ResDAG

Reservoir computing for PyTorch

ResDAG treats reservoir models as ordinary PyTorch layers. Compose them into arbitrary DAGs, fit readouts with a single algebraic solve, and run training and forecasting on the GPU.

First forecast — 5 min Composition handbook API

system ESN forecast

— — a 500-unit reservoir forecast, drawn live; a different system on each visit. Systems from TSDynamics

Ten lines to a trained forecaster

One teacher-forced pass collects reservoir states; one conjugate-gradient ridge solve fits the readout. A 500-unit forecaster trains in well under a second, which keeps wide hyperparameter searches cheap.

import resdag as rd

data = rd.utils.load_file("lorenz.npy")
splits = rd.utils.prepare_esn_data(
    data, warmup_steps=200, train_steps=5000, val_steps=2000)
warmup, train, target, f_warmup, val = splits

model = rd.models.ott_esn(
    reservoir_size=500, feedback_size=3, output_size=3)
rd.ESNTrainer(model).fit(
    (warmup,), (train,), targets={"output": target})

prediction = model.forecast(f_warmup, horizon=2000)

No. 1

Layers, not frameworks

Reservoirs, readouts and transforms are ordinary PyTorch modules composed with a functional API: parallel reservoirs, multiple heads, state augmentation — any DAG you can define.

No. 2

Structure is a function

Connectivity and input structure are pluggable functions: a registry of graph topologies and initializers, or any callable that builds a matrix.

No. 3

Honest mathematics

Every update equation, solver decision and timing convention is documented index by index, so the implementation can be validated against your own derivations.

Choose a track

Start Installation, a first trained forecaster, and the concepts the library is built on. Build Layers, readouts, premade architectures, and the initialization system. Work Training paths, forecasting with exogenous drivers, hyperparameter studies, GPU use. Theory The update equations, the ridge solvers, and the timing conventions, stated against the code. Reference The full public API, generated from the source docstrings. Project Citation, ecosystem, related work, contributing, and the changelog.