Backend redesign#63

Merged

renierts merged 59 commits into

devfrom

backend-redesign

Jul 20, 2026

Commits on Jul 17, 2026

Remove torch dependency and nn subpackage

renierts
and
claude
committed
Replace private/deprecated sklearn and scipy imports with public APIs

renierts
and
claude
committed
Migrate estimators to the scikit-learn 1.6+ developer API

renierts
and
claude
committed
Refactor metrics to use the public scikit-learn API

renierts
and
claude
committed
Update repository URLs to the PlasmaControl organization

renierts
and
claude
committed
Modernize packaging metadata for the scikit-learn 1.6+ upgrade

renierts
and
claude
committed
Modernize CI and test configuration

renierts
and
claude
committed
Modernize type hints and unify imports across the package

renierts
and
claude
committed
Clean up outdated coding practices

renierts
and
claude
committed
Fix non-reproducible sparse weight initialization

renierts
and
claude
committed
Follow scikit-learn conventions for constructor parameters

renierts
and
claude
committed

Commits on Jul 18, 2026

Make the package pass strict mypy and enforce it in CI

renierts
and
claude
committed
Modernize the test suite to match the package

renierts
and
claude
committed
Add backend redesign plan

renierts
and
claude
committed
Phase A0: declare torch dependency and add backend package scaffolding

renierts
and
claude
committed
Plan: note to drop the private-API disclaimer when pyrcn.nn goes public

renierts
and
claude
committed
Phase A1: add check_sequences normalizer (TDD)

renierts
and
claude
committed
Phase A2: torch reservoir module with NodeToNode parity (TDD)

renierts
and
claude
committed
Plan: document torch-module reuse strategy (D14)

renierts
and
claude
committed
Phase A2: rebuild the reservoir on nn.RNNCell (D14 reuse)

renierts
and
claude
committed
Phase A2: bidirectional reservoir with NodeToNode parity

renierts
and
claude
committed
Phase A2: torch-native reservoir weight init (TDD)

renierts
and
claude
committed
Phase A2: Euler + Hebbian reservoir variants (TDD)

renierts
and
claude
committed
Phase A2: torch InputFeatureMap (nn.Linear) + input init (TDD)

renierts
and
claude
committed
Fix InputToNode k_in: connect k_in inputs per hidden node

renierts
and
claude
committed
Phase A3: torch-native closed-form ridge readout (TDD)

renierts
and
claude
committed
Plan: mark Phase A0-A3 workstreams done

renierts
and
claude
committed
Phase A4: config-block -> torch-backend bridge (TDD)

renierts
and
claude
committed
Phase A4: wire ELM estimators onto the torch backend (fast path + fallback)

renierts
and
claude
committed

Commits on Jul 19, 2026

Phase A4: wire ESN estimators onto the torch backend (fast path + fallback)

renierts
and
claude
committed
Plan: mark A4 core (estimator torch integration) done

renierts
and
claude
committed
Phase A4b: ESN washout + predict initial_state / return_state

renierts
and
claude
committed
Plan: mark A4b (washout + state carry) done

renierts
and
claude
committed
Phase A5: drop the joblib/loky sequence-parallel path

renierts
and
claude
committed
Phase A5: parity harness + model-selection tests on the torch backend

renierts
and
claude
committed
Plan: mark A5 (parity + model-selection tests, drop joblib) done

renierts
and
claude
committed
Phase A6: finalize - remove dead merge operators, device test, deps

renierts
and
claude
committed
Plan: mark A6 (finalize) done; Phase A complete

renierts
and
claude
committed
Improve activation validation error message (#54)

renierts
and
claude
committed
P1: promote the torch backend to a public pyrcn.nn package

renierts
and
claude
committed
P1: document the public pyrcn.nn API (Sphinx page)

renierts
and
claude
committed
Plan: mark P1 (public pyrcn.nn) done

renierts
and
claude
committed
Phase B1 core: trainable LinearReadout + train_readout helper

renierts
and
claude
committed
Phase B1: gradient solver for the ELM estimators

renierts
and
claude
committed
Phase B1: gradient solver for the ESN estimators

renierts
and
claude
committed
Plan: mark Phase B1 (gradient readout solver) done

renierts
and
claude
committed

Commits on Jul 20, 2026

Phase B1: make gradient training reproducible via random_state

renierts
and
claude
committed
Plan: gradient training reproducible via random_state

renierts
and
claude
committed
Phase B1: tight gradient<->closed-form consistency demo (real reservoir)

renierts
and
claude
committed
Phase B2: trainable reservoir (end-to-end gradient training / BPTT)

renierts
and
claude
committed
Plan: mark Phase B2 (trainable reservoir) done

renierts
and
claude
committed
Update examples for the modernized API (no torch usage)

renierts
and
claude
committed
Examples: use matplotlib_inline for set_matplotlib_formats

renierts
and
claude
committed
Phase B2+: trainable_input flag + BPTT mini-batching (ESN)

renierts
and
claude
committed
Plan: mark B2+ (trainable_input + BPTT mini-batching) done

renierts
and
claude
committed
Docs: complete + modernize the Sphinx documentation

renierts
and
claude
committed
Unify activation set + extend gradient optimizers/losses

renierts
and
claude
committed
Plan: refresh roadmap status (through Phase B) + open tasks

renierts
and
claude
committed
Merge remote-tracking branch 'origin/dev' into backend-redesign
renierts
committed