ponder-lab / Hybridize-Functions-Refactoring / #1714

72%

main: 72%

push

travis-pro

Add `InputSignature.toTensorSpecList` formatter (#564)

* Add `InputSignature.toTensorSpecList` formatter

Phase 1 of #563. Adds the pure formatter that converts an `InputSignature`
to a Python source-code list of `tf.TensorSpec(shape=..., dtype=...)` calls
suitable for the `input_signature=` keyword argument of `@tf.function(...)`.
Handles dim discrimination (concrete `NumericDim` → integer literal; every
other dim sentinel → `None`), the rank-1 trailing-comma tuple form, scalar
empty-tuple form, all known dtypes, and the three import-prefix shapes
(`"tf."`, `"tensorflow."`, `""`).

Subsequent phases of #563 will wire this into `Function.convertToHybrid` and
implement the `Transformation.RECONFIGURE` case.

* Address Copilot findings on `toTensorSpecList`: shape-⊤ and locale

Two real defects flagged on PR #564:

1. `inferInputSignature()` can return `TensorType(dtype, null)` for shape-⊤
   (rank disagrees across contexts; see `Function.inferSpec` Step 2). The
   formatter dereferenced `t.getDims()` unconditionally, NPE-ing on this
   case. Now emits `shape=None`, matching TF's `tf.TensorSpec(shape=None,
   ...)` semantics for unknown rank.
2. `String.toLowerCase()` defaults to the JVM locale; under Turkish locale
   it lower-cases `I` to a non-ASCII dotless-i, corrupting `INT32` → `ınt32`
   and `STRING` → `strıng`. Now uses `Locale.ROOT`.

Two new tests pin the shape-⊤ path and lock locale-independence (sets the
default to `tr_TR` for the duration of the assertion).

* Drop unneeded `(List<Dimension<?>>) null` cast in `testShapeTopRendersAsNone`

13 of 13 new or added lines in 1 file covered. (100.0%)

1207 of 1665 relevant lines covered (72.49%)

0.72 hits per line