ponder-lab / Hybridize-Functions-Refactoring / #3534

89%

push

github

Recognize a Keras call override through the engine's class hierarchy, pinning the built-in-layer case on wala/ML#1098 (#1102)

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

The #881 rule recognizes a `call` override only when PyDev's hierarchy
names a `Layer` or `Model` base by its last segment, and PyDev stops at a
base written through `tf.keras`. Walk the engine's bases as well, matching
its `Layer` and `Model` summary classes exactly, so a `tf.keras.Sequential`
subclass is recognized. Subclasses of built-in layers are pinned at
today's wrong signature until the layer summaries declare a superclass
(wala/ML#1098). Refs #1099.

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

* Cover the engine's Keras-class walk with no dead branch

Patch coverage flagged the walk's missing-class return and the catch
around its call in the processor. Walk type names instead of classes,
relying on the engine registering every name it hands out, and run the
walk inside the supplied-parameter step's existing catch. Add a non-Keras
diamond over `object`, which revisits a shared base, follows `object` to
the root, and keeps `training` omittable. Refs #1099.

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

* Skip a name with no class in the engine's Keras-class walk

The walk's first name is derived from the declaring class's, and the
bases are resolved at the loader (wala/ML#1006), whose cross-module
resolution depends on translation order (wala/ML#944), so neither is
promised to be a class of the hierarchy. Skip such a name on the revisit
line, leaving the PyDev walk to decide, rather than dereferencing null.
Refs #1099.

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

---------

Co-authored-by: Claude Opus... (continued)

26 of 26 new or added lines in 2 files covered. (100.0%)

5212 of 5883 relevant lines covered (88.59%)

0.89 hits per line