v0.12.0.0

[0.12.0.0] - 2022-07-27

Component

• Don't exclude _init_args from deepcopy (#2380)
• Set correct Parameter owner on Component deepcopy

Composition

• Fix TERMINAL node detection (#2384)
• Add_projection: also detect feedback from node roles (#2429)

ControlMechanism

• Set function default to Identity
• Set monitor_for_control default to []

DriftOnASphereIntegrator

• Change default dimension to 3 (#2428)

JSON/MDF

• Determine fixed onnx noise values at test time (#2441)

KWTAMechanism

• Set matrix according to tests (#2398)

MDF

• Support v0.3.4; import from MDF classes
• Update interface methods
• Add yaml_summary
• Rename JSONDumpable to MDFSerializable
• Update all documentation from JSON to MDF
• Rename json files to mdf
• Remove manual arg substitution in expressions

MemoryFunction

• Use Buffer _update_default_variable fix
• _update_default_variable: handle ragged arrays

Parameter

• Remove delivery_condition from _additional_param_attr_properties
• Add specify_none flag

Parameters

• Fix Parameter inheritance when setting attributes (#2402)
• Allow specification in inner class or init

Testing

• Add checks for ControlMechanism allocation parameter
• Make sure all 'cuda' marked tests are also marked 'llvm'
• Add cmdline option to select compiler fp precision
• Add fp32 variant
• Consolidate matrix ops tests
• Drop fixed dimension vector matrix multiply
• Use pnlvm.ir instead of importing llvmlite ir again
• Convert arguments to the correct fp precision
• Consolidate spelling of 'Philox' to easily identify philox tests
• Convert parameters to correct type in PTX tests
• Parameters: add check for _additional_param_attr_properties
• Make sure inputs are representable in fp32
• Use higher tolerance when running in fp32 mode
• Refactor test parameters
• Add fp32 expected values
• Add Philox fp32 results
• Remove dead code
• Add fp32 results
• Add testing for fp32 results and fp32 precision
• Adjust probabilities for storage/retrieve
• Pass PRNG type as explicit test parameter
• Add special case result for PTX fp32 test
• Refactor tests to use pytest.param(id=)
• MDF: correct os-specific filepaths

Components

• Add default_variable as constructor_argument where missing
• Add check_user_specified decorator for _user_specified

Function/OptimizationFunction

• Allow array based SampleIterator in compiled code path
• Move Python specific code to non-compiled else branch

Functions/GridSearch

• Rename '_run_grid' -> '_search_grid'
• Refactor parallel compiled path to reuse existing '_evaluate' implementation
• Refactor parallel compiled path to reuse existing '_evaluate' implementation (#2445)

Functions/OptimizationFunction

• Disable aggregation in compiled mode

Github-actions

• Use v0.0.0.0 as dummy tag to compare documentation on pull requests (#2376)
• Bump actions/github-script from 5 to 6 (#2317)
• Bump actions/download-artifact from 2 to 3 (#2378)
• Bump actions/setup-python from 3 to 4 (#2425)
• Install latest pip and wheel when testing new dependencies (#2444)
• Drop python3.7 macos job (#2453)

Github-actions/install-pnl

• Fix lookup of dep packages (#2392)

Github-actions/pnl-ci-docs

• Skip removing zero tag if the tag creation was skipped (#2393)

Github-actions/pnl-install

• Check if a newly bumped dependency is rolled back during installation (#2390)

Llvm

• Split allocating new space in params with history from populating it
• Add support for different floating point precision conversions
• Add option to use 32bit floating point representation (#2400)
• Add support for more SoftMax output types (#2418)
• Always emit debuginfo with LLVM IR
• Cleanup codegen of mechanisms that modify function return value (#2421)
• Improve parallel execution handling (#2443)

Llvm/ConditionGenerator

• Refactor structure/initializer methods to not use type to detect composition

Llvm/UDF

• Reuse metadata from variable builder in codegen builder
• Update debug code location for every parsed AST node
• Add detailed line debugging information to generated IR (#2419)

Llvm/builder_context

• Use llvm.ir module directly
• Simplify debug info handling
• Use load_extract_scalar_array_one helper to load PRNG seed

Llvm/builtins

• Split 'is_close' builtin implementation by type
• Change 'tanh' formula to better handle large input
• Change 'coth' formula to better handle large input
• Change 'csch' formula to better handle large input
• Drop dead code

Llvm/debug

• Print parallel execution time when "time_stat" is set

Llvm/execution

• Add support for structs to _element_dtype
• Check for errors in jobs submitted to thread pool executor

Mechanism/ControlMechanism

• Drop DefaultAllocationFunction
• Remove obsolete workaround

Mechanisms/ControlMechanism

• Drop unused import of psyneulink.llvm

Requirements

• Update pillow requirement from <9.1.0 to <9.2.0 (#2370)
• Update pandas requirement from <=1.4.1 to <1.4.3 (#2373)
• Update autograd requirement from <=1.3 to <1.5 (#2379)
• Update graphviz requirement from <0.20.0 to <0.21.0 (#2389)
• Update networkx requirement from <2.8 to <2.9 (#2383)
• Update pytest requirement from <7.1.2 to <7.1.3 (#2396)
• Update psyneulink-sphinx-theme requirement (#2426)
• Update graph-scheduler requirement (#2391)
• Update pillow requirement from <9.2.0 to <9.3.0 (#2436)
• Update pandas requirement from <1.4.3 to <1.4.4 (#2434)
• Update elfi requirement from <0.8.4 to <0.8.5 (#2427)
• Update matplotlib requirement from <3.5.2 to <3.5.3 (#2403)
• Update torch requirement from <1.9.0 to <1.12.0 (#2446)
• Update pint requirement from <0.18 to <0.20 (#2447)
• Update llvmlite requirement from <0.39 to <0.40 (#2451)
• Bump pycuda to <2023 (#2452)
• Update numpy requirement from <1.21.4 to <1.21.7 (#2454)

Utilities

• Combine signature caches