v0.6.0

What’s Changed

Sampling Backprop Op

• Add a SumOpSampleBackprop so that sampling is done easily.
• Add a zero_evidence_inference method to SequentialSumProductNetwork that can be used to infer values for all variables. Combined with SumOpSampleBackprop this results in samples from the network.
• Add a notebook with an example of how to sample from an SPN

Simplify PoonDomingosMeanOfQuantileSplit

• Remove support for normalization within. Expect data to be normalized.
• Accept tf.data.Dataset to allow for more consistency with class and method signatures of other library components.

Add PoonDomingosStddevOfQuantileSplit

• Add initializer for leaf scales based on same logic as PoonDomingosMeanOfQuantileSplit for leaf locations.

Other

• Fixes in docs
• Loosen dependency constraints so that installing libspn-keras is less likely to result in conflicts

v0.5.2

What’s Changed

      * Fix issues with log conv (#27) @jostosh

v0.5.1

What’s Changed

• Fixes Conv2DSum forward pass
• Adds Clip constraint
• Adds constraint parameter to NormalLeaf layer
• Adds option to normalize with cross-sample statistics for NormalizeStandardScore layer

v0.5.0

What’s Changed

• Add normalized greater equal epsilon constraint
• Simplify API for building random SPNs. No need to determine factors array for PermuteAndPadScopesRandom anymore when using libspn_keras.models.SequentialSumProductNetwork.
• Normalize EM updates and use 'gliding averages'
• Add notebook for benchmarking against Einsum Networks
• Add Dirichlet initializer for accumulators

Modular Sum Op control and global configs

What’s Changed

• minor fixes to actions (#14) @jostosh
• Add tooling and modular sum op control (#13) @jostosh
• Use batch first always (#12) @jostosh
• Feature/add accumulators location scale leafs (#10) @jostosh
• add image (#7) @jostosh
• Refactor dense ratspn (#6) @jostosh
• Setup, README and minor fixes (#2) @jostosh