PrincetonUniversity / PsyNeuLink / 6224391157 / 1

85%

master: 85%

push

github-actions

Feat/learning nested (#2801)

• composition.py,
  autodiffcomposition.py,
  compositionrunner.py:
  - refactor to get targets directly in PyTorch mode, rather than creating PNL learning components

• composition.py
  - _parse_sender_spec(): check for CIM if sender == None and assign to sender.composition if so
  - various methods: learning_projection arg -> is_learning_projection
  - standardize passing of is_learning_projection arg to add_projection for all learning creation methods
  • _parse_sender_spec():  sender = sender.owner -> just projection.sender
  - _get_covariate_info(): change test for length of variable from >= 1 to > 1
  - _get_back_prop_error_sources(): handle nested comp (for use by autodiff, not regular Composition)
  - error message for nested composition in Python learning
  - infer_backpropagation_learning_pathway: move to autodiffcomposition.py
  - get_required_roles_by_node(): added method
  - _create_CIM_ports(): 
    - deal with errant input_ports and 
    - fix bug in validating equal number of input_ports to output_ports
    - assign learnable=False to MappingProjections
    - mods to clean up handling of projections to output_CIM, prunining extraneous ones that we getting assigned and left in place when no longer appropriate, and more precisely determining when a node should be considered an OUTPUT and/or PROBE note, all in support of better handling multiple projections from OUTPUT Nodes of nested Compositions.
  - get_nested_node_CIM_port():  
    - if an OUTPUT NODE of a nested Composition has a port that does *not* project to the output_CIM (i.e., it projects to some other Node in the nested composition), and it is referenced by an OCM of the outer composition, then assign it as NodeRole.PROBE
  - _create_non_terminal_backprop_learning_components():
    - deal with duplication of error_projections in overlapping pathways
    - refactor to deal with addition of error_signals to existing learning_... (continued)

12955 of 16046 branches covered (0.0%)

Branch coverage included in aggregate %.

31011 of 35844 relevant lines covered (86.52%)

0.87 hits per line