Skip to content

Version 2.8.1

Latest
Latest

Choose a tag to compare

View all tags
@chaoming0625 chaoming0625 released this 09 Jul 05:33
V2.8.1
6abc7ad
This commit was created on GitHub.com and signed with GitHub’s verified signature.
GPG key ID: B5690EEEBB952194
Verified
Learn about vigilant mode.

Release Date: July 9, 2026

Version 2.8.1 is a patch release that continues the 2.8.0 correctness focus. Its centerpiece is a second audited, library-wide bug-fix sweep (#868) — 18 confirmed defects across dnn, the integrators, encoders, initializers, connectivity, optimizers, dynold synaptic plasticity, the offline/online training algorithms, the object-transform layer, and the simulation runners — each backed by a co-located regression test. The release also repairs adaptive pooling on up-sampling, fixes a PoissonGroup spiking regression, and moves version metadata into a dedicated brainpy._version module.

Bug Fixes

Audited correctness sweep (#868)

Normalization and encoding

  • LayerNorm: normalize over the trailing normalized_shape dimensions instead of the leading batch/time axes.
  • WeightedPhaseEncoder: use a float power base so phase weights are non-zero (the encoder previously never spiked).

Integrators

  • Stratonovich Euler–Heun SDE: scale the predictor's Wiener increment by sqrt(dt), making it consistent with the corrector.

Initializers

  • TruncatedNormal: default to a finite 2-sigma truncation and accept None bounds as ±inf.
  • DOGDecay: produce a symmetric kernel on non-square grids and honor the normalize flag.
  • _compute_fans: handle 0-D/1-D shapes, fixing VarianceScaling initializers on bias vectors.

Connectivity

  • FixedProb.build_csr: emit a valid full-range CSR indptr when pre_ratio < 1.
  • FixedTotalNum: accept a fractional num as a connection fraction.
  • GridConn: de-duplicate periodic-boundary COO edges on small grids.

Optimizers

  • MomentumNesterov: apply the Nesterov look-ahead (it was previously identical to plain Momentum).
  • Adadelta: honor the learning rate.
  • LARS: clamp the degenerate zero-norm trust ratio to 1.

Training algorithms and runners

  • Ridge/LinearRegression: fix an IRLS while_loop seed that returned the untrained initial weights.
  • LogisticRegression: build the Newton diagonal with pure JAX ops, avoiding the __jax_array__ abstractification error.
  • OnlineTrainer: stop crashing on scalar monitor histories and keep time-major output in data_first_axis='T' mode.
  • BPTT.fit: accept a valid iterable dataset of exactly two (x, y) batches.
  • DSRunner.predict: forward shared_args in the memory-efficient path.

Object transforms

  • vector_grad: support the parameterized-decorator form (func=None).

Synaptic plasticity

  • dynold short-term plasticity (both implementations): remove a spurious source term so facilitation decays correctly when there are no presynaptic spikes.

Other fixes

  • dnn pooling: fix an adaptive-pooling ZeroDivisionError when a spatial dimension is smaller than its target (up-sampling), and add numpy typing stubs (#867).
  • dyn: PoissonGroup now spikes correctly (boolean rand_like dtype regression) (#862).
  • Fixed a SupportSTDP.stdp_update keyword typo and a DSRunner deprecation warning (#864).
  • Refreshed documentation notebooks and fixed clear_buffer/pmap runtime bugs (#863).

Build and Tooling

  • Drove package-wide mypy to zero errors (#863).
  • Moved __version__/__version_info__ into a dedicated brainpy._version module (#865) and read the packaged version statically at build time (#866).
  • Bumped codecov/codecov-action from 5 to 7 (#861).

Full Changelog: V2.8.0...V2.8.1

Assets 2
Loading