rel-2.1.0

The following changes were made in this release:

• Added logsumexp and logaddexp
• Numba backend improvements (e.g. less indirect function calls in generated code)
• Added NumPy Generator support for RandomVariables
• Added missing hermitian option to MatrixPinv
• Renamed aesara.tensor.nlinalg.lstsq class to Lstsq
• Updated aesara.tensor.slinalg.Solve to match the current SciPy interface
• Work around squeeze bug in SciPy samplers
• Added betainc Op
• Added cases for log1msigm and Softplus in local_exp_log opt
• Added stabilize opt for log1mexp
• Added a log1mexp Op
• Added log1pexp alias to softplus
• Added Numba support for RandomVariable Ops
• Introduced an aesara.tensor.linalg module that combines the nlinalg and slinalg modules
• Removed Hints
• Add Numba conversions for aesara.tensor.nlinalg.pinv and inv
• Preserve broadcastable dimensions in params_broadcast_shapes
• Added shape_tuple helper function
• Fixed broadcasting bug in broadcast_shape_iter

rel-2.0.12

The following changes were made in this release:

• Added aesara.tensor.take_along_axis
• Deprecated aesara.tensor.math.abs_ and replaced it with aesara.tensor.math.abs

rel-2.0.11

The following changes were made in this release:

• OpFromGraph.infer_shape now works with non-TensorTypes
• RandomVariable.__str__ now includes __props__ entries
• RandomVariable.destroy_map has been corrected
• aesara.scalar.basic_scipy has been renamed to aesara.scalar.math
• Numba conversion of Sum, Prod, and Switch have been fixed
• Numerous documentation fixes and updates
• Numba conversion of CAReduce no longer uses experimental Numba features
• Numba conversion of Elemwise Ops now support in-place updates

Repository changes

• Aesara has been moved to the aesara-devs organization
• master has been renamed to main

rel-2.0.10

The following changes were made in this release:

• aesara.tensor.math.inv was renamed to aesara.tensor.math.reciprocal
• Added broadcast-only DimShuffle case to aesara.tensor.random.opts.local_dimshuffle_rv_lift
• Created local_rv_size_lift local optimization
• Misc. documentation updates

rel-2.0.9

The following changes were made in this release:

• Explicitly add gcc DLL directory on Windows for Python >= 3.8
• Added chi-squared RandomVariable
• Fixed an attribute error in Scan's exception handling
• Reverted FunctionGraph replacement warnings
• Renamed aesara.graph.toolbox to aesara.graph.features
• Moved is_same_graph_with_merge and is_same_graph to aesara.graph.opt_utils
• Added aesara.graph.opt_utils.optimize_graph function to simplify running optimizations/rewrites on graphs
• Added atleast_*d and expand_dims functions
• Added broadcast_arrays function
• Renamed Query to OptimizationQuery and DB to OptimizationDatabase
• Exposed aesara.tensor.random.basic via aesara.tensor.random, so, now, random variables can be created via import aesara.tensor as aet; aet.random.*
• NumPy dtype objects (e.g. np.int32, np.float64) can be used when specifying dtypes to tensor and scalar construction functions and cast

rel-2.0.8

The following changes were made in this release:

• Added optimizations for log(exp(x)), exp(log(x)), and exp(log1p(x))
• AdvancedIncSubtensor Ops (e.g. x[idx] = ...) can now be in-place optimized
• Sigmoid and Softplus have been moved to aesara.scalar.basic_scipy, and expit was added as an alias to sigmoid
• Added a Numba graph compilation backend and a NUMBA compile mode option (e.g. aesara.function(..., mode="NUMBA"))

rel-2.0.7

The following changes were made in this release:

• SpecifyShape now works for scalar shapes
• aesara.tensor.as_tensor_variable now takes an optional dtype argument
• JAX mode now uses jax.vmap for Elemwise Ops and constructs much more concise FunctionGraph conversions

rel-2.0.6

The following changes were made in this release:

• broadcast_to, diff, and unique are now exposed at aesara.tensor package level
• The FunctionGraph constructor now has options for automatically determining the inputs and providing an existing clones dict to be updated in-place
• More numpy.random distribution Ops have been added: geometric, hypergeometric, lognormal, triangular, weibull, logistic, vonmises
• Op.view_map and Op.destroy_map are now present by default
• A measure was added to prevent BroadcastTo outputs from being updated in-place

rel-2.0.5

The following changes were made in this release:

• Added Gumbel and Pareto RandomVariables
• Catch JAX errors caused by attempting to disable omnistaging in newer versions

rel-2.0.4

The following changes were made in this release:

• Removed tests module from Python package
• MultinomialRV will now broadcast across its parameters
• Fixed a bug in CategoricalRV when size is a scalar