def __init__(self, valid=None, invalid=None, valid_equivalent=None):
'''
Check if variables can be expressed without using variables in invalid.
init_valid_equivalent provides a dictionary mapping some invalid
variables to valid ones that can be used instead.
'''
if valid is None:
valid = []
if invalid is None:
invalid = []
if valid_equivalent is None:
valid_equivalent = OrderedDict()
# Nodes that are valid to have in the graph computing outputs
self.valid = set(valid)
# Nodes that are NOT valid to have in the graph computing outputs
self.invalid = set(invalid)
# Mapping from invalid variables to equivalent valid ones.
self.valid_equivalent = valid_equivalent.copy()
self.valid.update(valid_equivalent.values())
self.invalid.update(valid_equivalent.keys())
def __init__(self, valid=None, invalid=None, valid_equivalent=None):
'''
Check if variables can be expressed without using variables in invalid.
init_valid_equivalent provides a dictionary mapping some invalid
variables to valid ones that can be used instead.
'''
if valid is None:
valid = []
if invalid is None:
invalid = []
if valid_equivalent is None:
valid_equivalent = OrderedDict()
# Nodes that are valid to have in the graph computing outputs
self.valid = set(valid)
# Nodes that are NOT valid to have in the graph computing outputs
self.invalid = set(invalid)
# Mapping from invalid variables to equivalent valid ones.
self.valid_equivalent = valid_equivalent.copy()
self.valid.update(valid_equivalent.values())
self.invalid.update(valid_equivalent.keys())
def __init__(self, valid=None, invalid=None, valid_equivalent=None):
if valid is None:
valid = []
if invalid is None:
invalid = []
if valid_equivalent is None:
valid_equivalent = OrderedDict()
# Nodes that are valid to have in the graph computing outputs
self.valid = set(valid)
# Nodes that are NOT valid to have in the graph computing outputs
self.invalid = set(invalid)
# Mapping from invalid variables to equivalent valid ones.
self.valid_equivalent = valid_equivalent.copy()
self.valid.update(list(valid_equivalent.values()))
self.invalid.update(list(valid_equivalent.keys()))
def __init__(self, valid=None, invalid=None, valid_equivalent=None):
if valid is None:
valid = []
if invalid is None:
invalid = []
if valid_equivalent is None:
valid_equivalent = OrderedDict()
# Nodes that are valid to have in the graph computing outputs
self.valid = set(valid)
# Nodes that are NOT valid to have in the graph computing outputs
self.invalid = set(invalid)
# Mapping from invalid variables to equivalent valid ones.
self.valid_equivalent = valid_equivalent.copy()
self.valid.update(list(valid_equivalent.values()))
self.invalid.update(list(valid_equivalent.keys()))
class Rebroadcast(gof.Op):
"""
Change the input's broadcastable fields in some predetermined way.
See Also
--------
unbroadcast <theano.tensor.unbroadcast>
addbroadcast <theano.tensor.addbroadcast>
patternbroadcast <theano.tensor.patternbroadcast>
Notes
-----
Works inplace and works for CudaNdarrayType.
Example
-------
`Rebroadcast((0, True), (1, False))(x)` would make `x` broadcastable in
axis 0 and not broadcastable in axis 1.
"""
view_map = {0: [0]}
_f16_ok = True
# Mapping from Type to C code (and version) to use.
# In the C code, the name of the input variable is %(iname)s,
# the output variable is %(oname)s.
c_code_and_version = {}
check_input = False
__props__ = ("axis", )
def __init__(self, *axis):
# Sort them to make sure we merge all possible case.
items = sorted(axis)
self.axis = OrderedDict(items)
for axis, broad in iteritems(self.axis):
assert isinstance(
axis,
(numpy.integer, int)), ("Rebroadcast needs integer axes. Got ",
axis)
assert isinstance(broad, bool), (
"Rebroadcast needs bool for new broadcast pattern. Got ",
broad)
def __hash__(self):
# Need special __hash__ as dict aren't hashable.
# no ambiguity because each item key is unique
items = sorted(iteritems(self.axis))
return hash((type(self), tuple(items)))
def __str__(self):
if len(self.axis) == 0:
broadcast_pattern = []
else:
broadcast_pattern = [
'?' for i in xrange(1 + max(self.axis.keys()))
]
for k, v in iteritems(self.axis):
broadcast_pattern[k] = str(int(v))
return '%s{%s}' % (self.__class__.__name__,
','.join(broadcast_pattern))
def make_node(self, x):
if self.axis.keys() and (x.ndim <= max(self.axis.keys())):
raise ValueError('Trying to rebroadcast non-existent dimension')
t = x.type.clone(broadcastable=[
self.axis.get(i, b) for i, b in enumerate(x.type.broadcastable)
])
return gof.Apply(self, [x], [t()])
def perform(self, node, inp, out_):
x, = inp
out, = out_
for axis, value in iteritems(self.axis):
if value and x.shape[axis] != 1:
raise ValueError('Dimension %s in Rebroadcast\'s input was'
' supposed to be 1 (got %s instead)' %
(axis, x.shape[axis]))
out[0] = x
def grad(self, inp, grads):
x, = inp
gz, = grads
# restore the broadcasting pattern of the input
return Rebroadcast(*[(axis, x.type.broadcastable[axis])
for axis, value in iteritems(self.axis)])(gz),
def infer_shape(self, node, ishapes):
assert len(ishapes) == 1
l = []
one = theano.tensor.basic.constant(1)
for ax in xrange(len(ishapes[0])):
if self.axis.get(ax, False):
l.append(one)
else:
l.append(ishapes[0][ax])
return [tuple(l)]
def R_op(self, inputs, eval_points):
if eval_points[0] is None:
return [None]
return self(*eval_points, **dict(return_list=True))
def c_code(self, node, nodename, inp, out, sub):
iname, = inp
oname, = out
fail = sub['fail']
itype = node.inputs[0].type.__class__
if itype in self.c_code_and_version:
code, version = self.c_code_and_version[itype]
final_code = ""
for axis, value in iteritems(self.axis):
if value:
final_code += code % locals()
return final_code + """
Py_XDECREF(%(oname)s);
%(oname)s = %(iname)s;
Py_XINCREF(%(oname)s);
""" % locals()
return super(Rebroadcast, self).c_code(node, nodename, inp, out, sub)
def c_code_cache_version(self):
version = []
# If any of the c code is unversionned, we have to return ()
# Else, we will return a list of (type name, version) pairs.
for t, (c, v) in sorted(iteritems(self.c_code_and_version),
key=lambda pair: str(pair[0])):
if not v:
warnings.warn("Type %s has C code for Rebroadcast, but it "
"has no version. You should add a 'version' "
"keyword arg when calling "
"register_rebroadcast_c_code." % t,
stacklevel=2)
return ()
version.append((str(t), v))
if version:
version.append(1)
return tuple(version)
class Rebroadcast(gof.Op):
"""
Change the input's broadcastable fields in some predetermined way.
See Also
--------
unbroadcast <theano.tensor.unbroadcast>
addbroadcast <theano.tensor.addbroadcast>
patternbroadcast <theano.tensor.patternbroadcast>
Notes
-----
Works inplace and works for CudaNdarrayType.
Example
-------
`Rebroadcast((0, True), (1, False))(x)` would make `x` broadcastable in
axis 0 and not broadcastable in axis 1.
"""
view_map = {0: [0]}
_f16_ok = True
# Mapping from Type to C code (and version) to use.
# In the C code, the name of the input variable is %(iname)s,
# the output variable is %(oname)s.
c_code_and_version = {}
check_input = False
__props__ = ("axis",)
def __init__(self, *axis):
# Sort them to make sure we merge all possible case.
items = sorted(axis)
self.axis = OrderedDict(items)
for axis, broad in iteritems(self.axis):
if not isinstance(axis, (numpy.integer, integer_types)):
raise TypeError("Rebroadcast needs integer axes. "
"Got {}".format(axis))
if not isinstance(broad, (numpy.bool_, bool)):
raise TypeError("Rebroadcast needs bool for new broadcast "
"pattern. Got {}".format(broad))
def __hash__(self):
# Need special __hash__ as dict aren't hashable.
# no ambiguity because each item key is unique
items = sorted(iteritems(self.axis))
return hash((type(self), tuple(items)))
def __str__(self):
if len(self.axis) == 0:
broadcast_pattern = []
else:
broadcast_pattern = ['?' for i
in xrange(1 + max(self.axis.keys()))]
for k, v in iteritems(self.axis):
broadcast_pattern[k] = str(int(v))
return '%s{%s}' % (self.__class__.__name__,
','.join(broadcast_pattern))
def make_node(self, x):
if self.axis.keys() and (x.ndim <= max(self.axis.keys())):
raise ValueError('Trying to rebroadcast non-existent dimension')
t = x.type.clone(
broadcastable=[self.axis.get(i, b)
for i, b in enumerate(x.type.broadcastable)])
return gof.Apply(self, [x], [t()])
def perform(self, node, inp, out_):
x, = inp
out, = out_
for axis, value in iteritems(self.axis):
if value and x.shape[axis] != 1:
raise ValueError('Dimension %s in Rebroadcast\'s input was'
' supposed to be 1 (got %s instead)' %
(axis, x.shape[axis]))
out[0] = x
def grad(self, inp, grads):
x, = inp
gz, = grads
# restore the broadcasting pattern of the input
return Rebroadcast(*[(axis, x.type.broadcastable[axis])
for axis, value in iteritems(self.axis)])(gz),
def infer_shape(self, node, ishapes):
assert len(ishapes) == 1
l = []
one = theano.tensor.basic.constant(1)
for ax in xrange(len(ishapes[0])):
if self.axis.get(ax, False):
l.append(one)
else:
l.append(ishapes[0][ax])
return [tuple(l)]
def R_op(self, inputs, eval_points):
if eval_points[0] is None:
return [None]
return self(*eval_points, **dict(return_list=True))
def c_code(self, node, nodename, inp, out, sub):
iname, = inp
oname, = out
fail = sub['fail']
itype = node.inputs[0].type.__class__
if itype in self.c_code_and_version:
code, version = self.c_code_and_version[itype]
final_code = ""
for axis, value in iteritems(self.axis):
if value:
final_code += code % locals()
return final_code + """
Py_XDECREF(%(oname)s);
%(oname)s = %(iname)s;
Py_XINCREF(%(oname)s);
""" % locals()
return super(Rebroadcast, self).c_code(node, nodename, inp, out, sub)
def c_code_cache_version(self):
version = []
# If any of the c code is unversionned, we have to return ()
# Else, we will return a list of (type name, version) pairs.
for t, (c, v) in sorted(iteritems(self.c_code_and_version),
key=lambda pair: str(pair[0])):
if not v:
warnings.warn("Type %s has C code for Rebroadcast, but it "
"has no version. You should add a 'version' "
"keyword arg when calling "
"register_rebroadcast_c_code." % t,
stacklevel=2)
return ()
version.append((str(t), v))
if version:
version.append(1)
return tuple(version)
