48cb1ae640
Adds a dependency on wrapt to TensorFlow's pip package. It's not a very large dependency, and this is a usability win when subclassing TensorFlow types (Model, Module, Layer, etc.).
wrapt fixes issues with CPython's direct access to dictionaries by inheriting the memory layout of the wrapped object, so we can now pass isinstance(obj, dict) checks while staying correct (i.e. {}.update(obj) won't look at the possibly-stale wrapper object's memory). There are several new oddities with the wrapt approach, but overall this is much better.
We're already doing dict wrapping, just not in a way that passes isinstance(obj, dict) checks. We need it to support restore-on-create for variables added to objects while executing eagerly.
I tried switching _ListWrapper to wrapt too. It works in most Python versions (with some tweaks to TF to accommodate the change), but apparently in Python 3.4 isinstance(obj, (list, tuple)) is not functioning properly. There are no correctness issues with actually subclassing list, and it means type(obj) shows up as a subclass of list, so this isn't too bad. Since ObjectProxy copies the memory layout of the wrapped object we can't do both of these at the same time.
PiperOrigin-RevId: 243118363
0 lines
Python
0 lines
Python