_dim" (int): Dimension of the encoding layer
+            - "dropout_rate" (float): Dropout rate for regularization
+            - "activation" (str): Activation function used in the layers
+            - "inputsize" (int): Size of the input data
+
+        Notes
+        -----
+        This method returns a copy of the current configuration parameters. Modifications
+        to the returned dictionary will not affect the original model configuration.
+
+        Examples
+        --------
+        >>> model = MyModel()
+        >>> config = model.get_config()
+        >>> print(config['num_filters'])
+        32
+        """
+        return {
+            "num_filters": self.num_filters,
+            "kernel_size": self.kernel_size,
+            "encoding_dim": self.encoding_dim,
+            "dropout_rate": self.dropout_rate,
+            "activation": self.activation,
+            "inputsize": self.inputsize,
+        }
+
+
+class CRNNDLFilter(DeepLearningFilter):
+    def __init__(
+        self,
+        encoding_dim: int = 10,
+        num_filters: int = 10,
+        kernel_size: int = 5,
+        dilation_rate: int = 1,
+        *args,
+        **kwargs,
+    ) -> None:
+        """
+        Initialize CRNNDLFilter layer.
+
+        Parameters
+        ----------
+        encoding_dim : int, optional
+            Dimension of the encoding layer, by default 10
+        num_filters : int, optional
+            Number of filters in the convolutional layer, by default 10
+        kernel_size : int, optional
+            Size of the convolutional kernel, by default 5
+        dilation_rate : int, optional
+            Dilation rate for the convolutional layer, by default 1
+        *args : tuple
+            Variable length argument list
+        **kwargs : dict
+            Arbitrary keyword arguments
+
+        Returns
+        -------
+        None
+            This method does not return any value
+
+        Notes
+        -----
+        This constructor initializes a CRNN (Convolutional Recurrent Neural Network)
+        with dilated filters. The layer type is set to "crnn" and configuration
+        parameters are stored in infodict for later reference.
+
+        Examples
+        --------
+        >>> layer = CRNNDLFilter(encoding_dim=20, num_filters=15, kernel_size=3)
+        >>> print(layer.nettype)
+        'crnn'
+        """
+        self.num_filters = num_filters
+        self.kernel_size = kernel_size
+        self.dilation_rate = dilation_rate
+        self.encoding_dim = encoding_dim
+        self.nettype = "crnn"
+        self.infodict["nettype"] = self.nettype
+        self.infodict["num_filters"] = self.num_filters
+        self.infodict["kernel_size"] = self.kernel_size
+        self.infodict["encoding_dim"] = self.encoding_dim
+        super(CRNNDLFilter, self).__init__(*args, **kwargs)
+
+    def getname(self):
+        """
+        Generate and configure model name and path based on configuration parameters.
+
+        This method constructs a descriptive model name string based on various configuration
+        parameters and creates the corresponding model directory path. The generated name
+        includes information about window size, encoding dimensions, filters, kernel size,
+        epochs, thresholds, step size, and activation function.
+
+        Parameters
+        ----------
+        self : object
+            The instance containing configuration parameters for model naming.
+
+        Returns
+        -------
+        None
+            This method modifies instance attributes in-place and does not return a value.
+
+        Notes
+        -----
+        The generated model name follows a consistent naming convention:
+        'model_crnn_pytorch_wXXX_enXXX_fnXX_flXX_eXXX_tX_ctX_sX_activation'
+        where XXX represents zero-padded numeric values and X represents single digits.
+
+        Additional suffixes are appended based on:
+        - usebadpts: '_usebadpts' if True
+        - excludebysubject: '_excludebysubject' if True
+        - namesuffix: '_{suffix}' if not None
+
+        Examples
+        --------
+        >>> model = ModelClass()
+        >>> model.window_size = 100
+        >>> model.encoding_dim = 128
+        >>> model.num_filters = 32
+        >>> model.kernel_size = 5
+        >>> model.num_epochs = 100
+        >>> model.excludethresh = 0.5
+        >>> model.corrthresh = 0.8
+        >>> model.step = 10
+        >>> model.activation = 'relu'
+        >>> model.modelroot = '/path/to/models'
+        >>> model.getname()
+        >>> print(model.modelname)
+        'model_crnn_pytorch_w100_en128_fn32_fl05_e100_t0.5_ct0.8_s10_relu'
+        """
+        self.modelname = "_".join(
+            [
+                "model",
+                "crnn",
+                "pytorch",
+                "w" + str(self.window_size).zfill(3),
+                "en" + str(self.encoding_dim).zfill(3),
+                "fn" + str(self.num_filters).zfill(2),
+                "fl" + str(self.kernel_size).zfill(2),
+                "e" + str(self.num_epochs).zfill(3),
+                "t" + str(self.excludethresh),
+                "ct" + str(self.corrthresh),
+                "s" + str(self.step),
+                self.activation,
+            ]
+        )
+        if self.usebadpts:
+            self.modelname += "_usebadpts"
+        if self.excludebysubject:
+            self.modelname += "_excludebysubject"
+        if self.namesuffix is not None:
+            self.modelname += "_" + self.namesuffix
+        self.modelpath = os.path.join(self.modelroot, self.modelname)
+
+        try:
+            os.makedirs(self.modelpath)
+        except OSError:
+            pass
+
+    def makenet(self):
+        """
+        Create and initialize a CRNN model for neural network training.
+
+        This method initializes a CRNN (Convolutional Recurrent Neural Network) model
+        using the specified configuration parameters and moves it to the designated
+        device (CPU or GPU).
+
+        Parameters
+        ----------
+        self : object
+            The instance of the class containing this method. Expected to have the
+            following attributes:
+            - num_filters : int
+                Number of filters in the convolutional layers
+            - kernel_size : int or tuple
+                Size of the convolutional kernel
+            - encoding_dim : int
+                Dimension of the encoding layer
+            - dropout_rate : float
+                Dropout rate for regularization
+            - activation : str or callable
+                Activation function to use
+            - inputsize : tuple
+                Input dimensions for the model
+            - device : torch.device
+                Device to move the model to (CPU or GPU)
+
+        Returns
+        -------
+        None
+            This method does not return any value. It initializes the model attribute
+            of the class instance.
+
+        Notes
+        -----
+        The method assumes that the CRNNModel class is properly imported and available
+        in the namespace. The model is automatically moved to the device specified
+        in self.device.
+
+        Examples
+        --------
+        >>> class MyModel:
+        ...     def __init__(self):
+        ...         self.num_filters = 32
+        ...         self.kernel_size = 3
+        ...         self.encoding_dim = 128
+        ...         self.dropout_rate = 0.2
+        ...         self.activation = 'relu'
+        ...         self.inputsize = (1, 28, 28)
+        ...         self.device = torch.device('cpu')
+        ...
+        ...     def makenet(self):
+        ...         self.model = CRNNModel(
+        ...             self.num_filters,
+        ...             self.kernel_size,
+        ...             self.encoding_dim,
+        ...             self.dropout_rate,
+        ...             self.activation,
+        ...             self.inputsize,
+        ...         )
+        ...         self.model.to(self.device)
+        ...
+        >>> model = MyModel()
+        >>> model.makenet()
+        >>> print(model.model)
+        """
+        self.model = CRNNModel(
+            self.num_filters,
+            self.kernel_size,
+            self.encoding_dim,
+            self.dropout_rate,
+            self.activation,
+            self.inputsize,
+        )
+        self.model.to(self.device)
+
+
+class LSTMModel(nn.Module):
+    def __init__(self, num_units, num_layers, dropout_rate, window_size, inputsize):
+        """
+        Initialize the LSTMModel with specified architecture parameters.
+
+        Parameters
+        ----------
+        num_units : int
+            Number of units in each LSTM layer
+        num_layers : int
+            Number of LSTM layers in the model
+        dropout_rate : float
+            Dropout rate for LSTM layers (applied only if num_layers > 1)
+        window_size : int
+            Size of the sliding window used for sequence processing
+        inputsize : int
+            Dimensionality of input features
+
+        Returns
+        -------
+        None
+            Initializes the LSTMModel instance with the specified architecture
+
+        Notes
+        -----
+        This constructor creates a bidirectional LSTM model with residual connections.
+        The model uses LSTM layers with bidirectional processing and time-distributed
+        dense layers for output transformation. Dropout is applied between layers
+        when multiple layers are present.
+
+        Examples
+        --------
+        >>> model = LSTMModel(num_units=128, num_layers=2, dropout_rate=0.2,
+        ...                   window_size=10, inputsize=20)
+        >>> print(model)
+        """
+        super(LSTMModel, self).__init__()
+
+        self.num_units = num_units
+        self.num_layers = num_layers
+        self.dropout_rate = dropout_rate
+        self.window_size = window_size
+        self.inputsize = inputsize
+
+        self.lstm_layers = nn.ModuleList()
+        self.dense_layers = nn.ModuleList()
+
+        for _ in range(num_layers):
+            # Bidirectional LSTM
+            self.lstm_layers.append(
+                nn.LSTM(
+                    inputsize if len(self.lstm_layers) == 0 else inputsize,
+                    num_units,
+                    batch_first=True,
+                    bidirectional=True,
+                    dropout=dropout_rate if num_layers > 1 else 0,
+                )
+            )
+            # Time-distributed dense layer
+            self.dense_layers.append(nn.Linear(num_units * 2, inputsize))
+
+    def forward(self, x):
+        """
+        Forward pass through LSTM and dense layers.
+
+        Apply a sequence of LSTM layers followed by dense layers to the input tensor,
+        with appropriate dimension permutations to maintain correct data flow.
+
+        Parameters
+        ----------
+        x : torch.Tensor
+            Input tensor with shape (batch, channels, length) containing the sequential data.
+
+        Returns
+        -------
+        torch.Tensor
+            Output tensor with shape (batch, channels, length) after processing through
+            LSTM and dense layers.
+
+        Notes
+        -----
+        The function performs the following operations:
+        1. Permutes input from (batch, channels, length) to (batch, length, channels)
+        2. Processes through LSTM layers sequentially
+        3. Applies dense layers to each time step
+        4. Permutes output back to (batch, channels, length)
+
+        Examples
+        --------
+        >>> import torch
+        >>> # Assuming self.lstm_layers and self.dense_layers are initialized
+        >>> x = torch.randn(32, 128, 100)  # batch=32, channels=128, length=100
+        >>> output = model.forward(x)
+        >>> output.shape
+        torch.Size([32, 128, 100])
+        """
+        # x is (batch, channels, length), convert to (batch, length, channels)
+        x = x.permute(0, 2, 1)
+
+        for lstm, dense in zip(self.lstm_layers, self.dense_layers):
+            x, _ = lstm(x)
+            # Apply dense layer across time steps
+            x = dense(x)
+
+        # Convert back to (batch, channels, length)
+        x = x.permute(0, 2, 1)
+
+        return x
+
+    def get_config(self):
+        """
+        Get the configuration parameters of the model.
+
+        Returns
+        -------
+        dict
+            A dictionary containing the model configuration parameters with the following keys:
+            - "num_units" (int): Number of units in each layer
+            - "num_layers" (int): Number of layers in the model
+            - "dropout_rate" (float): Dropout rate for regularization
+            - "window_size" (int): Size of the sliding window for sequence processing
+            - "inputsize" (int): Size of the input features
+
+        Notes
+        -----
+        This method returns a copy of the internal configuration parameters.
+        The returned dictionary can be used to recreate the model with the same configuration.
+
+        Examples
+        --------
+        >>> config = model.get_config()
+        >>> print(config['num_units'])
+        128
+        >>> new_model = ModelClass(**config)
+        """
+        return {
+            "num_units": self.num_units,
+            "num_layers": self.num_layers,
+            "dropout_rate": self.dropout_rate,
+            "window_size": self.window_size,
+            "inputsize": self.inputsize,
+        }
+
+
+class LSTMDLFilter(DeepLearningFilter):
+    def __init__(self, num_units: int = 16, *args, **kwargs) -> None:
+        """
+        Initialize the LSTMDLFilter layer.
+
+        Parameters
+        ----------
+        num_units : int, optional
+            Number of units in the LSTM layer, by default 16
+        *args
+            Variable length argument list passed to parent class
+        **kwargs
+            Arbitrary keyword arguments passed to parent class
+
+        Returns
+        -------
+        None
+            This method initializes the instance and does not return any value
+
+        Notes
+        -----
+        This constructor sets up the LSTM layer with specified number of units and
+        initializes the network type identifier. The infodict is updated with both
+        the network type and number of units for tracking purposes.
+
+        Examples
+        --------
+        >>> layer = LSTMDLFilter(num_units=32)
+        >>> print(layer.num_units)
+        32
+        >>> print(layer.nettype)
+        'lstm'
+        """
+        self.num_units = num_units
+        self.nettype = "lstm"
+        self.infodict["nettype"] = self.nettype
+        self.infodict["num_units"] = self.num_units
+        super(LSTMDLFilter, self).__init__(*args, **kwargs)
+
+    def getname(self):
+        """
+        Generate and configure model name and path based on current parameters.
+
+        This method constructs a descriptive model name string using various
+        hyperparameters and configuration settings. It then creates the
+        corresponding directory path and ensures it exists.
+
+        Parameters
+        ----------
+        self : object
+            The instance containing model configuration attributes.
+
+        Returns
+        -------
+        None
+            This method modifies instance attributes in-place and does not return a value.
+
+        Notes
+        -----
+        The generated model name follows a specific format:
+        "model_lstm_pytorch_wXXX_lYY_nuZZZ_dDD_rdDD_eFFF_tT_ctTT_sS"
+        where XXX, YY, ZZZ, DD, FF, T, TT, S represent formatted parameter values.
+
+        Examples
+        --------
+        >>> model = MyModel()
+        >>> model.window_size = 100
+        >>> model.num_layers = 2
+        >>> model.num_units = 128
+        >>> model.dropout_rate = 0.2
+        >>> model.num_epochs = 100
+        >>> model.excludethresh = 0.5
+        >>> model.corrthresh = 0.8
+        >>> model.step = 1
+        >>> model.excludebysubject = True
+        >>> model.getname()
+        >>> print(model.modelname)
+        'model_lstm_pytorch_w100_l02_nu128_d02_rd02_e100_t05_ct08_s1_excludebysubject'
+        """
+        self.modelname = "_".join(
+            [
+                "model",
+                "lstm",
+                "pytorch",
+                "w" + str(self.window_size).zfill(3),
+                "l" + str(self.num_layers).zfill(2),
+                "nu" + str(self.num_units),
+                "d" + str(self.dropout_rate),
+                "rd" + str(self.dropout_rate),
+                "e" + str(self.num_epochs).zfill(3),
+                "t" + str(self.excludethresh),
+                "ct" + str(self.corrthresh),
+                "s" + str(self.step),
+            ]
+        )
+        if self.excludebysubject:
+            self.modelname += "_excludebysubject"
+        self.modelpath = os.path.join(self.modelroot, self.modelname)
+
+        try:
+            os.makedirs(self.modelpath)
+        except OSError:
+            pass
+
+    def makenet(self):
+        """
+        Create and initialize an LSTM model for neural network training.
+
+        This method initializes an LSTMModel with the specified architecture parameters
+        and moves the model to the designated device (CPU or GPU).
+
+        Parameters
+        ----------
+        self : object
+            The instance containing the following attributes:
+            - num_units : int
+                Number of units in each LSTM layer
+            - num_layers : int
+                Number of LSTM layers in the model
+            - dropout_rate : float
+                Dropout rate for regularization
+            - window_size : int
+                Size of the input window for time series data
+            - inputsize : int
+                Size of the input features
+            - device : torch.device
+                Device to move the model to (e.g., 'cpu' or 'cuda')
+
+        Returns
+        -------
+        None
+            This method does not return any value. It initializes the model attribute
+            and moves it to the specified device.
+
+        Notes
+        -----
+        The method creates an LSTMModel instance with the following parameters:
+        - num_units: Number of hidden units in LSTM layers
+        - num_layers: Number of stacked LSTM layers
+        - dropout_rate: Dropout probability for regularization
+        - window_size: Input sequence length
+        - inputsize: Feature dimension of input data
+
+        Examples
+        --------
+        >>> # Assuming self is an instance with required attributes
+        >>> self.makenet()
+        >>> # Model is now initialized and moved to specified device
+        """
+        self.model = LSTMModel(
+            self.num_units,
+            self.num_layers,
+            self.dropout_rate,
+            self.window_size,
+            self.inputsize,
+        )
+        self.model.to(self.device)
+
+
+class HybridModel(nn.Module):
+    def __init__(
+        self,
+        num_filters,
+        kernel_size,
+        num_units,
+        num_layers,
+        dropout_rate,
+        activation,
+        inputsize,
+        window_size,
+        invert,
+    ):
+        """
+        Initialize the HybridModel with configurable CNN-LSTM architecture.
+
+        Parameters
+        ----------
+        num_filters : int
+            Number of filters in the convolutional layers.
+        kernel_size : int
+            Size of the convolutional kernel.
+        num_units : int
+            Number of units in the LSTM layers.
+        num_layers : int
+            Total number of layers in the model.
+        dropout_rate : float
+            Dropout rate for regularization.
+        activation : str
+            Activation function to use; options are 'relu' or 'tanh'.
+        inputsize : int
+            Size of the input features.
+        window_size : int
+            Size of the sliding window for input data.
+        invert : bool
+            If True, applies CNN first followed by LSTM. Otherwise, applies LSTM first followed by CNN.
+
+        Returns
+        -------
+        None
+            This method initializes the model's layers and components but does not return any value.
+
+        Notes
+        -----
+        The model supports two architectures:
+        - If `invert=False`: LSTM → CNN
+        - If `invert=True`: CNN → LSTM
+
+        Examples
+        --------
+        >>> model = HybridModel(
+        ...     num_filters=64,
+        ...     kernel_size=3,
+        ...     num_units=128,
+        ...     num_layers=3,
+        ...     dropout_rate=0.2,
+        ...     activation="relu",
+        ...     inputsize=10,
+        ...     window_size=100,
+        ...     invert=True
+        ... )
+        """
+        super(HybridModel, self).__init__()
+
+        self.num_filters = num_filters
+        self.kernel_size = kernel_size
+        self.num_units = num_units
+        self.num_layers = num_layers
+        self.dropout_rate = dropout_rate
+        self.activation = activation
+        self.inputsize = inputsize
+        self.window_size = window_size
+        self.invert = invert
+
+        # Get activation function
+        if activation == "relu":
+            act_fn = nn.ReLU
+        elif activation == "tanh":
+            act_fn = nn.Tanh
+        else:
+            act_fn = nn.ReLU
+
+        self.layers = nn.ModuleList()
+
+        if invert:
+            # CNN first, then LSTM
+            # Input layer
+            self.layers.append(nn.Conv1d(inputsize, num_filters, kernel_size, padding="same"))
+            self.layers.append(nn.BatchNorm1d(num_filters))
+            self.layers.append(nn.Dropout(dropout_rate))
+            self.layers.append(act_fn())
+
+            # Intermediate CNN layers
+            for _ in range(num_layers - 2):
+                self.layers.append(
+                    nn.Conv1d(num_filters, num_filters, kernel_size, padding="same")
+                )
+                self.layers.append(nn.BatchNorm1d(num_filters))
+                self.layers.append(nn.Dropout(dropout_rate))
+                self.layers.append(act_fn())
+
+            # LSTM layer
+            self.lstm = nn.LSTM(
+                num_filters, num_units, batch_first=True, bidirectional=True, dropout=dropout_rate
+            )
+            self.lstm_dense = nn.Linear(num_units * 2, inputsize)
+
+        else:
+            # LSTM first, then CNN
+            self.lstm = nn.LSTM(
+                inputsize, num_units, batch_first=True, bidirectional=True, dropout=dropout_rate
+            )
+            self.lstm_dense = nn.Linear(num_units * 2, inputsize)
+            self.lstm_dropout = nn.Dropout(dropout_rate)
+
+            # Intermediate CNN layers
+            for _ in range(num_layers - 2):
+                self.layers.append(nn.Conv1d(inputsize, num_filters, kernel_size, padding="same"))
+                self.layers.append(nn.BatchNorm1d(num_filters))
+                self.layers.append(nn.Dropout(dropout_rate))
+                self.layers.append(act_fn())
+
+            # Output layer
+            self.output_conv = nn.Conv1d(
+                num_filters if num_layers > 2 else inputsize,
+                inputsize,
+                kernel_size,
+                padding="same",
+            )
+
+    def forward(self, x):
+        """
+        Forward pass of the model with optional CNN-LSTM hybrid architecture.
+
+        This method implements a flexible forward pass that can operate in two modes
+        depending on the `invert` flag. When `invert` is True, the sequence processing
+        follows CNN → LSTM → CNN structure. When `invert` is False, the sequence
+        processing follows LSTM → CNN structure.
+
+        Parameters
+        ----------
+        x : torch.Tensor
+            Input tensor of shape (batch_size, channels, sequence_length) or
+            (batch_size, sequence_length, channels) depending on the processing path.
+
+        Returns
+        -------
+        torch.Tensor
+            Output tensor with the same batch dimension as input, with shape
+            dependent on the specific layers and processing path used.
+
+        Notes
+        -----
+        The function handles different tensor permutations based on the processing
+        order:
+        - CNN → LSTM path: permutes from (B, C, L) to (B, L, C) for LSTM, then back
+        - LSTM → CNN path: permutes from (B, C, L) to (B, L, C) for LSTM, then back
+        The `invert` flag determines which processing order is used.
+
+        Examples
+        --------
+        >>> model = MyModel()
+        >>> x = torch.randn(32, 10, 128)  # batch_size=32, seq_len=10, features=128
+        >>> output = model.forward(x)
+        >>> print(output.shape)
+        torch.Size([32, 10, 256])
+        """
+        if self.invert:
+            # Apply CNN layers
+            for layer in self.layers:
+                x = layer(x)
+
+            # LSTM expects (batch, seq_len, features)
+            x = x.permute(0, 2, 1)
+            x, _ = self.lstm(x)
+            x = self.lstm_dense(x)
+
+            # Convert back to (batch, channels, length)
+            x = x.permute(0, 2, 1)
+
+        else:
+            # LSTM first
+            x = x.permute(0, 2, 1)
+            x, _ = self.lstm(x)
+            x = self.lstm_dense(x)
+            x = self.lstm_dropout(x)
+            x = x.permute(0, 2, 1)
+
+            # CNN layers
+            for layer in self.layers:
+                x = layer(x)
+
+            # Output layer
+            if hasattr(self, "output_conv"):
+                x = self.output_conv(x)
+
+        return x
+
+    def get_config(self):
+        """
+        Get the configuration parameters of the model.
+
+        Returns
+        -------
+        dict
+            A dictionary containing all configuration parameters with their current values:
+            - num_filters: int, number of filters in the convolutional layers
+            - kernel_size: int, size of the convolutional kernel
+            - num_units: int, number of units in the dense layers
+            - num_layers: int, number of layers in the model
+            - dropout_rate: float, dropout rate for regularization
+            - activation: str or callable, activation function to use
+            - inputsize: int, size of the input features
+            - window_size: int, size of the sliding window
+            - invert: bool, whether to invert the input data
+
+        Notes
+        -----
+        This method returns a copy of the internal configuration dictionary.
+        Modifications to the returned dictionary will not affect the original model configuration.
+
+        Examples
+        --------
+        >>> config = model.get_config()
+        >>> print(config['num_filters'])
+        32
+        """
+        return {
+            "num_filters": self.num_filters,
+            "kernel_size": self.kernel_size,
+            "num_units": self.num_units,
+            "num_layers": self.num_layers,
+            "dropout_rate": self.dropout_rate,
+            "activation": self.activation,
+            "inputsize": self.inputsize,
+            "window_size": self.window_size,
+            "invert": self.invert,
+        }
+
+
+class HybridDLFilter(DeepLearningFilter):
+    def __init__(
+        self,
+        invert: bool = False,
+        num_filters: int = 10,
+        kernel_size: int = 5,
+        num_units: int = 16,
+        *args,
+        **kwargs,
+    ) -> None:
+        """
+        Initialize HybridDLFilter layer.
+
+        Parameters
+        ----------
+        invert : bool, default=False
+            If True, inverts the filter response. If False, applies normal filtering.
+        num_filters : int, default=10
+            Number of filters to apply in the convolutional layer.
+        kernel_size : int, default=5
+            Size of the convolutional kernel.
+        num_units : int, default=16
+            Number of units in the dense layer.
+        *args
+            Variable length argument list.
+        **kwargs
+            Arbitrary keyword arguments.
+
+        Returns
+        -------
+        None
+            This method initializes the HybridDLFilter instance and does not return any value.
+
+        Notes
+        -----
+        This constructor sets up a hybrid deep learning filter that combines convolutional
+        and dense layers. The infodict dictionary is populated with configuration parameters
+        for tracking and logging purposes.
+
+        Examples
+        --------
+        >>> filter_layer = HybridDLFilter(
+        ...     invert=True,
+        ...     num_filters=20,
+        ...     kernel_size=3,
+        ...     num_units=32
+        ... )
+        """
+        self.invert = invert
+        self.num_filters = num_filters
+        self.kernel_size = kernel_size
+        self.num_units = num_units
+        self.nettype = "hybrid"
+        self.infodict["nettype"] = self.nettype
+        self.infodict["num_filters"] = self.num_filters
+        self.infodict["kernel_size"] = self.kernel_size
+        self.infodict["invert"] = self.invert
+        self.infodict["num_units"] = self.num_units
+        super(HybridDLFilter, self).__init__(*args, **kwargs)
+
+    def getname(self):
+        """
+        Generate and configure the model name and path based on current parameters.
+
+        This method constructs a descriptive model name string using various
+        hyperparameters and configuration settings. The generated name follows
+        a standardized format that includes window size, layer count, filter count,
+        kernel size, number of units, dropout rates, number of epochs, threshold
+        values, step size, and activation function. The method also creates the
+        corresponding model directory path and ensures it exists.
+
+        Parameters
+        ----------
+        self : object
+            The instance of the class containing the model configuration attributes.
+            Required attributes include:
+            - window_size : int
+            - num_layers : int
+            - num_filters : int
+            - kernel_size : int
+            - num_units : int
+            - dropout_rate : float
+            - num_epochs : int
+            - excludethresh : float
+            - corrthresh : float
+            - step : int
+            - activation : str
+            - invert : bool
+            - excludebysubject : bool
+            - modelroot : str
+
+        Returns
+        -------
+        None
+            This method modifies the instance attributes in-place:
+            - self.modelname : str
+            - self.modelpath : str
+
+        Notes
+        -----
+        The model name format follows this pattern:
+        "model_hybrid_pytorch_wXXX_lYY_fnZZ_flZZ_nuZZ_dZZ_rdZZ_eXXX_tX_ctX_sX_activation"
+        where XXX, YY, ZZ, etc. represent zero-padded numerical values.
+
+        Additional suffixes are appended based on:
+        - "_invert" if self.invert is True
+        - "_excludebysubject" if self.excludebysubject is True
+
+        Examples
+        --------
+        >>> model = MyModel()
+        >>> model.window_size = 100
+        >>> model.num_layers = 2
+        >>> model.getname()
+        >>> print(model.modelname)
+        'model_hybrid_pytorch_w100_l02_fn08_fl08_nu128_d05_rd05_e100_t05_ct08_s1_relu'
+        """
+        self.modelname = "_".join(
+            [
+                "model",
+                "hybrid",
+                "pytorch",
+                "w" + str(self.window_size).zfill(3),
+                "l" + str(self.num_layers).zfill(2),
+                "fn" + str(self.num_filters).zfill(2),
+                "fl" + str(self.kernel_size).zfill(2),
+                "nu" + str(self.num_units),
+                "d" + str(self.dropout_rate),
+                "rd" + str(self.dropout_rate),
+                "e" + str(self.num_epochs).zfill(3),
+                "t" + str(self.excludethresh),
+                "ct" + str(self.corrthresh),
+                "s" + str(self.step),
+                self.activation,
+            ]
+        )
+        if self.invert:
+            self.modelname += "_invert"
+        if self.excludebysubject:
+            self.modelname += "_excludebysubject"
+        self.modelpath = os.path.join(self.modelroot, self.modelname)
+
+        try:
+            os.makedirs(self.modelpath)
+        except OSError:
+            pass
+
+    def makenet(self):
+        """
+        Create and initialize a hybrid neural network model.
+
+        This method constructs a HybridModel with the specified architecture parameters
+        and moves it to the designated device (CPU or GPU).
+
+        Parameters
+        ----------
+        self : object
+            The instance containing the following attributes:
+            - num_filters : int
+                Number of filters in the convolutional layers
+            - kernel_size : int
+                Size of the convolutional kernels
+            - num_units : int
+                Number of units in the dense layers
+            - num_layers : int
+                Number of layers in the model
+            - dropout_rate : float
+                Dropout rate for regularization
+            - activation : str or callable
+                Activation function to use
+            - inputsize : int
+                Size of the input features
+            - window_size : int
+                Size of the sliding window
+            - invert : bool
+                Whether to invert the model architecture
+
+        Returns
+        -------
+        None
+            This method does not return any value. It initializes the model attribute
+            and moves it to the specified device.
+
+        Notes
+        -----
+        The method assumes that the instance has all required attributes set before
+        calling. The model is moved to the device specified by `self.device`.
+
+        Examples
+        --------
+        >>> model = MyModel()
+        >>> model.num_filters = 32
+        >>> model.kernel_size = 3
+        >>> model.num_units = 64
+        >>> model.num_layers = 2
+        >>> model.dropout_rate = 0.2
+        >>> model.activation = 'relu'
+        >>> model.inputsize = 10
+        >>> model.window_size = 5
+        >>> model.invert = False
+        >>> model.device = 'cuda'
+        >>> model.makenet()
+        >>> print(model.model)
+        """
+        self.model = HybridModel(
+            self.num_filters,
+            self.kernel_size,
+            self.num_units,
+            self.num_layers,
+            self.dropout_rate,
+            self.activation,
+            self.inputsize,
+            self.window_size,
+            self.invert,
+        )
+        self.model.to(self.device)
+
+
+def filtscale(
+    data: np.ndarray,
+    scalefac: float = 1.0,
+    reverse: bool = False,
+    hybrid: bool = False,
+    lognormalize: bool = True,
+    epsilon: float = 1e-10,
+    numorders: int = 6,
+) -> tuple[np.ndarray, float] | np.ndarray:
+    """
+    Apply or reverse a scaling transformation to spectral data.
+
+    This function performs either forward or inverse scaling of input data,
+    typically used in signal processing or spectral analysis. In forward mode,
+    it computes the FFT of the input data and applies normalization and scaling
+    to the magnitude and phase components. In reverse mode, it reconstructs
+    the original time-domain signal from scaled magnitude and phase components.
+
+    Parameters
+    ----------
+    data : np.ndarray
+        Input time-domain signal or scaled spectral data depending on `reverse` flag.
+    scalefac : float, optional
+        Scaling factor used in normalization. Default is 1.0.
+    reverse : bool, optional
+        If True, performs inverse transformation to reconstruct the original signal.
+        If False, performs forward transformation. Default is False.
+    hybrid : bool, optional
+        If True, returns a hybrid output combining original signal and magnitude.
+        Only applicable in forward mode. Default is False.
+    lognormalize : bool, optional
+        If True, applies logarithmic normalization to the magnitude. Default is True.
+    epsilon : float, optional
+        Small constant added to magnitude before log to avoid log(0). Default is 1e-10.
+    numorders : int, optional
+        Number of orders used in normalization scaling. Default is 6.
+
+    Returns
+    -------
+    tuple[np.ndarray, float] or np.ndarray
+        - If `reverse` is False: Returns a tuple of (scaled_data, scalefac).
+          `scaled_data` is a stacked array of magnitude and phase (or original signal
+          and magnitude in hybrid mode).
+        - If `reverse` is True: Returns the reconstructed time-domain signal as
+          a numpy array.
+
+    Notes
+    -----
+    - In forward mode, the function computes the FFT of `data`, normalizes the
+      magnitude, and scales it to a range suitable for further processing.
+    - In reverse mode, the function reconstructs the time-domain signal using
+      inverse FFT from the provided scaled magnitude and phase components.
+    - The `hybrid` mode is useful for certain types of signal visualization or
+      feature extraction where both time-domain and frequency-domain information
+      are needed.
+
+    Examples
+    --------
+    >>> import numpy as np
+    >>> from scipy import fftpack
+    >>> x = np.random.randn(1024)
+    >>> scaled_data, scalefac = filtscale(x)
+    >>> reconstructed = filtscale(scaled_data, scalefac=scalefac, reverse=True)
+    """
+    if not reverse:
+        specvals = fftpack.fft(data)
+        if lognormalize:
+            themag = np.log(np.absolute(specvals) + epsilon)
+            scalefac = np.max(themag)
+            themag = (themag - scalefac + numorders) / numorders
+            themag[np.where(themag < 0.0)] = 0.0
+        else:
+            scalefac = np.std(data)
+            themag = np.absolute(specvals) / scalefac
+        thephase = np.angle(specvals)
+        thephase = thephase / (2.0 * np.pi) - 0.5
+        if hybrid:
+            return np.stack((data, themag), axis=1), scalefac
+        else:
+            return np.stack((themag, thephase), axis=1), scalefac
+    else:
+        if hybrid:
+            return data[:, 0]
+        else:
+            thephase = (data[:, 1] + 0.5) * 2.0 * np.pi
+            if lognormalize:
+                themag = np.exp(data[:, 0] * numorders - numorders + scalefac)
+            else:
+                themag = data[:, 0] * scalefac
+            specvals = themag * np.exp(1.0j * thephase)
+            return fftpack.ifft(specvals).real
+
+
+def tobadpts(name: str) -> str:
+    """
+    Convert a filename to its corresponding bad points filename.
+
+    This function takes a filename string and replaces the '.txt' extension
+    with '_badpts.txt' to create a new filename for bad points data.
+
+    Parameters
+    ----------
+    name : str
+        The input filename string, typically ending with '.txt'.
+
+    Returns
+    -------
+    str
+        The converted filename with '_badpts.txt' extension instead of '.txt'.
+
+    Notes
+    -----
+    This function is useful for creating consistent naming conventions for
+    bad points data files that correspond to original data files.
+
+    Examples
+    --------
+    >>> tobadpts("data.txt")
+    'data_badpts.txt'
+
+    >>> tobadpts("results.txt")
+    'results_badpts.txt'
+
+    >>> tobadpts("output.txt")
+    'output_badpts.txt'
+    """
+    return name.replace(".txt", "_badpts.txt")
+
+
+def targettoinput(name: str, targetfrag: str = "xyz", inputfrag: str = "abc") -> str:
+    """
+    Replace target fragment with input fragment in a string.
+
+    Parameters
+    ----------
+    name : str
+        The input string to perform replacement on.
+    targetfrag : str, default='xyz'
+        The fragment to be replaced in the input string.
+    inputfrag : str, default='abc'
+        The fragment to replace the target fragment with.
+
+    Returns
+    -------
+    str
+        The modified string with targetfrag replaced by inputfrag.
+
+    Notes
+    -----
+    This function uses Python's built-in string replace method, which replaces
+    all occurrences of the target fragment with the input fragment.
+
+    Examples
+    --------
+    >>> targettoinput("hello xyz world")
+    'hello abc world'
+
+    >>> targettoinput("test xyz xyz test", "xyz", "123")
+    'test 123 123 test'
+
+    >>> targettoinput("abcdef", "cde", "XXX")
+    'abXXXf'
+    """
+    LGR.debug(f"replacing {targetfrag} with {inputfrag}")
+    return name.replace(targetfrag, inputfrag)
+
+
+def getmatchedtcs(
+    searchstring: str,
+    usebadpts: bool = False,
+    targetfrag: str = "xyz",
+    inputfrag: str = "abc",
+    debug: bool = False,
+) -> tuple[list[str], int]:
+    """
+    Find and validate matched timecourse files based on a search pattern.
+
+    This function searches for timecourse files matching the given search string,
+    verifies their completeness by checking for associated info files, and
+    determines the length of the timecourses from the first valid file.
+
+    Parameters
+    ----------
+    searchstring : str
+        A glob pattern to match target timecourse files.
+    usebadpts : bool, optional
+        Flag indicating whether bad points should be used (default is False).
+    targetfrag : str, optional
+        Target fragment identifier (default is "xyz").
+    inputfrag : str, optional
+        Input fragment identifier (default is "abc").
+    debug : bool, optional
+        If True, prints debug information including matched files (default is False).
+
+    Returns
+    -------
+    tuple[list[str], int]
+        A tuple containing:
+        - List of matched and validated file paths.
+        - Length of the timecourses (number of timepoints).
+
+    Notes
+    -----
+    The function expects timecourse files to have a corresponding info file
+    with the same base name but with "_info" appended. Only files with complete
+    info files are considered valid.
+
+    Examples
+    --------
+    >>> matched_files, tc_length = getmatchedtcs("data/*cardiac*.tsv")
+    >>> print(f"Found {len(matched_files)} files with {tc_length} timepoints")
+    """
+    # list all of the target files
+    fromfile = sorted(glob.glob(searchstring))
+    if debug:
+        print(f"searchstring: {searchstring} -> {fromfile}")
+
+    # make sure all timecourses exist
+    # we need cardiacfromfmri_25.0Hz as x, normpleth as y, and perhaps badpts
+    matchedfilelist = []
+    for targetname in fromfile:
+        infofile = targetname.replace("_desc-stdrescardfromfmri_timeseries", "_info")
+        if os.path.isfile(infofile):
+            matchedfilelist.append(targetname)
+            print(f"{targetname} is complete")
+            LGR.debug(matchedfilelist[-1])
+        else:
+            print(f"{targetname} is incomplete")
+    print(f"found {len(matchedfilelist)} matched files")
+
+    # find out how long the files are
+    (
+        samplerate,
+        starttime,
+        columns,
+        inputarray,
+        compression,
+        columnsource,
+    ) = tide_io.readbidstsv(
+        matchedfilelist[0],
+        colspec="cardiacfromfmri_25.0Hz,normpleth",
+    )
+    print(f"{inputarray.shape=}")
+    tclen = inputarray.shape[1]
+    LGR.info(f"tclen set to {tclen}")
+    return matchedfilelist, tclen
+
+
+def readindata(
+    matchedfilelist: list[str],
+    tclen: int,
+    targetfrag: str = "xyz",
+    inputfrag: str = "abc",
+    usebadpts: bool = False,
+    startskip: int = 0,
+    endskip: int = 0,
+    corrthresh: float = 0.5,
+    readlim: int | None = None,
+    readskip: int | None = None,
+    debug: bool = False,
+) -> (
+    tuple[np.ndarray, np.ndarray, list[str]] | tuple[np.ndarray, np.ndarray, list[str], np.ndarray]
+):
+    """
+    Read and process time-series data from a list of matched files.
+
+    This function reads cardiac and plethysmographic time-series data from a list of
+    files, performs quality checks, and returns the data in arrays suitable for
+    training or analysis. It supports filtering based on correlation thresholds,
+    NaN values, and signal standard deviations, and allows for optional skipping
+    of data at the start and end of each time series.
+
+    Parameters
+    ----------
+    matchedfilelist : list of str
+        List of file paths to be processed. Each file should contain time-series data
+        in a format compatible with `tide_io.readbidstsv`.
+    tclen : int
+        Length of the time series to be read from each file.
+    targetfrag : str, optional
+        Fragment identifier for target files, used in naming conversions. Default is "xyz".
+    inputfrag : str, optional
+        Fragment identifier for input files, used in naming conversions. Default is "abc".
+    usebadpts : bool, optional
+        If True, include a third array with bad point indicators. Default is False.
+    startskip : int, optional
+        Number of samples to skip at the beginning of each time series. Default is 0.
+    endskip : int, optional
+        Number of samples to skip at the end of each time series. Default is 0.
+    corrthresh : float, optional
+        Minimum correlation threshold between raw and plethysmographic signals.
+        Files with lower correlation are excluded. Default is 0.5.
+    readlim : int, optional
+        Maximum number of files to read. If None, all files are read. Default is None.
+    readskip : int, optional
+        Number of files to skip at the beginning of the file list. If None, no files are skipped. Default is None.
+    debug : bool, optional
+        If True, print debug information for each file. Default is False.
+
+    Returns
+    -------
+    tuple of (np.ndarray, np.ndarray, list[str]) or (np.ndarray, np.ndarray, list[str], np.ndarray)
+        - `x1`: Array of shape `(tclen, count)` containing x-time series data.
+        - `y1`: Array of shape `(tclen, count)` containing y-time series data.
+        - `names`: List of file names that passed quality checks.
+        - `bad1`: Optional array of shape `(tclen, count)` with bad point indicators if `usebadpts=True`.
+
+    Notes
+    -----
+    - Files with NaNs, short data, extreme standard deviations, or low correlation are excluded.
+    - The function logs information about excluded files for debugging and quality control.
+    - The `startskip` and `endskip` parameters are applied after filtering and before returning the data.
+
+    Examples
+    --------
+    >>> x, y, names = readindata(filelist, tclen=1000)
+    >>> x, y, names, bad = readindata(filelist, tclen=1000, usebadpts=True)
+    """
+    LGR.info(
+        "readindata called with usebadpts, startskip, endskip, readlim, readskip, targetfrag, inputfrag = "
+        f"{usebadpts} {startskip} {endskip} {readlim} {readskip} {targetfrag} {inputfrag}"
+    )
+    # allocate target arrays
+    LGR.info("allocating arrays")
+    s = len(matchedfilelist[readskip:])
+    if readlim is not None:
+        if s > readlim:
+            LGR.info(f"trimming read list to {readlim} from {s}")
+            s = readlim
+    x1 = np.zeros((tclen, s))
+    y1 = np.zeros((tclen, s))
+    names = []
+    if usebadpts:
+        bad1 = np.zeros((tclen, s))
+
+    # now read the data in
+    count = 0
+    LGR.info("checking data")
+    lowcorrfiles = []
+    nanfiles = []
+    shortfiles = []
+    strangemagfiles = []
+    for i in range(readskip, readskip + s):
+        lowcorrfound = False
+        nanfound = False
+        LGR.info(f"processing {matchedfilelist[i]}")
+
+        # read the info dict first
+        infodict = tide_io.readdictfromjson(
+            matchedfilelist[i].replace("_desc-stdrescardfromfmri_timeseries", "_info")
+        )
+        if infodict["corrcoeff_raw2pleth"] < corrthresh:
+            lowcorrfound = True
+            lowcorrfiles.append(matchedfilelist[i])
+        (
+            samplerate,
+            starttime,
+            columns,
+            inputarray,
+            compression,
+            columnsource,
+        ) = tide_io.readbidstsv(
+            matchedfilelist[i],
+            colspec="cardiacfromfmri_25.0Hz,normpleth",
+        )
+        tempy = inputarray[1, :]
+        tempx = inputarray[0, :]
+
+        if np.any(np.isnan(tempy)):
+            LGR.info(f"NaN found in file {matchedfilelist[i]} - discarding")
+            nanfound = True
+            nanfiles.append(matchedfilelist[i])
+        if np.any(np.isnan(tempx)):
+            nan_fname = targettoinput(
+                matchedfilelist[i], targetfrag=targetfrag, inputfrag=inputfrag
+            )
+            LGR.info(f"NaN found in file {nan_fname} - discarding")
+            nanfound = True
+            nanfiles.append(nan_fname)
+        strangefound = False
+        if not (0.5 < np.std(tempx) < 20.0):
+            strange_fname = matchedfilelist[i]
+            LGR.info(
+                f"file {strange_fname} has an extreme cardiacfromfmri standard deviation - discarding"
+            )
+            strangefound = True
+            strangemagfiles.append(strange_fname)
+        if not (0.5 < np.std(tempy) < 20.0):
+            LGR.info(
+                f"file {matchedfilelist[i]} has an extreme normpleth standard deviation - discarding"
+            )
+            strangefound = True
+            strangemagfiles.append(matchedfilelist[i])
+        shortfound = False
+        ntempx = tempx.shape[0]
+        ntempy = tempy.shape[0]
+        if ntempx < tclen:
+            short_fname = matchedfilelist[i]
+            LGR.info(f"file {short_fname} is short - discarding")
+            shortfound = True
+            shortfiles.append(short_fname)
+        if ntempy < tclen:
+            LGR.info(f"file {matchedfilelist[i]} is short - discarding")
+            shortfound = True
+            shortfiles.append(matchedfilelist[i])
+        if (
+            (ntempx >= tclen)
+            and (ntempy >= tclen)
+            and (not nanfound)
+            and (not shortfound)
+            and (not strangefound)
+            and (not lowcorrfound)
+        ):
+            x1[:tclen, count] = tempx[:tclen]
+            y1[:tclen, count] = tempy[:tclen]
+            names.append(matchedfilelist[i])
+            if debug:
+                print(f"{matchedfilelist[i]} included:")
+            if usebadpts:
+                bad1[:tclen, count] = inputarray[2, :]
+            count += 1
+        else:
+            print(f"{matchedfilelist[i]} excluded:")
+            if ntempx < tclen:
+                print("\tx data too short")
+            if ntempy < tclen:
+                print("\ty data too short")
+            print(f"\t{nanfound=}")
+            print(f"\t{shortfound=}")
+            print(f"\t{strangefound=}")
+            print(f"\t{lowcorrfound=}")
+    LGR.info(f"{count} runs pass file length check")
+    if len(lowcorrfiles) > 0:
+        LGR.info("files with low raw/pleth correlations:")
+        for thefile in lowcorrfiles:
+            LGR.info(f"\t{thefile}")
+    if len(nanfiles) > 0:
+        LGR.info("files with NaNs:")
+        for thefile in nanfiles:
+            LGR.info(f"\t{thefile}")
+    if len(shortfiles) > 0:
+        LGR.info("short files:")
+        for thefile in shortfiles:
+            LGR.info(f"\t{thefile}")
+    if len(strangemagfiles) > 0:
+        LGR.info("files with extreme standard deviations:")
+        for thefile in strangemagfiles:
+            LGR.info(f"\t{thefile}")
+
+    print(f"training set contains {count} runs of length {tclen}")
+    if usebadpts:
+        return (
+            x1[startskip:-endskip, :count],
+            y1[startskip:-endskip, :count],
+            names[:count],
+            bad1[startskip:-endskip, :count],
+        )
+    else:
+        return (
+            x1[startskip:-endskip, :count],
+            y1[startskip:-endskip, :count],
+            names[:count],
+        )
+
+
+def prep(
+    window_size: int,
+    step: int = 1,
+    excludethresh: float = 4.0,
+    usebadpts: bool = False,
+    startskip: int = 200,
+    endskip: int = 200,
+    excludebysubject: bool = True,
+    thesuffix: str = "sliceres",
+    thedatadir: str = "/data/frederic/physioconn/output_2025",
+    inputfrag: str = "abc",
+    targetfrag: str = "xyz",
+    corrthresh: float = 0.5,
+    dofft: bool = False,
+    readlim: int | None = None,
+    readskip: int | None = None,
+    countlim: int | None = None,
+    debug: bool = False,
+) -> (
+    tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray, int, int, int]
+    | tuple[np.ndarray, np.ndarray, np.ndarray, np.ndarray, int, int, int, np.ndarray, np.ndarray]
+):
+    """
+    Prepare time-series data for training and validation by reading, normalizing,
+    windowing, and splitting into batches.
+
+    This function reads physiological time-series data from JSON files, normalizes
+    the data, and organizes it into overlapping windows for model training and
+    validation. It supports filtering by subject or by window, and can optionally
+    apply FFT transformations to the data.
+
+    Parameters
+    ----------
+    window_size : int
+        Size of the sliding window used to segment time series data.
+    step : int, optional
+        Step size for sliding window (default is 1).
+    excludethresh : float, optional
+        Threshold for excluding data points based on maximum absolute value
+        (default is 4.0).
+    usebadpts : bool, optional
+        If True, includes bad points in the data processing (default is False).
+    startskip : int, optional
+        Number of time points to skip at the beginning of each time series
+        (default is 200).
+    endskip : int, optional
+        Number of time points to skip at the end of each time series
+        (default is 200).
+    excludebysubject : bool, optional
+        If True, exclude subjects with any region exceeding `excludethresh`;
+        otherwise, exclude windows (default is True).
+    thesuffix : str, optional
+        Suffix used in file search pattern (default is "sliceres").
+    thedatadir : str, optional
+        Directory path where the data files are stored (default is
+        "/data/frederic/physioconn/output_2025").
+    inputfrag : str, optional
+        Fragment identifier for input data (default is "abc").
+    targetfrag : str, optional
+        Fragment identifier for target data (default is "xyz").
+    corrthresh : float, optional
+        Correlation threshold for data filtering (default is 0.5).
+    dofft : bool, optional
+        If True, apply FFT transformation to the data (default is False).
+    readlim : int, optional
+        Limit on number of time points to read (default is None).
+    readskip : int, optional
+        Number of time points to skip when reading data (default is None).
+    countlim : int, optional
+        Maximum number of subjects to include (default is None).
+    debug : bool, optional
+        If True, enable debug logging (default is False).
+
+    Returns
+    -------
+    tuple of (np.ndarray, np.ndarray, np.ndarray, np.ndarray, int, int, int)
+        If `dofft` is False:
+            - train_x : Training input data (shape: [n_windows, window_size, 1])
+            - train_y : Training target data (shape: [n_windows, window_size, 1])
+            - val_x : Validation input data (shape: [n_windows, window_size, 1])
+            - val_y : Validation target data (shape: [n_windows, window_size, 1])
+            - N_subjs : Number of subjects
+            - tclen : Total time points after skipping
+            - batchsize : Number of windows per subject
+
+        tuple of (np.ndarray, np.ndarray, np.ndarray, np.ndarray, int, int, int, np.ndarray, np.ndarray)
+        If `dofft` is True:
+            - train_x : Training input data (shape: [n_windows, window_size, 2])
+            - train_y : Training target data (shape: [n_windows, window_size, 2])
+            - val_x : Validation input data (shape: [n_windows, window_size, 2])
+            - val_y : Validation target data (shape: [n_windows, window_size, 2])
+            - N_subjs : Number of subjects
+            - tclen : Total time points after skipping
+            - batchsize : Number of windows per subject
+            - Xscale_fourier : Fourier scaling for input data
+            - Yscale_fourier : Fourier scaling for target data
+
+    Notes
+    -----
+    - Data normalization is performed using median absolute deviation (MAD).
+    - The function supports both window-based and subject-based exclusion strategies.
+    - If `usebadpts` is True, bad points are included in the output arrays.
+    - FFT transformations are applied using a helper function `filtscale`.
+
+    Examples
+    --------
+    >>> train_x, train_y, val_x, val_y, N_subjs, tclen, batchsize = prep(
+    ...     window_size=100,
+    ...     step=10,
+    ...     excludethresh=3.0,
+    ...     excludebysubject=True,
+    ...     dofft=False
+    ... )
+    """
+    searchstring = os.path.join(thedatadir, "*", "*_desc-stdrescardfromfmri_timeseries.json")
+
+    # find matched files
+    matchedfilelist, tclen = getmatchedtcs(
+        searchstring,
+        usebadpts=usebadpts,
+        targetfrag=targetfrag,
+        inputfrag=inputfrag,
+        debug=debug,
+    )
+    # print("matchedfilelist", matchedfilelist)
+    print("tclen", tclen)
+
+    # read in the data from the matched files
+    print("about to read in data")
+    if usebadpts:
+        x, y, names, bad = readindata(
+            matchedfilelist,
+            tclen,
+            corrthresh=corrthresh,
+            targetfrag=targetfrag,
+            inputfrag=inputfrag,
+            usebadpts=True,
+            startskip=startskip,
+            endskip=endskip,
+            readlim=readlim,
+            readskip=readskip,
+        )
+    else:
+        x, y, names = readindata(
+            matchedfilelist,
+            tclen,
+            corrthresh=corrthresh,
+            targetfrag=targetfrag,
+            inputfrag=inputfrag,
+            startskip=startskip,
+            endskip=endskip,
+            readlim=readlim,
+            readskip=readskip,
+        )
+    print("finished reading in data")
+    LGR.info(f"xshape, yshape: {x.shape} {y.shape}")
+
+    # normalize input and output data
+    LGR.info("normalizing data")
+    LGR.info(f"count: {x.shape[1]}")
+    if LGR.getEffectiveLevel() <= logging.DEBUG:
+        # Only take these steps if the logger is set to DEBUG.
+        for thesubj in range(x.shape[1]):
+            LGR.debug(
+                f"prenorm sub {thesubj} min, max, mean, std, MAD x, y: "
+                f"{thesubj} "
+                f"{np.min(x[:, thesubj])} {np.max(x[:, thesubj])} {np.mean(x[:, thesubj])} "
+                f"{np.std(x[:, thesubj])} {mad(x[:, thesubj])} {np.min(y[:, thesubj])} "
+                f"{np.max(y[:, thesubj])} {np.mean(y[:, thesubj])} {np.std(x[:, thesubj])} "
+                f"{mad(y[:, thesubj])}"
+            )
+
+    y -= np.mean(y, axis=0)
+    themad = mad(y, axis=0)
+    for thesubj in range(themad.shape[0]):
+        if themad[thesubj] > 0.0:
+            y[:, thesubj] /= themad[thesubj]
+
+    x -= np.mean(x, axis=0)
+    themad = mad(x, axis=0)
+    for thesubj in range(themad.shape[0]):
+        if themad[thesubj] > 0.0:
+            x[:, thesubj] /= themad[thesubj]
+
+        if LGR.getEffectiveLevel() <= logging.DEBUG:
+            # Only take these steps if the logger is set to DEBUG.
+            for thesubj in range(x.shape[1]):
+                LGR.debug(
+                    f"postnorm sub {thesubj} min, max, mean, std, MAD x, y: "
+                    f"{thesubj} "
+                    f"{np.min(x[:, thesubj])} {np.max(x[:, thesubj])} {np.mean(x[:, thesubj])} "
+                    f"{np.std(x[:, thesubj])} {mad(x[:, thesubj])} {np.min(y[:, thesubj])} "
+                    f"{np.max(y[:, thesubj])} {np.mean(y[:, thesubj])} {np.std(x[:, thesubj])} "
+                    f"{mad(y[:, thesubj])}"
+                )
+
+    # now decide what to keep and what to exclude
+    thefabs = np.fabs(x)
+    if not excludebysubject:
+        N_pts = x.shape[0]
+        N_subjs = x.shape[1]
+        windowspersubject = np.int64((N_pts - window_size - 1) // step)
+        LGR.info(
+            f"{N_subjs} subjects with {N_pts} points will be evaluated with "
+            f"{windowspersubject} windows per subject with step {step}"
+        )
+        usewindow = np.zeros(N_subjs * windowspersubject, dtype=np.int64)
+        subjectstarts = np.zeros(N_subjs, dtype=np.int64)
+        # check each window
+        numgoodwindows = 0
+        LGR.info("checking windows")
+        subjectnames = []
+        for subj in range(N_subjs):
+            subjectstarts[subj] = numgoodwindows
+            subjectnames.append(names[subj])
+            LGR.info(f"{names[subj]} starts at {numgoodwindows}")
+            for windownumber in range(windowspersubject):
+                if (
+                    np.max(
+                        thefabs[
+                            step * windownumber : (step * windownumber + window_size),
+                            subj,
+                        ]
+                    )
+                    <= excludethresh
+                ):
+                    usewindow[subj * windowspersubject + windownumber] = 1
+                    numgoodwindows += 1
+        LGR.info(
+            f"found {numgoodwindows} out of a potential {N_subjs * windowspersubject} "
+            f"({100.0 * numgoodwindows / (N_subjs * windowspersubject)}%)"
+        )
+
+        for subj in range(N_subjs):
+            LGR.info(f"{names[subj]} starts at {subjectstarts[subj]}")
+
+        LGR.info("copying data into windows")
+        Xb = np.zeros((numgoodwindows, window_size, 1))
+        Yb = np.zeros((numgoodwindows, window_size, 1))
+        if usebadpts:
+            Xb_withbad = np.zeros((numgoodwindows, window_size, 1))
+        LGR.info(f"dimensions of Xb: {Xb.shape}")
+        thiswindow = 0
+        for subj in range(N_subjs):
+            for windownumber in range(windowspersubject):
+                if usewindow[subj * windowspersubject + windownumber] == 1:
+                    Xb[thiswindow, :, 0] = x[
+                        step * windownumber : (step * windownumber + window_size), subj
+                    ]
+                    Yb[thiswindow, :, 0] = y[
+                        step * windownumber : (step * windownumber + window_size), subj
+                    ]
+                    if usebadpts:
+                        Xb_withbad[thiswindow, :, 0] = bad[
+                            step * windownumber : (step * windownumber + window_size),
+                            subj,
+                        ]
+                    thiswindow += 1
+
+    else:
+        # now check for subjects that have regions that exceed the target
+        themax = np.max(thefabs, axis=0)
+
+        cleansubjs = np.where(themax < excludethresh)[0]
+
+        totalcount = x.shape[1] + 0
+        cleancount = len(cleansubjs)
+        if countlim is not None:
+            if cleancount > countlim:
+                LGR.info(f"reducing count to {countlim} from {cleancount}")
+                cleansubjs = cleansubjs[:countlim]
+
+        x = x[:, cleansubjs]
+        y = y[:, cleansubjs]
+        cleannames = []
+        for theindex in cleansubjs:
+            cleannames.append(names[theindex])
+        if usebadpts:
+            bad = bad[:, cleansubjs]
+        subjectnames = cleannames
+
+        LGR.info(f"after filtering, shape of x is {x.shape}")
+
+        N_pts = y.shape[0]
+        N_subjs = y.shape[1]
+
+        X = np.zeros((1, N_pts, N_subjs))
+        Y = np.zeros((1, N_pts, N_subjs))
+        if usebadpts:
+            BAD = np.zeros((1, N_pts, N_subjs))
+
+        X[0, :, :] = x
+        Y[0, :, :] = y
+        if usebadpts:
+            BAD[0, :, :] = bad
+
+        windowspersubject = int((N_pts - window_size - 1) // step)
+        LGR.info(
+            f"found {windowspersubject * cleancount} out of a potential "
+            f"{windowspersubject * totalcount} "
+            f"({100.0 * cleancount / totalcount}%)"
+        )
+        LGR.info(f"{windowspersubject} {cleancount} {totalcount}")
+
+        Xb = np.zeros((N_subjs * windowspersubject, window_size, 1))
+        LGR.info(f"dimensions of Xb: {Xb.shape}")
+        for j in range(N_subjs):
+            LGR.info(
+                f"sub {j} ({cleannames[j]}) min, max X, Y: "
+                f"{j} {np.min(X[0, :, j])} {np.max(X[0, :, j])} {np.min(Y[0, :, j])} "
+                f"{np.max(Y[0, :, j])}"
+            )
+            for i in range(windowspersubject):
+                Xb[j * windowspersubject + i, :, 0] = X[0, step * i : (step * i + window_size), j]
+
+        Yb = np.zeros((N_subjs * windowspersubject, window_size, 1))
+        LGR.info(f"dimensions of Yb: {Yb.shape}")
+        for j in range(N_subjs):
+            for i in range(windowspersubject):
+                Yb[j * windowspersubject + i, :, 0] = Y[0, step * i : (step * i + window_size), j]
+
+        if usebadpts:
+            Xb_withbad = np.zeros((N_subjs * windowspersubject, window_size, 2))
+            LGR.info(f"dimensions of Xb_withbad: {Xb_withbad.shape}")
+            for j in range(N_subjs):
+                LGR.info(f"packing data for subject {j}")
+                for i in range(windowspersubject):
+                    Xb_withbad[j * windowspersubject + i, :, 0] = X[
+                        0, step * i : (step * i + window_size), j
+                    ]
+                    Xb_withbad[j * windowspersubject + i, :, 1] = BAD[
+                        0, step * i : (step * i + window_size), j
+                    ]
+            Xb = Xb_withbad
+
+        subjectstarts = [i * windowspersubject for i in range(N_subjs)]
+        for subj in range(N_subjs):
+            LGR.info(f"{names[subj]} starts at {subjectstarts[subj]}")
+
+    LGR.info(f"Xb.shape: {Xb.shape}")
+    LGR.info(f"Yb.shape: {Yb.shape}")
+
+    if dofft:
+        Xb_fourier = np.zeros((N_subjs * windowspersubject, window_size, 2))
+        LGR.info(f"dimensions of Xb_fourier: {Xb_fourier.shape}")
+        Xscale_fourier = np.zeros((N_subjs, windowspersubject))
+        LGR.info(f"dimensions of Xscale_fourier: {Xscale_fourier.shape}")
+        Yb_fourier = np.zeros((N_subjs * windowspersubject, window_size, 2))
+        LGR.info(f"dimensions of Yb_fourier: {Yb_fourier.shape}")
+        Yscale_fourier = np.zeros((N_subjs, windowspersubject))
+        LGR.info(f"dimensions of Yscale_fourier: {Yscale_fourier.shape}")
+        for j in range(N_subjs):
+            LGR.info(f"transforming subject {j}")
+            for i in range((N_pts - window_size - 1)):
+                (
+                    Xb_fourier[j * windowspersubject + i, :, :],
+                    Xscale_fourier[j, i],
+                ) = filtscale(X[0, step * i : (step * i + window_size), j])
+                (
+                    Yb_fourier[j * windowspersubject + i, :, :],
+                    Yscale_fourier[j, i],
+                ) = filtscale(Y[0, step * i : (step * i + window_size), j])
+
+    limit = np.int64(0.8 * Xb.shape[0])
+    LGR.info(f"limit: {limit} out of {len(subjectstarts)}")
+    # find nearest subject start
+    firstvalsubject = np.abs(subjectstarts - limit).argmin()
+    LGR.info(f"firstvalsubject: {firstvalsubject}")
+    perm_train = np.random.permutation(np.int64(np.arange(subjectstarts[firstvalsubject])))
+    perm_val = np.random.permutation(
+        np.int64(np.arange(subjectstarts[firstvalsubject], Xb.shape[0]))
+    )
+
+    LGR.info("training subjects:")
+    for i in range(0, firstvalsubject):
+        LGR.info(f"\t{i} {subjectnames[i]}")
+    LGR.info("validation subjects:")
+    for i in range(firstvalsubject, len(subjectstarts)):
+        LGR.info(f"\t{i} {subjectnames[i]}")
+
+    perm = range(Xb.shape[0])
+
+    batchsize = windowspersubject
+
+    if dofft:
+        train_x = Xb_fourier[perm[:limit], :, :]
+        train_y = Yb_fourier[perm[:limit], :, :]
+
+        val_x = Xb_fourier[perm[limit:], :, :]
+        val_y = Yb_fourier[perm[limit:], :, :]
+        LGR.info(f"train, val dims: {train_x.shape} {train_y.shape} {val_x.shape} {val_y.shape}")
+        return (
+            train_x,
+            train_y,
+            val_x,
+            val_y,
+            N_subjs,
+            tclen - startskip - endskip,
+            batchsize,
+            Xscale_fourier,
+            Yscale_fourier,
+        )
+    else:
+        train_x = Xb[perm_train, :, :]
+        train_y = Yb[perm_train, :, :]
+
+        val_x = Xb[perm_val, :, :]
+        val_y = Yb[perm_val, :, :]
+
+        LGR.info(f"train, val dims: {train_x.shape} {train_y.shape} {val_x.shape} {val_y.shape}")
+        return (
+            train_x,
+            train_y,
+            val_x,
+            val_y,
+            N_subjs,
+            tclen - startskip - endskip,
+            batchsize,
+        )