Changelog

Changes to the combined snnlab Python distribution are recorded here. Package versions use MAJOR.MINOR.PATCH. Component and serialized-schema versions identify separate compatibility contracts.

[Unreleased]Copy Anchor Link

ChangedCopy Anchor Link

1. Removed the root release-helper scripts and separate migration/versioning documents. Release preparation is manual, with instructions in the README; CI and publishing read the package version directly.

[0.2.0] - 2026-10-04Copy Anchor Link

ChangedCopy Anchor Link

1. Removed Scira AI and Cursor from the documentation’s Open menu.

2. ExecutionSpec now defaults to the graph executor. Typed requests requiring legacy routing must explicitly set executor="legacy"; the CLI retains its existing legacy default.

3. ExecutionSpec now exposes one input_bindings sequence accepting DenseArrayBinding, EventStreamBinding, PoissonInputBinding and DatasetSnapshotBinding through the public InputBinding type alias. Removed the separate inputs, event_bindings, poisson_bindings and dataset_binding constructor arguments; callers must migrate to typed bindings. Existing input compatibility rules and serialized execution protocols are retained.

4. PoissonInputBinding.batch_size now defaults to 1. Required fields precede optional fields in its constructor; positional callers must migrate to the new order (input_id, steps_count, rates_hz, seed, batch_size=1, categorical=False) or use keyword arguments.

5. Replaced ExecutionSpec.recording and recording_fields with diagnostics: bool = True. Declared outputs always return; only explicitly exposed diagnostics return by default, and diagnostics=False disables them for simulation, inference and training. Renamed ExecutionResult.recordings to diagnostics, replaced the graph CLI --recording profile with --diagnostics / --no-diagnostics, and added Projection.conductance for explicit diagnostic exposure. Training regularizers and runtime continuation state remain independent of diagnostic retention.

6. The training example now learns both input-to-E and E-to-readout weights, with explicit fast-sigmoid surrogate gradients, separate learning rates and gradient clipping. Input weights remain constrained to be non-negative.

7. ExecutionSpec exposes epochs, batch_size, shuffle, updates, save_final_checkpoint and save_selected_checkpoint directly. Graph training rejects these settings inside options; callers must move them to constructor fields. The CLI adapter, examples and API reference use the direct fields. Inference-specific options remain in options.

8. The PyTorch Integration example now composes the bundled SNN with an external Linear(2, 8) → ReLU → Linear(8, 2) head. Both modules train together, share a saved state dictionary, and appear in the introductory diagram.

9. Consolidated input documentation: Network covers declarations, and ExecutionSpec covers binding types and compatibility rules. Removed the standalone Inputs reference from navigation; former routes redirect to ExecutionSpec.

10. lang.LIF now creates a supported current-based LIF specification (also available explicitly as CUBA_LIF), replacing the previously unsupported lif declaration. Projection weights and trace ports use synapse-specific units: uS/.conductance for conductance, nA/.current for current.

AddedCopy Anchor Link

1. ExecutionResult.numpy(batch=None) returns named output and diagnostic arrays through NumpyExecutionResult, with an execution-derived time_ms axis. Optional batch selection uses declared signal axes, supporting both time-series and reduced outputs. Arrays are independent copies; original tensors and gradients remain intact. Quickstart now uses this API for plotting.

2. Added a runnable spike-pattern training example and Training documentation page with an explicit E-to-readout connection and named w_out weights, per-epoch training/validation loss and accuracy curves, and saved artifacts for later inference. SpikeCount.parameters now includes its readout weight so training recipes can select it directly.

3. Graph train now returns baseline and completed-epoch loss, accuracy and component metrics in result.metrics["epochs"]. ExecutionSpec.validation accepts a ValidationSpec containing held-out bindings and targets for evaluation without optimizer updates. Training now demonstrates one call handling all epochs and one final checkpoint save.

4. Added a paired Inference example and documentation page alongside Training. It loads Training’s saved bundle and learned checkpoint, classifies fresh spike patterns without retraining, and saves predictions, checkpoint provenance, a network diagram and response plots.

5. Added a PyTorch Integration walkthrough and examples/pytorch/training.py. The example wraps Training’s saved graph in an ordinary nn.Module, trains fresh weights using PyTorch data loaders and an external optimization loop, enforces graph constraints, plots per-epoch metrics, and saves/reloads a standard PyTorch state dictionary for test inference.

6. Expanded API Reference with Lang, Sim and Viz subgroups covering authoring, components, parameters, training recipes, compilation, operations/readouts, execution results, PyTorch integration, diagrams and plotting. Existing API URLs redirect to the grouped pages.

7. Added CUBA_LIF, ExponentialCurrent and the nA unit, with configurable rest/reset/initial voltage, refractory steps, surrogate gradients, training and runtime continuation. Current and conductance families are checked for compatibility.

8. Added versioned named registrations in snnlab.extensions for neurons, synapses, initializers, constraints, operations, objectives, regularizers, optimizers, surrogates and dataset encoders. Bundles retain definition/config dependencies without embedding code. Custom tensor state supports save/load continuation, regression objectives retain real targets, and external PyTorch loops can call GraphExecutor.enforce_constraints().

9. Added a runnable Customisation example and documentation comparing standard current LIF and a registered adaptive neuron using matched weights, custom initialization, input/response/adaptation plots and an introductory diagram.

10. Added a standalone current-based LIF simulation tutorial and runnable example, with input, spikes, voltage and current plotted together; Customisation follows it with registered adaptive dynamics and a custom weight initializer.

FixedCopy Anchor Link

1. Zero-delay feedforward connections between populations now retain a valid spike history for execution and runtime continuation.
2. SignalLike.id is now read-only, matching immutable Signal objects and readout properties. Network outputs and cross-entropy objectives share the same protocol, eliminating incorrect editor type errors for valid signals. Examples also check optional checkpoint and time-axis results before use.
3. Documentation navigation, search and LLM exports now use the current content collections instead of a module-level snapshot. Content and sidebar metadata changes invalidate development routes so newly added pages appear without restarting the server.

[0.1.1] - 2026-10-04Copy Anchor Link

AddedCopy Anchor Link

1. Static Fumadocs documentation with KaTeX, search, GitHub Pages and Cloudflare hosting.
2. Complete source-derived API references for lang, sim and viz, with automated drift checks.
3. Six runnable general examples for bundles, simulation, input replay, training, checkpoint resume and retained-signal plotting.
4. A changelog, single-source package version and release preparation/check helper.
5. Automatic PyPI publishing on version-source changes to main, using Trusted Publishing and release tags after package checks pass.

ChangedCopy Anchor Link

1. The documentation opens at the site root; previous /docs/ paths redirect there.
2. Documentation builds use Astro static output with Fumadocs React islands in place of Next.js.
3. Package builds read the runtime version directly, and source archives exclude local documentation dependencies, caches and generated output.

[0.1.0] - 2026-10-03Copy Anchor Link

AddedCopy Anchor Link

1. Initial combined Python distribution extracted from Pinglab, exposing snnlab.lang, snnlab.sim and snnlab.viz.
2. Portable graph authoring and validated bundles, graph-native and legacy execution, surrogate-gradient training, retained artifacts and visualization utilities.
3. Portable tests and package checks, retaining existing numerical defaults, schemas and backend identifiers.

The historical baseline is Git tag v0.1.0. Changes above it remain unreleased until a subsequent version is prepared and tagged.