ponder-lab / Hybridize-Functions-Refactoring / #2539

82%

main: 83%

push

github

Omit defaulted non-tensor parameters from an inferred input signature (#788)

* Omit defaulted non-tensor parameters from an inferred input signature

`computeInputSignature` treated every non-tensor parameter as fatal, which is
stricter than TensorFlow. A `TensorSpec` is required per *required* argument,
not per argument: on TF 2.9.3, `@tf.function(input_signature=[tf.TensorSpec(
shape=[2], dtype=tf.float32)])` over `def f(x, training=True)` is accepted and
returns, while the same signature over `def f(x, training)` raises `TypeError:
The decorated function f has 2 required argument(s), but tf.function was only
passed an input_signature of length 1`. TensorFlow's own message states the
rule. So a defaulted `training` was declining signatures TensorFlow would take.

Omitting the parameter is not unconditional, though, because it changes the
hybridized function's arity: the function then accepts exactly what the
signature names, and `f(tf.constant([1.0, 2.0]), False)` starts raising
`TypeError: f(x, training) specifies 1 positional arguments, but got 2`. Two
conditions gate it, and each gets its own reason so the two stay separable:

- `DEFAULTED_PARAMETER_SUPPLIED`: some call site passes the parameter, so
  omitting it would break that caller. Also covers ignorance (no call-graph
  node, unresolvable predecessor, non-Python invoke), since absent evidence
  must not read as evidence of absence.
- `DEFAULTED_PARAMETER_PRECEDES_SPEC`: nobody passes it, but a later parameter
  contributes a spec. An input signature covers parameters positionally, so
  omitting this one would apply the later spec to it. Python's grammar forces
  defaults to trail the required parameters but still permits a defaulted
  tensor parameter after a defaulted non-tensor one, which is the
  `def call(self, x, training=False, mask=None)` Keras shape.

`NON_TENSOR_PARAMETER` narrows to the required case, which is genuinely
unsatisfiable.

The supply check is the first analysis here ... (continued)

75 of 83 new or added lines in 4 files covered. (90.36%)

2213 of 2696 relevant lines covered (82.08%)

0.82 hits per line