import contextlib
import functools
import operator
import sys
import threading
import numpy
import six
import chainer
from chainer.backends import cuda
_thread_local = threading.local()
@contextlib.contextmanager
def get_function_check_context(f):
try:
default = _thread_local.current_function
except AttributeError:
default = None
_thread_local.current_function = f
try:
yield
finally:
_thread_local.current_function = default
class TypeInfo(object):
"""Type information of an input/gradient array.
It contains type information of an array, such as the shape of array and
the number of dimensions.
This information is independent of CPU or GPU array.
"""
def __init__(self, shape, dtype):
self.shape = shape
self.dtype = dtype
self.ndim = len(shape)
@property
def size(self):
return functools.reduce(operator.mul, self.shape, 1)
class TypeInfoTuple(tuple):
"""Type information of input/gradient tuples.
It is a sub-class of tuple containing :class:`TypeInfo`. The i-th element
of this object contains type information of the i-th input/gradient data.
As each element is :class:`Expr`, you can easily check its validity.
"""
def size(self):
"""Returns an expression representing its length.
Returns:
Expr: An expression object representing length of the tuple.
"""
return Variable(len(self), '{0}.size'.format(self.name))
class LightTypeInfoTuple(tuple):
"""Type information of input/gradient tuples for light-weight check.
It is a sub-class of tuple containing :class:`TypeInfo`. The i-th element
of this object contains type information of the i-th input/gradient data.
"""
def size(self):
"""Returns its length.
Returns:
int: Length of the tuple.
"""
return len(self)
def get_types(data, name, accept_none, *, shapes=None):
assert isinstance(data, tuple)
if shapes is None:
shapes = tuple([x.shape for x in data])
info = TypeInfoTuple(
_get_type(name, i, x, accept_none, shape)
for i, (x, shape) in enumerate(zip(data, shapes)))
# I don't know a method to set an attribute in an initializer of tuple.
info.name = name
return info
def get_light_types(data, *, shapes=None):
assert(isinstance(data, tuple))
if shapes is None:
data_ = data
else:
# For non-default memory format (e.g. NHWC), shapes of data are
# different from the semantic shapes (e.g. NCHW). In such cases
# semantic shapes are explicitly given as `shapes` argument.
# If it is given, TypeInfos with modified shapes are wrapped.
data_ = tuple([
TypeInfo(shape, x.dtype) for x, shape in zip(data, shapes)])
return LightTypeInfoTuple(data_)
def _get_type(name, index, array, accept_none, shape):
var = '{0}[{1}]'.format(name, index)
if accept_none and array is None:
# case that gradient is not given
return Variable(TypeInfo((), None), var)
assert isinstance(array, chainer.get_array_types())
return Variable(TypeInfo(shape, array.dtype), var)
def _make_un_operator(exp, priority, func):
def f(x):
return UnaryOperator(priority, x, exp, func)
return f
def _make_bin_operator(exp, priority, func, right_associative=False):
def f(x, y):
return BinaryOperator(priority, x, y, exp, func, right_associative)
return f
def _make_bool_operator(exp, inv, func):
def f(x, y):
return BoolBinaryOperator(x, y, exp, inv, func)
return f
def _flip(f):
return lambda x, y: f(y, x)
class Expr(object):
"""Abstract syntax tree of an expression.
It represents an abstract syntax tree, and isn't a value. You can get its
actual value with :meth:`eval` function, and get syntax representation with
the :meth:`__str__` method.
Each comparison operator (e.g. ``==``) generates a new :class:`Expr` object
which represents the result of comparison between two expressions.
.. admonition:: Example
Let ``x`` and ``y`` be instances of :class:`Expr`, then ::
>>> x = Variable(1, 'x')
>>> y = Variable(1, 'y')
>>> c = (x == y)
is also an instance of :class:`Expr`. To evaluate and get its value,
call :meth:`eval` method::
>>> c.eval()
True
Call ``str`` function to get a representation of the original
equation::
>>> str(c)
'x == y'
You can actually compare an expression with a value::
>>> (x == 1).eval()
True
Note that you can't use boolean operators such as ``and``, as they try
to cast expressions to boolean values::
>>> z = Variable(1, 'z')
>>> x == y and y == z # raises an error
Traceback (most recent call last):
RuntimeError: Don't convert Expr to bool. Please call Expr.eval \
method to evaluate expression.
"""
def __init__(self, priority):
self.priority = priority
def eval(self):
"""Evaluates the tree to get actual value.
Behavior of this function depends on an implementation class.
For example, a binary operator ``+`` calls the ``__add__`` function
with the two results of :meth:`eval` function.
"""
raise NotImplementedError()
def __getattr__(self, name):
return GetAttr(self, name)
def __getitem__(self, key):
return GetItem(self, key)
def __call__(self, *args):
return Call(self, args)
def __nonzero__(self):
# When a user calls a boolean operator like `(x == y and z == w)`,
# `and` operator evaluate the first expression.
# If it returns `True` (and it's default behavior), the `and` operator
# returns *the second expression*, not a boolean value.
# So, `(x == y and z == w)` returns the result of `z == w`, and
# `(x == y and z == w).expect()` raise no errors but only checks
# `z == w`. It is confusing.
# See also:
# https://docs.python.org/3/library/stdtypes.html
msg = ('An Expr instance cannot be evaluated as bool. '
'Please use chainer.utils.type_check.eval() to evaluate an '
'expression.')
raise RuntimeError(msg)
def __bool__(self):
self.__nonzero__()
__eq__ = _make_bool_operator('==', '!=', operator.__eq__)
__ne__ = _make_bool_operator('!=', '==', operator.__ne__)
__lt__ = _make_bool_operator('<', '>=', operator.__lt__)
__le__ = _make_bool_operator('<=', '>', operator.__le__)
__gt__ = _make_bool_operator('>', '<=', operator.__gt__)
__ge__ = _make_bool_operator('>=', '<', operator.__ge__)
# Please refer the Python documentation to know priority of operators.
# https://docs.python.org/3/reference/expressions.html
__add__ = _make_bin_operator('+', 4, operator.__add__)
__radd__ = _flip(__add__)
__sub__ = _make_bin_operator('-', 4, operator.__sub__)
__rsub__ = _flip(__sub__)
__mul__ = _make_bin_operator('*', 5, operator.__mul__)
__rmul__ = _flip(__mul__)
if sys.version_info < (3, 0, 0):
__div__ = _make_bin_operator('/', 5, operator.__div__) # type: ignore # NOQA
__rdiv__ = _flip(__div__)
else:
__truediv__ = _make_bin_operator('/', 5, operator.__truediv__)
__rtruediv__ = _flip(__truediv__)
__floordiv__ = _make_bin_operator('//', 5, operator.__floordiv__)
__rfloordiv__ = _flip(__floordiv__)
__mod__ = _make_bin_operator('%', 5, operator.__mod__)
__rmod__ = _flip(__mod__)
# Only '**' operator is right-associative
__pow__ = _make_bin_operator('**', 7, operator.__mod__,
right_associative=True)
__lshift__ = _make_bin_operator('<<', 3, operator.__lshift__)
__rlshift__ = _flip(__lshift__)
__rshift__ = _make_bin_operator('>>', 3, operator.__rshift__)
__rrshift__ = _flip(__rshift__)
__and__ = _make_bin_operator('&', 2, operator.__and__)
__rand__ = _flip(__and__)
__xor__ = _make_bin_operator('^', 1, operator.__xor__)
__rxor__ = _flip(__xor__)
__or__ = _make_bin_operator('|', 0, operator.__or__)
__ror__ = _flip(__or__)
__neg__ = _make_un_operator('-', 6, operator.__neg__)
__pos__ = _make_un_operator('+', 6, operator.__pos__)
__invert__ = _make_un_operator('~', 6, operator.__invert__)
def _eval_expr(v):
if isinstance(v, Expr):
return v.eval()
elif isinstance(v, list):
return list(map(_eval_expr, v))
elif isinstance(v, tuple):
return tuple(map(_eval_expr, v))
else:
return v
def _repr(v):
if isinstance(v, Expr):
return str(v)
elif isinstance(v, list):
return '[{0}]'.format(', '.join(map(_repr, v)))
elif isinstance(v, tuple):
if len(v) == 0:
return '()'
elif len(v) == 1:
return '({0},)'.format(_repr(v[0]))
else:
return '({0})'.format(', '.join(map(_repr, v)))
else:
return repr(v)
class Atom(Expr):
def __init__(self):
super(Atom, self).__init__(8)
class Constant(Atom):
def __init__(self, value):
super(Constant, self).__init__()
self.value = value
def __str__(self):
return _repr(self.value)
def eval(self):
return self.value
class Variable(Atom):
def __init__(self, value, name):
super(Variable, self).__init__()
self.value = value
self.name = name
def __str__(self):
return self.name
def eval(self):
return self.value
class GetAttr(Atom):
def __init__(self, obj, name):
super(GetAttr, self).__init__()
self.obj = obj
self.name = name
def __str__(self):
if isinstance(self.name, str):
return '{0}.{1}'.format(_repr(self.obj), self.name)
elif (isinstance(self.name, Constant) and
isinstance(self.name.value, str)):
return '{0}.{1}'.format(_repr(self.obj), self.name.value)
else:
return 'getattr({0}, {1})'.format(_repr(self.obj),
_repr(self.name))
def eval(self):
return getattr(_eval_expr(self.obj), _eval_expr(self.name))
def _str_subscript(exp):
if exp is Ellipsis:
return '...'
elif isinstance(exp, slice):
def key_str(v):
return '' if v is None else _repr(v)
if exp.step is None:
return '{0}:{1}'.format(key_str(exp.start),
key_str(exp.stop))
else:
return '{0}:{1}:{2}'.format(key_str(exp.start),
key_str(exp.stop),
key_str(exp.step))
elif isinstance(exp, tuple):
return ', '.join(map(_str_subscript, exp))
else:
return _repr(exp)
class GetItem(Atom):
def __init__(self, obj, key):
super(GetItem, self).__init__()
self.obj = obj
self.key = key
def __str__(self):
key = _str_subscript(self.key)
return '{0}[{1}]'.format(_repr(self.obj), key)
def eval(self):
return _eval_expr(self.obj)[_eval_expr(self.key)]
class Call(Atom):
def __init__(self, obj, args):
assert isinstance(args, tuple)
super(Call, self).__init__()
self.obj = obj
self.args = args
def __str__(self):
return '{0}({1})'.format(_repr(self.obj),
', '.join(map(_repr, self.args)))
def eval(self):
args = map(_eval_expr, self.args)
func = _eval_expr(self.obj)
return func(*args)
class UnaryOperator(Expr):
def __init__(self, priority, term, exp, func):
super(UnaryOperator, self).__init__(priority)
self.term = term
self.exp = exp
self.func = func
def eval(self):
return self.func(_eval_expr(self.term))
def __str__(self):
exp = _repr(self.term)
if isinstance(self.term, Expr) and self.term.priority < self.priority:
exp = '(' + exp + ')'
return self.exp + exp
class BinaryOperator(Expr):
def __init__(self, priority, lhs, rhs, exp, func, right_associative=False):
super(BinaryOperator, self).__init__(priority)
self.lhs = lhs
self.rhs = rhs
self.exp = exp
self.func = func
self.right_associative = right_associative
def eval(self):
left = self._eval_left()
right = self._eval_right()
return self.func(left, right)
def _eval_left(self):
return _eval_expr(self.lhs)
def _eval_right(self):
return _eval_expr(self.rhs)
def __str__(self):
# When an infix operator is left-associative, we need to append parens
# when rhs has the same priority
# e.g. x << (y << z) != x << y << z
left = _repr(self.lhs)
if isinstance(self.lhs, Expr) and (
self.priority > self.lhs.priority or
(self.right_associative and
self.priority == self.lhs.priority)):
left = '(' + left + ')'
right = _repr(self.rhs)
if isinstance(self.rhs, Expr) and (
self.priority > self.rhs.priority or
(not self.right_associative and
self.priority == self.rhs.priority)):
right = '(' + right + ')'
return '{0} {2} {1}'.format(left, right, self.exp)
class Testable(object):
def expect(self):
raise NotImplementedError()
class BoolBinaryOperator(BinaryOperator, Testable):
def __init__(self, lhs, rhs, exp, inv, func):
BinaryOperator.__init__(self, -1, lhs, rhs, exp, func)
self.inv = inv
def expect(self):
left = self._eval_left()
right = self._eval_right()
if not self.func(left, right):
raise InvalidType(
'{0} {1} {2}'.format(self.lhs, self.exp, self.rhs),
'{0} {1} {2}'.format(left, self.inv, right))
class InvalidType(Exception):
"""Raised when types of data for forward/backward are invalid.
"""
def __init__(self, expect, actual, msg=None):
if msg is None:
msg = 'Expect: {0}\nActual: {1}'.format(expect, actual)
if (hasattr(_thread_local, 'current_function')
and _thread_local.current_function is not None):
msg = '''
Invalid operation is performed in: {0} (Forward)
{1}'''.format(_thread_local.current_function.label, msg)
super(InvalidType, self).__init__(msg)
self.expect = expect
self.actual = actual
def __reduce__(self):
msg, = self.args
return (InvalidType, (self.expect, self.actual, msg))
def _argname(in_types, names):
"""Assigns user friendly names for the input types.
This function also asserts that lengths of in_types and names are the
same.
Args:
in_types (tuple of TypeInfoTuple): Tuple of type information to assign
name to.
names (tuple of str): Human-readable names of ``in_types``.
"""
if len(in_types) != len(names):
raise InvalidType(
'{} argument(s)'.format(str(len(names))),
'{} argument(s)'.format(str(len(in_types))),
'Invalid number of arguments')
for in_type, name in zip(in_types, names):
if isinstance(in_type, Variable):
in_type.name = name
def expect(*bool_exprs):
"""Evaluates and tests all given expressions.
This function evaluates given boolean expressions in order. When at least
one expression is evaluated as ``False``, that means the given condition is
not satisfied.
You can check conditions with this function.
Args:
bool_exprs (tuple of Bool expressions): Bool expressions you want to
evaluate.
"""
if in_light_mode():
if not all(bool_exprs):
raise InvalidType('', '')
else:
for expr in bool_exprs:
assert isinstance(expr, Testable)
expr.expect()
def same_types(*arrays):
for x in arrays:
if not isinstance(x, numpy.ndarray):
break
else:
return True
for x in arrays:
if not isinstance(x, cuda.ndarray):
return False
return True
def eval(exp):
if in_light_mode():
return exp
else:
return exp.eval()
def make_variable(value, name):
if in_light_mode():
return value
else:
return Variable(value, name)
def _make_variable_from_array(array, name):
if not isinstance(array, chainer.get_array_types()):
raise InvalidType(
'isinstance({}, ndarray)'.format(name),
'type({}) == {}'.format(name, type(array)),
)
if in_light_mode():
return array
else:
return Variable(TypeInfo(array.shape, array.dtype), name)
class LightMode(object):
def __enter__(self):
_thread_local.light_mode = True
def __exit__(self, exc_type, exc_value, traceback):
_thread_local.light_mode = False
def _prod_impl(xs):
result = 1
for x in xs:
result *= x
return result
_prod = Variable(_prod_impl, 'prod')
light_mode = LightMode()
def in_light_mode():
try:
return _thread_local.light_mode
except AttributeError:
_thread_local.light_mode = False
return False
def prod(xs):
if in_light_mode():
return _prod_impl(xs)
else:
return _prod(xs)
def expect_broadcast_shapes(*shape_types):
"""Checks if shapes can be broadcasted together.
Args:
shapes_types: Type-checked shapes of the arrays to broadcast.
"""
shapes = [eval(s) for s in shape_types]
error = None
try:
# simulate the shape calculation using zero-sized arrays
numpy.broadcast(*[numpy.empty(s + (0,)) for s in shapes])
except ValueError:
msgs = ['cannot broadcast inputs of the following shapes:']
for shape_type, shape in six.moves.zip(shape_types, shapes):
msgs.append('{} = {}'.format(shape_type, shape))
error = InvalidType('', '', msg='\n'.join(msgs))
if error is not None:
raise error
