ponder-lab / Hybridize-Functions-Refactoring / #3583

89%

main: 90%

push

github

Refuse reads of a tensor field another method rebinds (#1093)

* Refuse reads of a tensor field another method rebinds

A method was converted when it read a tensor from a field of self that
another method rebinds on every call, so the converted method returned
the first call's state. Measured on TensorFlow 2.9.3, with each reader
alone under tf.function and its state set up twice with different
tensors, the second call returns the first call's state for distilled
TensorFlowTTS readers (Tacotron2Sampler.next_inputs, the location
sensitive attention's __call__ and _compute_window_mask, and
TFTacotronDecoder.step) and OpenNMT-tf's
AttentionalRNNDecoder._get_initial_state.

A new analysis collects the fields the function and every node it
reaches read into a value the tensor-type analysis types as a tensor,
using ModRef's ref sets, and refuses the function when a node outside
its closure, other than __init__ or build, writes one of them, per
ModRef's mod sets. A function that does fails a new precondition,
HAS_REBOUND_STATE_READS (39). The elements of a list, tuple,
dictionary, or set are not read as state, since tf.function passes a
container argument in afresh. A field set only in __init__ or build, a
string field used as an op name, a variable updated with assign, a list
argument's elements, and a holder created and read within one call are
left alone. A getter of an object created afresh on each call is
refused, since objects are told apart by allocation site.

OpenNMT-tf's make_features fills in the dictionary its caller passes,
which tf.function refuses with a ValueError. The side-effect
precondition already refuses it when the caller passes a dictionary it
built. It misses the caller that passes a copy, as OpenNMT-tf's Model
does, because dict.copy is not modeled and the copy points to nothing.
That case is pinned with a TODO, as is a reader whose memory is a
tensor multiplied by a float computed from a loop variable, whose
product has no tensor t... (continued)

84 of 91 new or added lines in 4 files covered. (92.31%)

5289 of 5967 relevant lines covered (88.64%)

0.89 hits per line