CHANGELOG.md

Changelog

All notable changes to this project will be documented in this file. The format is based on Keep a Changelog, and this project adheres to Semantic Versioning.

Unreleased

3.8.0

• Move functionality from CenteredRBMs.jl to this package.

3.7.0

• Move functionality from StandardizedRestrictedBoltzmannMachines.jl to this package.

3.6.0

• Move functionality from RestrictedBoltzmannMachinesHDF5.jl to this package, through the Extensions mechanism.

3.5.0

• Move functionality from CudaRBMs.jl to this package, through the Extensions mechanism.
• Move PottsGumbel to this package.

3.4.1

• Use default_rng in place of GLOBAL_RNG, which fixes an issue with CUDA e84438a.

3.2.5

• Compat with Optimisers v0.3.

3.2.0

• Patch CUDA issue with Potts due to https://github.com/JuliaGPU/CUDA.jl/issues/1957.
• callback now receives state too.

3.1.0

• visible_cgf, free_energy_h, free_energy_v, and gradients.

3.0.0

• BREAKING: rescale_hidden! returns true or false depending on whether the hidden units were rescaled or not.

2.2.0

• Initialization from data for dReLU, pReLU, xReLU.

2.1.1

• Fix mirror.

2.1.0

• Allow division of ∂RBM by a scalar.

2.0.4

• Always convert to weights eltype in ∂interaction_energy (previously, converted only in some cases).

2.0.3

• Convert to weights eltype in ∂interaction_energy (fix #10).

2.0.2

• Fix batch weight correction (https://github.com/cossio/RestrictedBoltzmannMachines.jl/commit/355b5faae78d268f083787c7f92a0e995eee6116).

2.0.1

• Close CUDA issue #20.

2.0.0

• This CHANGELOG file.
• Rescale weights to norm 1, instead of hidden unit activities to unit variances. This is a simpler way to settle the scale degeneracy between weights and hidden unit activities for continuous hidden units. Commit here.
  • Introduce rescale_weights!(rbm) to normalize weights attached to each hidden units.
  • Now pcd!(...; rescale=true, ...) uses rescale_weights!, instead of scaling hidden unit activities to unit variances.
  • BREAKING: Removed ρh, ϵh keyword arguments from pcd!, which used to control the tracking of hidden unit variances during training.
  • BREAKING: grad2var has been removed.
• Allow passing ps, state to pcd! to control which parameters are optimized. Now pcd! returns state, ps, which can be breaking. Commit here.

1.0.0

• Release v1.0.0.

0.39

• Now pcd!(...; iters=n, ...) performs n gradient updates. This replaces the epochs setting.
• pcd now uses the Optimisers framework instead of Flux. In particular, optim expects a Optimisers.AbstractRule.

0.38