ponder-lab / Hybridize-Functions-Refactoring / #2813

84%

main: 85%

push

github

Require a known rank for a parameter applied to a built-in Keras layer (#886)

An unknown-rank spec (shape=None) is sound as a type but not as an
input to a built-in layer: the layer's build reads a static axis off
the input shape (Conv2D the channel axis, Dense the last dimension) and
raises before the body runs, while the body itself contains no shape
read for the statically-read-axis walk to see (#883, observed on
TensorFlow2.0-Examples' CNN and deep_recommenders' FM).

A parameter-derived value applied to a call-graph target under the
summarized Keras namespace (or its $-prefixed trampoline, excluding
/class/ constructor nodes) now records a rank-only requirement,
resolving through the existing #809 machinery: only a dims-null spec
blocks, and a known rank with wildcard axes stays writable, which is
exactly what the layers admit. Ordinary modeled ops resolve to function
endpoints and stay outside the predicate.

The fixture pins the three directions: FM.call (differing call ranks,
Dense application) withholds its signature; Ranked.call (one rank,
wildcard batch) stays writable; op_only (differing ranks, tf.abs only)
keeps emitting an unknown-rank spec, so plain ops do not become
rank-sensitive. Validated two-sided: pre-fix FM.call records nothing
and the pin fails; post-fix the class runs 627/627.

Fixes #883.


Claude-Session: https://claude.ai/code/session_01X8W9PDNUSQFzZRUAYJmRiE

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

11 of 11 new or added lines in 1 file covered. (100.0%)

3215 of 3820 relevant lines covered (84.16%)

0.84 hits per line