ponder-lab / Hybridize-Functions-Refactoring / #4191

90%

push

github

Refuse colliding layer names only when a run saves or loads weights by name (#1279)

* Refuse colliding layer names only when a run saves or loads by name

The colliding-names shape of MAY_BUILD_LAYERS refused a method whenever two of its sublayers' variables could get the same name once decoration drops the enclosing layers' name scopes. The hazard is a save or a load keyed by name, such as to HDF5, which then fails with "name already exists" or mixes weights. The rule never asked whether the program keys weights by name, so it refused methods in programs that save a SavedModel or set weights by position, such as NLPGNN's GPT-2 MLP.call and deep_recommenders' FM.call.

The shape now refuses only when an entry point of the call graph that reaches one of the method's nodes also reaches a call in program code that may save or load weights by name. Such a call is one of an attribute named init_from_checkpoint or Saver; of save, save_weights or save_model whose save_format is HDF5 or not a constant, or else whose path may name an HDF5 file; of load_weights passing a by_name other than the constant False, or else whose path may name one; or of load_model or ModelCheckpoint whose path may name one. A path names no HDF5 file when it is a constant, or an os.path.join whose last part is a constant, that doesn't end in .h5, .hdf5 or .keras. A save_weights in the "tf" format keys nothing by name, but save and save_model still go by the path's suffix under "tf", since TensorFlow 2.9.3 saves them to HDF5 whenever the suffix says so. A method node that no entry point reaches, a call unpacking ** keywords, and any path the IR doesn't show as such a constant all count as keying by name.

Each test that pinned the colliding-names refusal without a name-keyed save now expects no refusal, and a copy of it whose fixture saves the model's weights to HDF5 keeps the refusal. New tests cover each kind of call that keys weights by name, the saves that don't, and a save in ano... (continued)

155 of 162 new or added lines in 1 file covered. (95.68%)

7095 of 7889 relevant lines covered (89.94%)

0.9 hits per line