ponder-lab / Hybridize-Functions-Refactoring / #3420

89%

Build Type

push

github

Committed by web-flow

Commit Message

Withhold the signature when a static dimension reaches a TensorFlow shape op (#1059)

* Withhold the signature when a static dimension reaches a TensorFlow shape op

A dimension read statically from an axis the signature leaves wild is None at trace time. The check already treated tf.reshape, tf.range and tf.image.resize as consuming it, but not the other TensorFlow ops that take a shape, so a function passing x.shape[0] to tf.ones kept a signature it fails under (#1058). Each of these raises on a None dimension in TensorFlow 2.9.3, and is now a sink: tf.ones, tf.zeros, tf.fill, tf.eye, tf.random.normal, tf.random.uniform, tf.random.truncated_normal, tf.broadcast_to, tf.tile, tf.one_hot and tf.split. So are the Python builtins range and int, which raise TypeError on None.

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

* Count a dimension only in the arguments it breaks

Each new shape op now names the arguments in which a None dimension raises at trace time, and only those count. Arguments that default to None accept it and change meaning instead, so counting them withheld signatures the function traces under: tf.random.uniform's maxval means 1 for a float dtype, tf.eye's num_columns means square, and tf.split's num, tf.one_hot's axis and on_value, and seed trace too (TensorFlow 2.9.3). The test gains two of those as controls that keep their signatures.

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

* Count tf.random.uniform's maxval unless its dtype is a float

With a float dtype, the default, a None maxval means 1. With an integer dtype it raises, since the samples come from [minval, maxval). So maxval counts unless the dtype, by keyword or position, resolves to a float type or is absent. A dtype that can't be resolved counts too, in the withholding direction.

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

Coverage Stats

42 of 43 new or added lines in 1 file covered. (97.67%)

1 existing line in 1 file now uncovered.

4852 of 5510 relevant lines covered (88.06%)

0.88 hits per line