ponder-lab / Hybridize-Functions-Refactoring / #3652

90%

push

github

Recognize a Keras call override only through the engine's class hierarchy (#1131)

* Recognize a Keras call override only through the engine's class hierarchy

A call override was also recognized when PyDev's hierarchy yielded a base whose last name segment is `Layer` or `Model` (#881). With the built-in layer summaries declaring `Layer` as their superclass (wala/ML#1098), the engine's walk recognizes direct `Layer` and `Model` subclasses, `Sequential` subclasses and built-in layer subclasses, so the match matters only for a class named like a Keras base. There it is wrong: a program-defined class extending a class named `Layer` took `training` as supplied by Keras and lost its signature. Drop the match, and pin that case.

Fixes #1107.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTnpjPjJj2aZy1ZkUD1eAn

* Pin Keras layers across modules, under unmodeled bases, and nested

With the engine's class hierarchy as the only source of a Keras call
override, pin three shapes of Keras layer:

- A base layer defined in another module is recognized, so `training`
  keeps the signature withheld. The engine's cross-module resolution
  depends on translation order (wala/ML#944), which this fixture doesn't
  reach under any base-module name tried.
- A layer extending `keras.layers.Layer` or `MaxPooling2D` is not
  recognized, since the engine has no class for either base
  (wala/ML#1127), and gets a signature covering only `x`, which raises.
  Matching a base's last name segment withheld the first.
- Layers defined in a function and in a class are recognized.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTnpjPjJj2aZy1ZkUD1eAn

* Except Embedding from the built-in layers the walk resolves to nothing

Embedding already declares Layer as its superclass in the engine's
summary, so the walk reaches Layer from its subclasses.

Co-Authored-By: Claude Op... (continued)

1 of 1 new or added line in 1 file covered. (100.0%)

5295 of 5973 relevant lines covered (88.65%)

0.89 hits per line