PrincetonUniversity / PsyNeuLink / 19093589251

85%

master: 85%

Build Type

push

github

Committed by web-flow

Commit Message

feat/autodiff/devel_additional_optimizations (#3379)

- Add customizations for multi-optimization in PyTorch
- EGO is now numerically identical to pytorch for any num_optimizations

- composition.py, autodiffcomposition.py
  - get_targets() -> get_target_nodes()

- autodiffcomposition.py
  - add custom_optimization, call_after_last_optimization and call_after_last_optimization
  - add exclude_from_gradient_calc

- composition.py
  - _validate_and_parse_additional_optimizations()
  - _after_first_optimization()
  - _after_last_optimization()
  - execute_in_additional_optimizations(): runs, needs to be validated
  - learn():  store passed in value of optimizations_per_minibatch to composition.optimizations_per_minibatch(context)
  - refactor execute_in_additional_optimizations to use value for optimization_num rather than reparameterization

- emcomposition.py
  - add store_on_optimizations Parameter
  - __init__(): change scheduling condition for storage_node from AllHaveRun to AfterNodes
  - allow field_weight to be specified as False (=None)

- compositionrunner.py
  - _batch_inputs():  add calls to call_after_last_optimization and call_after_last_optimization

- pytorchEMwrappers.py
  - store_memory():  restored assert, runs.
  - execute(): use store_on_optimizations

- remove prints
- fix codestyle formatting errors
- EMComposition: execute_in_additional_optimizations: handle no storage_node case
- tests: MDF: xfail default EMComposition (now uses AfterNodes by default, like GRU)
- CompExecution: handle varying shape of generator inputs
- emcomposition: fix missing DEFAULT_LEARNING_RATE import
- CompositionRunner: batch methods: fix return typehints
- fix various style errors

---------

Co-authored-by: jdcpni <jdc@princeton.edu>
Co-authored-by: Younes Strittmatter <younes_strittmatter@brown.edu>
Co-authored-by: Younes Strittmatter <ystrittm@gmail.com>

Run Details

10228 of 13318 branches covered (76.8%)

Branch coverage included in aggregate %.

81 of 99 new or added lines in 9 files covered. (81.82%)

2 existing lines in 2 files now uncovered.

35351 of 40599 relevant lines covered (87.07%)

0.87 hits per line