Covariance-based Convolutional Neural Network

Solver ID: CovCNN

Usage

from invert import Solver

# fwd = ...    (mne.Forward object)
# evoked = ... (mne.Evoked object)

solver = Solver("CovCNN")
solver.make_inverse_operator(fwd)
stc = solver.apply_inverse_operator(evoked)
stc.plot()

Overview

Supervised CNN that operates on sensor covariance features (optionally with shrinkage) to predict source activity on a cortical grid.

References

1. Lukas Hecker 2025, unpublished

API Reference

Bases: BaseSolver

Class for the Covariance-based Convolutional Neural Network (CovCNN) for EEG inverse solutions.

Source code in invert/solvers/neural_networks/covcnn.py

class SolverCovCNN(BaseSolver):
    """Class for the Covariance-based Convolutional Neural Network (CovCNN) for EEG inverse solutions."""

    meta = SolverMeta(
        acronym="CovCNN",
        full_name="Covariance-based Convolutional Neural Network",
        category="Neural Networks",
        description=(
            "Supervised CNN that operates on sensor covariance features (optionally "
            "with shrinkage) to predict source activity on a cortical grid."
        ),
        references=["Lukas Hecker 2025, unpublished"],
    )

    def __init__(self, name="Cov-CNN", reduce_rank=False, use_shrinkage=True, **kwargs):
        self.name = name
        self.use_shrinkage = use_shrinkage
        self.Uk = None
        self.model = None
        self.optimizer = None
        self.device = None
        return super().__init__(reduce_rank=reduce_rank, **kwargs)

    def make_inverse_operator(
        self,
        forward,
        simulation_config,
        *args,
        parcellator=None,
        n_filters="auto",
        n_dense_units=300,
        n_dense_layers=2,
        activation_function="tanh",
        output_activation="sigmoid",
        epochs=300,
        learning_rate=1e-3,
        loss="cosine_similarity",
        size_validation_set=256,
        epsilon=0.0,
        patience=100,
        cov_type="basic",
        alpha="auto",
        **kwargs,
    ):
        """Calculate inverse operator.

        Parameters
        ----------
        forward : mne.Forward
            The mne-python Forward model instance.
        simulation_config : SimulationConfig
            A SimulationConfig object for data generation containing all
            simulation parameters (batch_size, n_sources, n_orders, snr_range, etc.).
        parcellator : Parcellator
            The Parcellator object. Default None.
        n_filters : int
            Number of filters in the convolution layer. Default "auto" sets to n_channels.
        n_dense_units : int
            Number of units in dense layers. Default 300.
        n_dense_layers : int
            Number of dense layers. Default 2.
        activation_function : str
            The activation function of the hidden layers. Default "tanh".
        output_activation : str
            The activation function of the output layer. Default "sigmoid".
        epochs : int
            The number of epochs to train. Default 300.
        learning_rate : float
            The learning rate of the optimizer. Default 1e-3.
        loss : str
            The loss function of the neural network. Default "cosine_similarity".
        size_validation_set : int
            The size of validation data set. Default 256.
        epsilon : float
            The threshold at which to select sources as "active". Default 0.
        patience : int
            Early stopping patience. Default 100.
        cov_type : str
            The type of covariance matrix to compute.
            "basic" -> Y @ Y.T
            "SSM" -> Signal Subspace Matching covariance
            "riemann" -> Riemannian covariance
            Default "basic".
        alpha : float
            The regularization parameter. Default "auto".

        Return
        ------
        self : object returns itself for convenience
        """
        super().make_inverse_operator(forward, *args, alpha=alpha, **kwargs)
        n_channels, n_dipoles = self.leadfield.shape

        if n_filters == "auto":
            n_filters = n_channels
        self.forward = forward

        # Store simulation config
        self.simulation_config = simulation_config

        # Store Parameters
        # Architecture
        self.parcellator = parcellator
        self.n_filters = n_filters
        self.activation_function = activation_function
        self.output_activation = output_activation
        self.n_dense_layers = n_dense_layers
        self.n_dense_units = n_dense_units
        # Training
        self.batch_size = self.simulation_config.batch_size
        self.epochs = epochs
        self.learning_rate = learning_rate
        self.loss = loss
        self.size_validation_set = size_validation_set
        self.patience = patience
        # Training Data (from simulation_config)
        self.n_timepoints = self.simulation_config.n_timepoints
        self.n_sources = self.simulation_config.n_sources
        self.n_orders = self.simulation_config.n_orders
        self.batch_repetitions = self.simulation_config.batch_repetitions
        self.snr_range = self.simulation_config.snr_range
        self.add_forward_error = self.simulation_config.add_forward_error
        self.forward_error = self.simulation_config.forward_error
        self.inter_source_correlation = self.simulation_config.inter_source_correlation
        self.cov_type = cov_type
        self.correlation_mode = self.simulation_config.correlation_mode
        self.noise_color_coeff = self.simulation_config.noise_color_coeff
        # Inference
        self.epsilon = epsilon
        logger.info("Create Generator:..")
        self.create_generator()
        logger.info("Build Model:..")
        self.build_model()
        logger.info("Train Model:..")
        self.train_model()

        self.inverse_operators = []
        return self

    def apply_inverse_operator(self, mne_obj, prior=None) -> mne.SourceEstimate:
        """Apply the inverse operator.

        Parameters
        ----------
        mne_obj : [mne.Evoked, mne.Epochs, mne.io.Raw]
            The MNE data object.
        prior : numpy.ndarray
            The prior gamma vector. Default None.
        Return
        ------
        stc : mne.SourceEstimate
            The mne Source Estimate object.
        """
        data = self.unpack_data_obj(mne_obj)

        source_mat = self.apply_model(data, prior=prior)
        stc = self.source_to_object(source_mat)

        return stc

    def compute_covariance(self, Y: np.ndarray) -> np.ndarray:
        """Compute the covariance matrix of the data.

        Parameters
        ----------
        Y : numpy.ndarray
            The data matrix.

        Return
        ------
        C : numpy.ndarray
            The covariance matrix.
        """
        if self.cov_type == "basic":
            C = Y @ Y.T
        elif self.cov_type == "log":
            C = self.log_covariance(Y)
        elif self.cov_type == "recipsiicos":
            C = self.recipsiicos_covariance(Y)
        elif self.cov_type == "riemann":
            C = self.riemannian_covariance(Y)
        elif self.cov_type == "SSM":
            n_time = Y.shape[1]
            M_Y = Y.T @ Y
            YY = M_Y + 0.001 * (50 / n_time) * np.trace(M_Y) * np.eye(n_time)
            P_Y = (Y @ np.linalg.inv(YY)) @ Y.T
            C = P_Y.T @ P_Y
            # print("yes its new")
        else:
            msg = f"Covariance type '{self.cov_type}' not recognized. Use 'basic', 'SSM' or 'riemann' or provide a custom covariance matrix."
            raise ValueError(msg)

        if self.use_shrinkage:
            C = self._shrinkage_covariance(Y)

        return C

    def _shrinkage_covariance(self, Y: np.ndarray) -> np.ndarray:
        """Compute shrinkage covariance using OAS estimator."""
        lw = OAS(assume_centered=False)
        C = lw.fit(Y.T).covariance_
        return C

    def log_covariance(
        self,
        Y,
        eps=1e-6,  # jitter & eigenvalue floor
        demean=True,  # remove channel means before covariance
    ):
        """
        Compute a single log–covariance matrix from EEG data Y (m x T).

        Steps:
        1) (Optional) Demean channels over time
        3) Shrinkage: C_shrunk = (1-alpha)*C + alpha*(tr(C)/m) * I  [or Ledoit–Wolf]
        4) (Optional) Whitening: Cw = W^{-1/2} C_shrunk W^{-1/2}
        5) Matrix log via eigendecomposition: C_log = U log(Λ) U^T

        Args:
            Y: np.ndarray, shape (m, T)
            noise_cov: np.ndarray (m, m) SPD baseline noise covariance for whitening; if None, no whitening
            eps: float, numerical floor for SPD safety and eigenvalues
            demean: bool, subtract per-channel temporal mean before covariance

        Returns:
            C_log: np.ndarray, shape (m, m) — the log–covariance matrix
        """
        Y = np.asarray(Y)
        m, T = Y.shape
        assert m <= T, "For stable covariance estimation, prefer T >= m."

        # 1) Demean over time (recommended)
        X = Y - Y.mean(axis=1, keepdims=True) if demean else Y

        C = X @ X.T / T

        # Symmetrize (guard against tiny FP asymmetries)
        C = 0.5 * (C + C.T)

        C += eps * np.eye(m)

        # 5) Matrix log via eigen-decomposition (log-Euclidean map)
        C_log = self.compute_log_cov(C, eps=eps)

        return C_log

    def compute_log_cov(self, C: np.ndarray, eps: float = 1e-6) -> np.ndarray:
        evals, evecs = np.linalg.eigh(C)
        evals = np.clip(evals, eps, None)
        log_evals = np.log(evals)
        C_log = (evecs * log_evals) @ evecs.T  # evecs @ diag(log_evals) @ evecs.T
        C_log = 0.5 * (C_log + C_log.T)  # re-symmetrize
        return C_log

    def recipsiicos_covariance(
        self, Y: np.ndarray, pwr_energy: float = 0.95
    ) -> np.ndarray:
        """Compute the ReciPSIICOS covariance matrix of the data."""
        m, n = self.leadfield.shape
        # Center Y over time
        Y = Y - Y.mean(axis=1, keepdims=True)

        if self.Uk is None:  # type: ignore[has-type]
            Qp = self._Q_power(self.leadfield)  # (m^2, n)

            U, s, _ = np.linalg.svd(Qp, full_matrices=False)
            k = self._svd_rank_from_energy(s, pwr_energy)
            self.Uk = U[:, :k]  # (m^2, k)

        v = np.kron(Y, Y)  # (m^2,)
        v_tilde = self.Uk @ (self.Uk.T @ v)  # project
        Ct = np.reshape(v_tilde, (m, m))  # (m, m)

        # Spectral flip
        w, V = np.linalg.eigh(Ct)
        Ct = (V * np.abs(w)) @ V.T

        # Log-Euclidean map
        Ct_log = self.compute_log_cov(Ct, eps=1e-6)
        return Ct_log

    @staticmethod
    def _svd_rank_from_energy(s: np.ndarray, keep_energy: float) -> int:
        s2 = s**2
        cum = np.cumsum(s2) / np.sum(s2) if np.sum(s2) > 0 else np.ones_like(s2)
        return int(np.searchsorted(cum, keep_energy) + 1)

    @staticmethod
    def _Q_power(Lr: np.ndarray) -> np.ndarray:
        """
        Build power subspace columns: q_i = vec(g_i g_i^T) = kron(g_i, g_i).
        Returns Q_pwr: (r^2, n)
        """
        r, n = Lr.shape
        Q = np.empty((r * r, n), dtype=Lr.dtype)
        for i in range(n):
            gi = Lr[:, i]
            Q[:, i] = np.kron(gi, gi)
        return Q

    def riemannian_covariance(self, X, eps=1e-12):
        """
        Compute the Riemannian (log-mapped) covariance of EEG data.

        Parameters
        ----------
        X : ndarray of shape (n_channels, n_times)
            EEG data matrix.
        eps : float
            Small value for numerical stability.

        Returns
        -------
        Z : ndarray of shape (n_channels, n_channels)
            Tangent-space (Riemannian) covariance matrix.
        """
        # 1) Empirical covariance
        C = (X @ X.T) / X.shape[1]
        C = 0.5 * (C + C.T)  # enforce symmetry

        # 2) Reference covariance (for simplicity use C itself here)
        C0 = C.copy()

        # 3) Compute inverse sqrt and sqrt of C0
        eigvals, eigvecs = np.linalg.eigh(C0)
        eigvals = np.clip(eigvals, eps, None)
        C0_sqrt = (eigvecs * np.sqrt(eigvals)) @ eigvecs.T
        C0_invsqrt = (eigvecs * (1.0 / np.sqrt(eigvals))) @ eigvecs.T

        # 4) Log map: Z = C0^{1/2} log( C0^{-1/2} C C0^{-1/2} ) C0^{1/2}
        A = C0_invsqrt @ C @ C0_invsqrt
        vals, vecs = np.linalg.eigh(0.5 * (A + A.T))
        vals = np.clip(vals, eps, None)
        logA = (vecs * np.log(vals)) @ vecs.T
        Z = C0_sqrt @ logA @ C0_sqrt
        Z = 0.5 * (Z + Z.T)

        return Z

    def apply_model(self, data: np.ndarray, prior=None) -> np.ndarray:
        """Compute the inverse solution of the M/EEG data.

        Parameters
        ----------
        data : numpy.ndarray
            The M/EEG data matrix.
        prior : numpy.ndarray
            The prior gamma vector. Default None.
        Return
        ------
        x_hat : numpy.ndarray
            The source esimate.

        """
        if self.parcellator is not None:
            leadfield = self.parcellator.forward_parcellated["sol"]["data"]
        else:
            leadfield = self.leadfield

        y = deepcopy(data)
        y = y - y.mean(axis=1, keepdims=True)  # Temporal mean centering

        n_channels, n_times = y.shape

        # Compute Data Covariance Matrix
        C = self.compute_covariance(y)
        # Scale
        C /= abs(C).max()
        # C /= y.shape[1]

        # Add empty batch and (color-) channel dimension
        C = C[np.newaxis, np.newaxis, :, :].astype(np.float32, copy=False)

        # Get prior source covariance from model
        assert self.model is not None
        self.model.eval()
        device = self.device or get_torch_device()
        with torch.no_grad():
            gammas = (
                self.model(torch.as_tensor(C, dtype=torch.float32, device=device))
                .detach()
                .cpu()
                .numpy()[0]
            )
        # print(f"Gammas shape: {gammas.shape}")
        # gammas = np.maximum(gammas, 0)
        max_gamma = float(np.max(gammas))
        if max_gamma > 0:
            gammas /= max_gamma
        # gammas = gammas**2
        gammas = gammas ** (1)
        # gammas[gammas<self.epsilon] = 0

        if prior is not None:
            prior_max = float(np.max(prior))
            if prior_max > 0:
                gammas = gammas * (prior / prior_max)

        self.gammas = deepcopy(gammas)

        source_covariance = np.diag(gammas)

        # Minimum norm
        Sigma_y = leadfield @ source_covariance @ leadfield.T
        if self.alpha == "auto":
            r_grid = np.asarray(self.r_values, dtype=float)
        else:
            r_grid = np.asarray([float(self.alpha)], dtype=float)
        self.alphas = list(r_grid)

        inverse_operators = []
        for r in r_grid:
            reg_term = float(r) * np.trace(Sigma_y) / n_channels
            inverse_operator = (
                source_covariance
                @ leadfield.T
                @ np.linalg.inv(Sigma_y + reg_term * np.eye(n_channels))
            )
            inverse_operators.append(inverse_operator)

        self.inverse_operators = [
            InverseOperator(inverse_operator, self.name)
            for inverse_operator in inverse_operators
        ]

        if self.parcellator is not None:
            leadfield_parcellated = self.parcellator.forward_parcellated["sol"]["data"]
            leadfield_original = self.leadfield
            self.leadfield = leadfield_parcellated
            x_hat, _ = self.regularise_gcv(y)
            self.leadfield = leadfield_original
            return self.parcellator.decompress_data(x_hat)
        else:
            x_hat, _ = self.regularise_gcv(y)

        return x_hat

    def train_model(
        self,
    ):
        """Train the neural network model."""
        import time

        if self.model is None:
            raise RuntimeError("Model not initialized. Call build_model() first.")
        if self.optimizer is None:
            raise RuntimeError("Optimizer not initialized. Call build_model() first.")

        start_t = time.perf_counter()

        self.model.train()
        loss_fn = loss_from_name(self.loss)

        # Re-initialize generator (matches previous TF behaviour).
        self.create_generator()

        # Get validation data from generator (consume one repeated batch).
        for _ in range(self.batch_repetitions):
            x_val, y_val = next(self.generator)

        device = self.device or get_torch_device()
        x_val_t = torch.as_tensor(x_val, dtype=torch.float32, device=device)
        y_val_t = torch.as_tensor(y_val, dtype=torch.float32, device=device)

        history: dict[str, list[float]] = {"loss": [], "val_loss": []}
        best_val = float("inf")
        best_state = None
        patience_left = int(self.patience)

        total_epochs = int(self.epochs)
        log_every = 10
        logger.info(
            "Training %s on simulated data (epochs=%d, batch_size=%d, batch_repetitions=%d, lr=%s, loss=%s, patience=%d, device=%s)",
            self.__class__.__name__,
            total_epochs,
            int(getattr(self, "batch_size", -1)),
            int(getattr(self, "batch_repetitions", 1)),
            str(getattr(self, "learning_rate", "?")),
            str(getattr(self, "loss", "?")),
            int(getattr(self, "patience", 0)),
            str(device),
        )

        epochs_ran = 0
        for epoch in range(total_epochs):
            self.model.train()
            running = 0.0
            for _step in range(int(self.batch_repetitions)):
                x_batch, y_batch = next(self.generator)
                x_t = torch.as_tensor(x_batch, dtype=torch.float32, device=device)
                y_t = torch.as_tensor(y_batch, dtype=torch.float32, device=device)

                self.optimizer.zero_grad(set_to_none=True)
                y_pred = self.model(x_t)
                loss = loss_fn(y_pred, y_t)
                loss.backward()
                self.optimizer.step()
                running += float(loss.detach().cpu().item())

            train_loss = running / float(self.batch_repetitions)

            self.model.eval()
            with torch.no_grad():
                val_loss = float(loss_fn(self.model(x_val_t), y_val_t).cpu().item())

            history["loss"].append(train_loss)
            history["val_loss"].append(val_loss)
            epochs_ran = epoch + 1

            if val_loss < best_val:
                best_val = val_loss
                best_state = deepcopy(self.model.state_dict())
                patience_left = int(self.patience)
                logger.info(
                    "Epoch %d/%d - loss=%.6f val_loss=%.6f (new best)",
                    epoch + 1,
                    total_epochs,
                    train_loss,
                    val_loss,
                )
            else:
                patience_left -= 1
                if (epoch == 0) or ((epoch + 1) % log_every == 0):
                    logger.info(
                        "Epoch %d/%d - loss=%.6f val_loss=%.6f (patience_left=%d)",
                        epoch + 1,
                        total_epochs,
                        train_loss,
                        val_loss,
                        patience_left,
                    )
                if patience_left <= 0:
                    logger.info(
                        "Early stopping at epoch %d/%d (best_val=%.6f)",
                        epoch + 1,
                        total_epochs,
                        best_val,
                    )
                    break

        if best_state is not None:
            self.model.load_state_dict(best_state)
        self.model.eval()
        self.history = history
        elapsed_s = time.perf_counter() - start_t
        logger.info(
            "Training finished (epochs_ran=%d, best_val=%.6f, elapsed=%.2fs)",
            epochs_ran,
            best_val,
            elapsed_s,
        )

    def build_model(
        self,
    ):
        """Build the neural network model."""
        if self.parcellator is None:
            leadfield = self.leadfield
            n_channels, n_dipoles = leadfield.shape
        else:
            leadfield = self.parcellator.forward_parcellated["sol"]["data"]
            n_channels, n_dipoles = leadfield.shape

        self.device = get_torch_device()
        model = _CovCNNNet(
            leadfield,
            n_dense_layers=int(self.n_dense_layers),
            n_dense_units=int(self.n_dense_units),
            activation_function=str(self.activation_function),
            output_activation=str(self.output_activation),
        ).to(self.device)

        self.model = model
        self.optimizer = torch.optim.Adam(
            self.model.parameters(), lr=float(self.learning_rate)
        )

        logger.info(
            "Total number of trainable parameters: %d",
            count_trainable_parameters(self.model),
        )

    def create_generator(
        self,
    ):
        """Create the data generator used for the simulations."""

        # Create SimulationGenerator using the config
        sim_gen = SimulationGenerator(self.forward, config=self.simulation_config)

        # Wrap the generator to compute covariance and return (x_cov, y_mask)
        def wrapped_generator():
            for x, y, _info in sim_gen.generate():
                x_cov = np.stack([self.compute_covariance(xx) for xx in x], axis=0)
                max_abs = np.abs(x_cov).max(axis=(1, 2), keepdims=True)
                max_abs = np.where(max_abs == 0, 1.0, max_abs)
                x_cov = (x_cov / max_abs).astype(np.float32, copy=False)

                # Torch expects NCHW: (batch, 1, n_channels, n_channels)
                x_cov = x_cov[:, np.newaxis, :, :]

                # Use continuous values as target
                y_abs_mean = np.abs(y).mean(axis=2)
                y_scale = y_abs_mean.max(axis=1, keepdims=True)
                y_scale = np.where(y_scale == 0, 1.0, y_scale)
                y_continuous = (y_abs_mean / y_scale).astype(np.float32, copy=False)

                if self.parcellator is not None:
                    y_continuous = self.parcellator.compress_data(y_continuous.T).T

                yield x_cov, y_continuous

        self.generator = wrapped_generator()

__init__

__init__(
    name="Cov-CNN",
    reduce_rank=False,
    use_shrinkage=True,
    **kwargs,
)

Source code in invert/solvers/neural_networks/covcnn.py

def __init__(self, name="Cov-CNN", reduce_rank=False, use_shrinkage=True, **kwargs):
    self.name = name
    self.use_shrinkage = use_shrinkage
    self.Uk = None
    self.model = None
    self.optimizer = None
    self.device = None
    return super().__init__(reduce_rank=reduce_rank, **kwargs)

make_inverse_operator

make_inverse_operator(
    forward,
    simulation_config,
    *args,
    parcellator=None,
    n_filters="auto",
    n_dense_units=300,
    n_dense_layers=2,
    activation_function="tanh",
    output_activation="sigmoid",
    epochs=300,
    learning_rate=0.001,
    loss="cosine_similarity",
    size_validation_set=256,
    epsilon=0.0,
    patience=100,
    cov_type="basic",
    alpha="auto",
    **kwargs,
)

Calculate inverse operator.

Parameters:

Name	Type	Description	Default
forward	Forward	The mne-python Forward model instance.	required
simulation_config	SimulationConfig	A SimulationConfig object for data generation containing all simulation parameters (batch_size, n_sources, n_orders, snr_range, etc.).	required
parcellator	Parcellator	The Parcellator object. Default None.	None
n_filters	int	Number of filters in the convolution layer. Default "auto" sets to n_channels.	'auto'
n_dense_units	int	Number of units in dense layers. Default 300.	300
n_dense_layers	int	Number of dense layers. Default 2.	2
activation_function	str	The activation function of the hidden layers. Default "tanh".	'tanh'
output_activation	str	The activation function of the output layer. Default "sigmoid".	'sigmoid'
epochs	int	The number of epochs to train. Default 300.	300
learning_rate	float	The learning rate of the optimizer. Default 1e-3.	0.001
loss	str	The loss function of the neural network. Default "cosine_similarity".	'cosine_similarity'
size_validation_set	int	The size of validation data set. Default 256.	256
epsilon	float	The threshold at which to select sources as "active". Default 0.	0.0
patience	int	Early stopping patience. Default 100.	100
cov_type	str	The type of covariance matrix to compute. "basic" -> Y @ Y.T "SSM" -> Signal Subspace Matching covariance "riemann" -> Riemannian covariance Default "basic".	'basic'
alpha	float	The regularization parameter. Default "auto".	'auto'

Return

self : object returns itself for convenience

Source code in invert/solvers/neural_networks/covcnn.py

def make_inverse_operator(
    self,
    forward,
    simulation_config,
    *args,
    parcellator=None,
    n_filters="auto",
    n_dense_units=300,
    n_dense_layers=2,
    activation_function="tanh",
    output_activation="sigmoid",
    epochs=300,
    learning_rate=1e-3,
    loss="cosine_similarity",
    size_validation_set=256,
    epsilon=0.0,
    patience=100,
    cov_type="basic",
    alpha="auto",
    **kwargs,
):
    """Calculate inverse operator.

    Parameters
    ----------
    forward : mne.Forward
        The mne-python Forward model instance.
    simulation_config : SimulationConfig
        A SimulationConfig object for data generation containing all
        simulation parameters (batch_size, n_sources, n_orders, snr_range, etc.).
    parcellator : Parcellator
        The Parcellator object. Default None.
    n_filters : int
        Number of filters in the convolution layer. Default "auto" sets to n_channels.
    n_dense_units : int
        Number of units in dense layers. Default 300.
    n_dense_layers : int
        Number of dense layers. Default 2.
    activation_function : str
        The activation function of the hidden layers. Default "tanh".
    output_activation : str
        The activation function of the output layer. Default "sigmoid".
    epochs : int
        The number of epochs to train. Default 300.
    learning_rate : float
        The learning rate of the optimizer. Default 1e-3.
    loss : str
        The loss function of the neural network. Default "cosine_similarity".
    size_validation_set : int
        The size of validation data set. Default 256.
    epsilon : float
        The threshold at which to select sources as "active". Default 0.
    patience : int
        Early stopping patience. Default 100.
    cov_type : str
        The type of covariance matrix to compute.
        "basic" -> Y @ Y.T
        "SSM" -> Signal Subspace Matching covariance
        "riemann" -> Riemannian covariance
        Default "basic".
    alpha : float
        The regularization parameter. Default "auto".

    Return
    ------
    self : object returns itself for convenience
    """
    super().make_inverse_operator(forward, *args, alpha=alpha, **kwargs)
    n_channels, n_dipoles = self.leadfield.shape

    if n_filters == "auto":
        n_filters = n_channels
    self.forward = forward

    # Store simulation config
    self.simulation_config = simulation_config

    # Store Parameters
    # Architecture
    self.parcellator = parcellator
    self.n_filters = n_filters
    self.activation_function = activation_function
    self.output_activation = output_activation
    self.n_dense_layers = n_dense_layers
    self.n_dense_units = n_dense_units
    # Training
    self.batch_size = self.simulation_config.batch_size
    self.epochs = epochs
    self.learning_rate = learning_rate
    self.loss = loss
    self.size_validation_set = size_validation_set
    self.patience = patience
    # Training Data (from simulation_config)
    self.n_timepoints = self.simulation_config.n_timepoints
    self.n_sources = self.simulation_config.n_sources
    self.n_orders = self.simulation_config.n_orders
    self.batch_repetitions = self.simulation_config.batch_repetitions
    self.snr_range = self.simulation_config.snr_range
    self.add_forward_error = self.simulation_config.add_forward_error
    self.forward_error = self.simulation_config.forward_error
    self.inter_source_correlation = self.simulation_config.inter_source_correlation
    self.cov_type = cov_type
    self.correlation_mode = self.simulation_config.correlation_mode
    self.noise_color_coeff = self.simulation_config.noise_color_coeff
    # Inference
    self.epsilon = epsilon
    logger.info("Create Generator:..")
    self.create_generator()
    logger.info("Build Model:..")
    self.build_model()
    logger.info("Train Model:..")
    self.train_model()

    self.inverse_operators = []
    return self

apply_inverse_operator

apply_inverse_operator(
    mne_obj, prior=None
) -> mne.SourceEstimate

Apply the inverse operator.

Parameters:

Name	Type	Description	Default
mne_obj	[Evoked, Epochs, Raw]	The MNE data object.	required
prior	ndarray	The prior gamma vector. Default None.	None

Return

stc : mne.SourceEstimate The mne Source Estimate object.

Source code in invert/solvers/neural_networks/covcnn.py

def apply_inverse_operator(self, mne_obj, prior=None) -> mne.SourceEstimate:
    """Apply the inverse operator.

    Parameters
    ----------
    mne_obj : [mne.Evoked, mne.Epochs, mne.io.Raw]
        The MNE data object.
    prior : numpy.ndarray
        The prior gamma vector. Default None.
    Return
    ------
    stc : mne.SourceEstimate
        The mne Source Estimate object.
    """
    data = self.unpack_data_obj(mne_obj)

    source_mat = self.apply_model(data, prior=prior)
    stc = self.source_to_object(source_mat)

    return stc