Build · Architecture

classic_esn¶

The standard Echo State Network baseline. In this implementation the readout sees the raw input concatenated with the reservoir states, not the states alone.

Wiring¶

Input → Reservoir → Concatenate(Input, States) → Readout

The feedback signal drives a tanh reservoir, then rejoins its own states at the concatenation, so the readout (a CGReadoutLayer) solves ridge regression over feedback_size + reservoir_size features. The linear part of the input–output map reaches the readout unfiltered; the reservoir only has to supply what linearity cannot.

wiring — Model graph: the input feeds both the reservoir and the concatenation.

Use¶

import torch
from resdag.models import classic_esn
from resdag.training import ESNTrainer

series = torch.cumsum(0.1 * torch.randn(1, 1201, 3), dim=1)

model = classic_esn(reservoir_size=300, feedback_size=3, output_size=3)
ESNTrainer(model).fit(
    warmup_inputs=(series[:, :200],),
    train_inputs=(series[:, 200:1200],),
    targets={"output": series[:, 201:1201]},              # next-step targets
)
preds = model.forecast(series[:, :200], horizon=100)      # (1, 100, 3)

Parameters¶

Parameter	Default	Notes
`reservoir_size`, `feedback_size`, `output_size`	required	units, input dim, output dim
`topology`, `feedback_initializer`	`None`	any initialization spec
`spectral_radius`	`0.9`	the factory scales; the bare `ESNLayer` defaults to `None`
`leak_rate`	`1.0`	`1.0` = no leak
`activation`	`"tanh"`	also `"relu"`, `"sigmoid"`, `"identity"`
`bias`, `trainable`	`True`, `False`	random bias on; frozen reservoir
`readout_alpha`, `readout_bias`, `readout_name`	`1e-6`, `True`, `"output"`	ridge strength; `readout_name` keys the targets dict
`**reservoir_kwargs`	—	forwarded to `ESNLayer` (e.g. `bias_scaling`)

Choosing a baseline

Use classic_esn as the baseline against which augmented variants are compared. For chaotic attractor reconstruction, start from ott_esn instead.

Reference¶

H. Jaeger, The "echo state" approach to analysing and training recurrent neural networks, GMD Report 148, German National Research Center for Information Technology (2001).