pytorch/caffe2/python/regularizer_context.py
Tristan Rice dc7d8a889e caffe2: refactor context to allow being typed (#48340)
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/48340

This changes the context managed classes from using a decorator to define them to using inheritance. Inheritance allows the python static type checking to work correctly.

```
context.define_context()
class Bar(object): ...

context.define_context(allow_default=True)
class Foo(object): ...
```

becomes
```
class Foo(context.Managed): ...

class Bar(context.DefaultManaged): ...
```

Behavior differences:
* arg_name has been removed since it's not used anywhere
* classes need to call `super()` in `__enter__/__exit__` methods if they override (none do)

This also defines a context.pyi file to add types for python3. python2 support should not be affected

Test Plan:
ci

  buck test //caffe2/caffe2/python:context_test //caffe2/caffe2/python:checkpoint_test

Reviewed By: dongyuzheng

Differential Revision: D25133469

fbshipit-source-id: 16368bf723eeb6ce3308d6827f5ac5e955b4e29a
2020-11-30 18:31:14 -08:00

38 lines
1013 B
Python

# @package regularizer_context
# Module caffe2.python.regularizer_context
from caffe2.python import context
from caffe2.python.modifier_context import (
ModifierContext, UseModifierBase)
class RegularizerContext(ModifierContext, context.DefaultManaged):
"""
provide context to allow param_info to have different regularizers
"""
def has_regularizer(self, name):
return self._has_modifier(name)
def get_regularizer(self, name):
assert self.has_regularizer(name), (
"{} regularizer is not provided!".format(name))
return self._get_modifier(name)
class UseRegularizer(UseModifierBase):
'''
context class to allow setting the current context.
Example usage with layer:
regularizers = {'reg1': reg1, 'reg2': reg2}
with UseRegularizer(regularizers):
reg = RegularizerContext.current().get_regularizer('reg1')
layer(reg=reg)
'''
def _context_class(self):
return RegularizerContext