EchoStateNetwork#

class torchdyno.models.esn.esn.EchoStateNetwork[source]#

Bases: Module

Methods

__init__()

Initialize internal Module state, shared by both nn.Module and ScriptModule.

apply_reservoirs(x, initial_state)

Apply the reservoirs to the input.

evaluate(test_loader, score_fn[, washout])

Evaluate the model.

fit_from_matrices([l2])

Fit the readout layer from the stored matrices.

fit_readout(train_loader[, l2_value, ...])

Fit the readout layer.

forward(x[, initial_state, return_states])

Forward pass of the network.

Attributes

hidden_size

Returns the size of the hidden state.
__init__(input_size: int, output_size: int, layer_sizes: List[int], arch_type: Literal['stacked', 'multi', 'parallel'] = 'stacked', activation: str = 'tanh', leakage: float = 1.0, input_scaling: float = 0.9, rho: float = 0.99, bias: bool = False, kernel_initializer: str | Callable[[Size], Tensor] = 'uniform', recurrent_initializer: str | Callable[[Size], Tensor] = 'uniform', net_gain_and_bias: bool = False)[source]#
__init__(output_size: int, combine_reservoirs: Literal['stacked', 'multi', 'parallel', 'merge'] = 'stacked', joint_scaling: float = 1.0, independent_inputs: bool = False)

Initialize internal Module state, shared by both nn.Module and ScriptModule.

forward(x, initial_state=None, return_states=False)[source]#

Forward pass of the network.

Parameters:

  • x (torch.Tensor) – input tensor.

  • initial_state (List[torch.Tensor]) – initial state of the reservoirs.

  • return_states (bool, optional) – whether to return the states. Defaults to False.

Returns:

output tensor.

Return type:

torch.Tensor

apply_reservoirs(x, initial_state)[source]#

Apply the reservoirs to the input.

Parameters:

  • x (torch.Tensor) – input tensor.

  • initial_state (Optional[torch.Tensor]) – initial state of the reservoirs.

Returns:

output tensor.

Return type:

torch.Tensor

fit_readout(train_loader, l2_value=1e-09, washout=0, score_fn=None, mode=None, eval_on=None, store_matrices=False)[source]#

Fit the readout layer.

Parameters:

  • train_loader (DataLoader) – training data loader.

  • eval_loader (DataLoader) – evaluation data loader.

  • l2_values (List[float]) – list of L2 regularization values.

  • score_fn (Callable[[Tensor, Tensor], float]) – scoring function.

  • mode (Literal["min", "max"]) – optimization mode.

  • weights (Optional[List[float]], optional) – weights for the loss. Defaults to None.

  • preprocess_fn (Optional[Callable], optional) – preprocessing function. Defaults to None.

  • device (Optional[str], optional) – device to use. Defaults to None.

  • l2_value (float | List[float])

  • washout (int)

  • eval_on (Literal['train'] | ~torch.utils.data.dataloader.DataLoader | None)

  • store_matrices (bool)

fit_from_matrices(l2=None)[source]#

Fit the readout layer from the stored matrices.

Parameters:

l2 (float | None)

evaluate(test_loader, score_fn, washout=0)[source]#

Evaluate the model.

Parameters:

  • test_loader (DataLoader) – test data loader.

  • score_fn (Callable[[Tensor, Tensor], float]) – scoring function.

  • washout (int, optional) – washout steps. Defaults to 0.

Returns:

evaluation score.

Return type:

float

property hidden_size: int#

Returns the size of the hidden state.