Skip to content

ESNGenerator

The number of inputs (n_inputs) is always 1 and n_outputs is infered from passed data. It uses always feedback, so that is not a parameter anymore (always True).

Parameters:

Name Type Description Default
n_steps

int, default=100 Number of steps to generate pattern (used by predict method).

required
n_reservoir

int, optional, default=100 Number of reservoir neurons. Only used if W is not passed. If W is passed, n_reservoir gets overwritten with len(W). Either n_reservoir or W must be passed.

required
W

np.ndarray of shape (n_reservoir, n_reservoir), optional, default=None Reservoir weights matrix. If None, random weights are used (uniformly distributed around 0, ie., in [-0.5, 0.5). Be careful with the distribution of W values. Wrong W initialization might drastically affect test performance (even with reasonable good training fit). Spectral radius will be adjusted in all cases. Either n_reservoir or W must be passed.

required
spectral_radius

float, default=.99 Spectral radius of the reservoir weights matrix (W). Spectral radius will be adjusted in all cases (also with user specified W).

required
W_in

np.ndarray of shape (n_reservoir, 1+n_inputs) (1->bias), optional, default None. Input weights matrix by which input signal is multiplied. If None, random weights are used.

required
W_fb

np.ndarray of shape(n_reservoir, n_outputs), optional, default None. Feedback weights matrix by which feedback is multiplied in case of feedback.

required
sparsity

float, optional, default=0 Proportion of the reservoir matrix weights forced to be zero. Note that with default W (centered around 0), the actual sparsity will be slightly more than the specified. If W is passed, sparsity will be ignored.

required
noise

float, optional, default=0 Scaling factor of the (uniform) noise input added to neurons at each step. This is used for regularization purposes and should typically be very small, e.g. 0.0001 or 1e-5.

required
leak_rate

float, optional, default=1 Leaking rate applied to the neurons at each step. Default is 1, which is no leaking. 0 would be total leakeage.

required
bias

int, float or np.ndarray, optional, default=1 Value of the bias neuron, injected at each time to the reservoir neurons. If int or float, all neurons receive the same. If np.ndarray is must be of length n_reservoir.

required
activation

function (numba jitted), optional, default=tanh Non-linear activation function applied to the neurons at each step. For numba acceleration, it must be a jitted function. Basic activation functions as tanh, sigmoid, relu are already available in echoe.utils. Either use those or write a custom one decorated with numba njit.

required
activation_out

function, optional, default=identity Activation function applied to the outputs. In other words, it is assumed that targets = f(outputs). So the output produced must be transformed.

required
fit_only_states

bool,default=False If True, outgoing weights (W_out) are computed fitting only the reservoir states. Inputs and bias are still use to drive reservoir activity, but ignored for fitting W_out, both in the training and prediction phase.

required
regression_method

str, optional, default "pinv" (pseudoinverse). Method to solve the linear regression to find out outgoing weights. One of ["pinv", "ridge"]. If "ridge", ridge_* parameters will be used.

required
ridge_alpha

float, ndarray of shape (n_outputs,), default=None Regularization coefficient used for Ridge regression. Larger values specify stronger regularization. If an array is passed, penalties are assumed to be specific to the targets. Hence they must correspond in number. Default is None to make sure one deliberately sets this since it is a crucial parameter. See sklearn Ridge documentation for details.

required
ridge_fit_intercept

bool, optional, default=False If True, intercept is fit in Ridge regression. Default False. See sklearn Ridge documentation for details.

required
ridge_normalize

bool, default=False This parameter is ignored when fit_intercept is set to False. If True, the regressors X will be normalized before regression by subtracting the mean and dividing by the l2-norm. See sklearn Ridge documentation for details.

required
ridge_max_iter

int, default=None Maximum number of iterations for conjugate gradient solver. See sklearn Ridge documentation for details.

required
ridge_tol

float, default=1e-3 Precision of the solution. See sklearn Ridge documentation for details.

required
ridge_solver

str, optional, default="auto" Solver to use in the Ridge regression. One of ["auto", "svd", "cholesky", "lsqr", "sparse_cg", "sag", "saga"]. See sklearn Ridge documentation for details.

required
ridge_sample_weight

float or ndarray of shape (n_samples,), default=None Individual weights for each sample. If given a float, every sample will have the same weight. See sklearn Ridge documentation for details.

required
n_transient

int, optional, default=0 Number of activity initial steps removed (not considered for training) in order to avoid initial instabilities. Default is 0, but this is something one definitely might want to tweak.

required
random_state

int, RandomState instance, default=None The seed of the pseudo random number generator used to generate weight matrices, to generate noise inyected to reservoir neurons (regularization) and it is passed to the ridge solver in case regression_method=ridge. From sklearn: If int, random_state is the seed used by the random number generator; If RandomState instance, random_state is the random number generator; If None, the random number generator is the RandomState instance used by np.random.

required
store_states_train

bool, optional, default=False If True, time series series of reservoir neurons during training are stored in the object attribute states_train_.

required
store_states_pred

bool, optional, default=False If True, time series series of reservoir neurons during prediction are stored in the object attribute states_pred_.

required

Attributes:

- W_out_ : array of shape (n_outputs, n_inputs + n_reservoir + 1).
    Outgoing weights after fitting linear regression model to predict outputs.
- training_prediction_: array of shape (n_samples, n_outputs).
    Predicted output on training data.
- states_train_: array of shape (n_samples, n_reservoir), default False.
    If store_states_train is True, states matrix is stored for visualizing
    reservoir neurons activity during training.
- states_pred_: array of shape (n_samples, n_reservoir), default False.
    If store_states_pred is True, states matrix is stored for visualizing
    reservoir neurons activity during prediction (test).

fit(self, X=None, y=None)

Fit Echo State model, i.e., find outgoing weights matrix (W_out) for later prediction. Bias is appended automatically to the inputs.

Parameters:

Name Type Description Default
X

None, always ignored, API consistency It is ignored as only the target sequence matters (outputs). A sequence of zeros will be fed in - matching the len(outputs) as initial condition.

None
y

2D np.ndarray of shape (n_samples,) or (n_samples, n_outputs), default=None Target variable.

None

Returns:

Type Description
self

returns an instance of self.

Source code in echoes/esn/_generator.py
def fit(self, X=None, y=None) -> "ESNGenerator":
    """
    Fit Echo State model, i.e., find outgoing weights matrix (W_out) for later
    prediction.
    Bias is appended automatically to the inputs.

    Arguments:
        X: None, always ignored, API consistency
            It is ignored as only the target sequence matters (outputs).
            A sequence of zeros will be fed in - matching the len(outputs) as initial
            condition.
        y: 2D np.ndarray of shape (n_samples,) or (n_samples, n_outputs), default=None
            Target variable.

    Returns:
        self: returns an instance of self.
    """
    if X is not None:
        warnings.warn("X will be ignored – ESNGenerator only takes y for training")
    y = check_array(y, ensure_2d=False, dtype=np.float64)
    self._dtype_ = y.dtype
    if y.ndim == 1:
        y = y.reshape(-1, 1)
    outputs = y

    # Initialize matrices and random state
    self.random_state_ = check_random_state(self.random_state)
    # Pattern generation takes no input, thus hardcode for later
    # construction of matrices
    self.n_inputs_ = 1
    self.n_reservoir_ = len(self.W) if self.W is not None else self.n_reservoir
    self.n_outputs_ = outputs.shape[1]
    self.W_in_ = self._init_incoming_weights()
    self.W_ = self._init_reservoir_weights()
    self.W_fb_ = self._init_feedback_weights()

    check_model_params(self.__dict__)

    # Make inputs zero
    inputs = np.zeros(shape=(outputs.shape[0], self.n_inputs_), dtype=self._dtype_)

    check_consistent_length(inputs, outputs)  # sanity check

    # Initialize reservoir model
    self.reservoir_ = self._init_reservoir_neurons()

    states = self.reservoir_.harvest_states(inputs, outputs, initial_state=None)

    # Extend states matrix with inputs; i.e., make [h(t); x(t)]
    full_states = states if self.fit_only_states else np.hstack((states, inputs))

    # Solve for W_out using full states and outputs, excluding transient
    self.W_out_ = self._solve_W_out(
        full_states[self.n_transient :, :], outputs[self.n_transient :, :]
    )
    # Predict on training set (including the pass through the output nonlinearity)
    self.training_prediction_ = self.activation_out(full_states @ self.W_out_.T)

    # Keep last state for later
    self.last_state_ = states[-1, :]
    self.last_input_ = inputs[-1, :]
    self.last_output_ = outputs[-1, :]

    # Store reservoir activity
    if self.store_states_train:
        self.states_train_ = states
    return self

predict(self, X=None)

Last training state/input/output is used as initial test state/input/output and at each step the output of the network is reinjected as input for next prediction, thus no inputs are needed for prediction.

Parameters:

Name Type Description Default
X

None, always ignored, API consistency

None

Returns:

Type Description
outputs

2D np.ndarray of shape (n_steps, n_outputs) Predicted outputs.

Source code in echoes/esn/_generator.py
def predict(self, X=None) -> np.ndarray:
    """
    Last training state/input/output is used as initial test
    state/input/output and at each step the output of the network is reinjected
    as input for next prediction, thus no inputs are needed for prediction.

    Arguments:
        X: None, always ignored, API consistency

    Returns:
        outputs: 2D np.ndarray of shape (n_steps, n_outputs)
         Predicted outputs.
    """
    if X is not None:
        warnings.warn("X will be ignored – ESNGenerator takes no X for prediction")

    assert self.n_steps >= 1, "n_steps must be >= 1"
    n_steps = self.n_steps  # shorthand

    # Initialize predictions: begin with last state as first state
    inputs = np.zeros(shape=(n_steps, self.n_inputs_), dtype=self._dtype_)
    inputs = np.vstack([self.last_input_, inputs])
    states = np.vstack([
        self.last_state_, np.zeros((n_steps, self.n_reservoir_), dtype=self._dtype_)
    ])
    outputs = np.vstack([
        self.last_output_, np.zeros((n_steps, self.n_outputs_), dtype=self._dtype_)
    ])

    check_consistent_length(inputs, outputs)  # sanity check

    # Go through samples (steps) and predict for each of them
    for t in range(1, outputs.shape[0]):
        states[t, :] = self.reservoir_.update_state(
            state_t=states[t - 1, :], X_t=inputs[t, :], y_t=outputs[t - 1, :],
        )
        if self.fit_only_states:
            full_states = states[t, :]
        else:
            full_states = np.concatenate([states[t, :], inputs[t, :]])
        # Predict
        outputs[t, :] = self.W_out_ @ full_states

        # TODO: Update last_{input, states, outputs}_

    # Store reservoir activity
    if self.store_states_pred:
        self.states_pred_ = states[1:, :]  # discard first step (comes from fitting)

    # Apply output non-linearity
    outputs = self.activation_out(outputs)
    return outputs[1:, :]  # discard initial step (comes from training)

score(self, X=None, y=None, sample_weight=None)

Wrapper around sklearn r2_score with kwargs.

From sklearn: R^2 (coefficient of determination) regression score function. Best possible score is 1.0 and it can be negative (because the model can be arbitrarily worse). A constant model that always predicts the expected value of y, disregarding the input features, would get a R^2 score of 0.0.

Parameters:

Name Type Description Default
X

None Not used, present for API consistency. Generative ESN predicts purely based on its generative outputs.

None
y

2D np.ndarray of shape (n_samples, ) or (n_samples, n_outputs) Target sequence, true values of the outputs.

None
sample_weight

array-like of shape (n_samples,), default=None Sample weights.

None

Returns:

Type Description
score

float R2 score

Source code in echoes/esn/_generator.py
def score(self, X=None, y=None, sample_weight=None) -> float:
    """
    Wrapper around sklearn r2_score with kwargs.

    From sklearn:
      R^2 (coefficient of determination) regression score function.
      Best possible score is 1.0 and it can be negative (because the model can be
      arbitrarily worse).
      A constant model that always predicts the expected value of y,
      disregarding the input features, would get a R^2 score of 0.0.

    Arguments:
        X: None
            Not used, present for API consistency.
            Generative ESN predicts purely based on its generative outputs.
        y: 2D np.ndarray of shape (n_samples, ) or (n_samples, n_outputs)
            Target sequence, true values of the outputs.
        sample_weight: array-like of shape (n_samples,), default=None
            Sample weights.

    Returns:
        score: float
            R2 score
    """
    return r2_score(y, self.predict(), sample_weight=sample_weight)