cirkit 0.2.0

Release Notes

Documentation changes

#294 #303 #304 #323 Improved overall documentation, both high level (on the readthedocs website) and low level (docstrings)
#328 Added notebook about instantiating and learning a generative circuit classifier for MNIST images, which also showcase variable marginalization for robust classification in presence of missing features

New features

#312 Implemented logic circuits and the parsing of sentential decision diagrams (SDDs) from the PySDD library
#292 Perform batched variable marginalization by specifying a marginalization mask
#312 Added routine to graphically plot symbolic circuits
#309 Simplified overall symbolic representation by having a single sum layer that generalizes the previous dense and mixing layers
#324 Added the CP tensor factorization as circuit construction template

Additional new features

#300 The image data circuit instantiation template now supports Gaussians as input distributions
#325 Added more flexibility in the circuit construction templates by specifying additional hyperparameters arguments to the Parameterization data class
#316 The correctness and consistency of the symbolic representation is now checked before compilation
#313 Added indexed Categorical and embedding implementations, which are generally faster and more memory efficient than the previously one based on one hot encoding followed by an einsum
#330 Added optimization rule replacing kronecker products followed by sum in parameters into einsum

Bug fixes

#311 #319 Fixed NaNs arising from the implementations of Categorical and sum layers when the evaluation is in the log-space
#306 Fixed the region graph to circuit template method as to ensure circuits built with CP layers have a sum as output layer

Library management

#329 Updated requirements, now allowing numpy 2.x versions. Plus, add opt_einsum for possibly faster einsums

Detailed Release Notes (automatically generated)

• Made dense/mixing layer the last layer even if sum_product='cp' by @loreloc in #308
• Enable gaussian input in image_data function by @isurulucky in #301
• added indexed categorical layer evaluation by @gengala in #313
• Fix categorical nans by @loreloc in #317
• Update README.md by @andreasgrv in #315
• Elaborate on what is implemented in Cirkit by @andreasgrv in #321
• Refactor sum layers + consistency checks + bug fixes by @loreloc in #314
• Batched marginalisation mask by @andreasgrv in #318
• Plotting symbolic circuits using graphviz by @n28div in #322
• implemented logic circuit and parsing of SDD by @n28div in #312
• support passing additional activation parameters in parameterizations by @isurulucky in #325
• Cleaning up by @loreloc in #326
• Mkdocs math support + more and better docstrings + interfaces fix by @loreloc in #323
• CP factorization by @loreloc in #324
• Improve Documentation by @andreasgrv in #327
• MNIST Classification notebook by @andreasgrv in #328
• Avoid OOM and speed up training in SOS circuits by @loreloc in #330
• Upgrade to numpy 2.x and add opt_einsum by @loreloc in #329
• Fix unconditional sampling and add unit test by @loreloc in #332
• fix issues in logic circuits by @n28div in #333

New Contributors

• @n28div made their first contribution in #322

Full Changelog: 0.1.0...0.2.0

cirkit 0.1.0

First release of the cirkit library.

Technical release notes:

• multi-level APIs to construct circuits: low level - by manipulating the computational graph - and high level - by using region graphs and templates.
• initial symbolic representation of circuits, parameters and pipelines
• support for a first minimal set of circuit operators: concatenate, evidence, integrate, multiply, differentiate, conjugate
• simple PyTorch compilation backend supporting in-pipeline parameter sharing, folding, computational graph optimizations using a rule-matching procedure, and evaluation using semirings
• PyTorch implementation of bottom-up unconditional sampling and variable marginalization