Skip to content

P
pytensor
提交 bd11e130 authored 作者: Frédéric Bastien's avatar Frédéric Bastien
浏览文件
操作

Merge pull request #3074 from harlouci/flake8_v2

flake8
上级 cb08bc11 fc6d2310
隐藏空白字符变更
内嵌 并排
正在显示 包含 557 行增加 和 573 行删除
theano/tensor/basic.py
浏览文件 @ bd11e130
"""A `Type` and `Op` classes to work with numpy.ndarrays symbolically.""" """A `Type` and `Op` classes to work with numpy.ndarrays symbolically."""
__docformat__ = "restructuredtext en"
import sys import sys
import warnings import warnings
...@@ -29,7 +27,6 @@ from theano.printing import pprint, min_informative_str ...@@ -29,7 +27,6 @@ from theano.printing import pprint, min_informative_str
# For history # For history
from theano.compile import Rebroadcast, Shape, shape from theano.compile import Rebroadcast, Shape, shape
# We use these exceptions as well. # We use these exceptions as well.
import theano.scalar.sharedvar import theano.scalar.sharedvar
from theano.gradient import grad_undefined from theano.gradient import grad_undefined
...@@ -42,6 +39,8 @@ from theano.tensor.elemwise import Elemwise, DimShuffle, CAReduce, Sum ...@@ -42,6 +39,8 @@ from theano.tensor.elemwise import Elemwise, DimShuffle, CAReduce, Sum
import logging import logging
_logger = logging.getLogger("theano.tensor.basic") _logger = logging.getLogger("theano.tensor.basic")
__docformat__ = "restructuredtext en"
# This is needed as we will hide it later # This is needed as we will hide it later
python_complex = complex python_complex = complex
python_any = any python_any = any
...@@ -620,8 +619,8 @@ def get_scalar_constant_value(orig_v, elemwise=True, ...@@ -620,8 +619,8 @@ def get_scalar_constant_value(orig_v, elemwise=True,
ret = [[None]] ret = [[None]]
v.owner.op.perform(v.owner, const, ret) v.owner.op.perform(v.owner, const, ret)
return ret[0][0] return ret[0][0]
elif (isinstance(v.owner.op, theano.tensor.subtensor.Subtensor) elif (isinstance(v.owner.op, theano.tensor.subtensor.Subtensor) and
and v.ndim == 0): v.ndim == 0):
if isinstance(v.owner.inputs[0], TensorConstant): if isinstance(v.owner.inputs[0], TensorConstant):
cdata = tuple(v.owner.op.get_constant_idx(v.owner.inputs)) cdata = tuple(v.owner.op.get_constant_idx(v.owner.inputs))
try: try:
...@@ -1090,7 +1089,7 @@ scalar_from_tensor = ScalarFromTensor() ...@@ -1090,7 +1089,7 @@ scalar_from_tensor = ScalarFromTensor()
# to be removed as we get the epydoc routine-documenting thing going # to be removed as we get the epydoc routine-documenting thing going
#-JB 20080924 # -JB 20080924
def _conversion(real_value, name): def _conversion(real_value, name):
__oplist_tag(real_value, 'casting') __oplist_tag(real_value, 'casting')
real_value.__module__ = 'tensor.basic' real_value.__module__ = 'tensor.basic'
...@@ -1235,8 +1234,8 @@ class MaxAndArgmax(Op): ...@@ -1235,8 +1234,8 @@ class MaxAndArgmax(Op):
raise TypeError( raise TypeError(
"MaxAndArgmax needs a constant axis. Got %s" % axis) "MaxAndArgmax needs a constant axis. Got %s" % axis)
else: else:
assert (axis.dtype.startswith("int") assert (axis.dtype.startswith("int") or
or axis.dtype.startswith("uint")) axis.dtype.startswith("uint"))
axis = int(axis.data) axis = int(axis.data)
# we make the axis all positive to make the infer_shape work # we make the axis all positive to make the infer_shape work
# with negative axis # with negative axis
...@@ -1373,13 +1372,13 @@ class MaxAndArgmax(Op): ...@@ -1373,13 +1372,13 @@ class MaxAndArgmax(Op):
# Lebesgue measure, the result may be interpreted as weak gradient. # Lebesgue measure, the result may be interpreted as weak gradient.
# @note: This function should work correctly for L{vector}s. # @note: This function should work correctly for L{vector}s.
# (x, y), (gz, gw) # (x, y), (gz, gw)
# gz*dz/dx + gw*dw/dx, gz*dz/dy + gw*dw/dy # gz*dz/dx + gw*dw/dx, gz*dz/dy + gw*dw/dy
# gMax * dMax/dx + gArgMax * dArgMax/dx, # gMax * dMax/dx + gArgMax * dArgMax/dx,
# gMax * dMax/daxis + gArgMax * dArgMax/daxis # gMax * dMax/daxis + gArgMax * dArgMax/daxis
# g_max has one less dimension than x, so you need to complete # g_max has one less dimension than x, so you need to complete
# g_max to x's shape when axis=0 the broadcasting mechanism # g_max to x's shape when axis=0 the broadcasting mechanism
# does it automatically # does it automatically
x, axis = inp x, axis = inp
g_max, g_max_idx = grads g_max, g_max_idx = grads
...@@ -2078,7 +2077,7 @@ def chi2sf(x, k): ...@@ -2078,7 +2077,7 @@ def chi2sf(x, k):
# numpy.real(float32) return a view on the inputs. # numpy.real(float32) return a view on the inputs.
#@_scal_elemwise_with_nfunc('real', 1, 1) # @_scal_elemwise_with_nfunc('real', 1, 1)
@_scal_elemwise @_scal_elemwise
def real(z): def real(z):
"""Return real component of complex-valued tensor `z`""" """Return real component of complex-valued tensor `z`"""
...@@ -2116,7 +2115,7 @@ def complex_from_polar(abs, angle): ...@@ -2116,7 +2115,7 @@ def complex_from_polar(abs, angle):
# fill, _fill_inplace = _elemwise(scal.second, 'fill', # fill, _fill_inplace = _elemwise(scal.second, 'fill',
#"""fill WRITEME (elemwise)""") # """fill WRITEME (elemwise)""")
@_scal_elemwise @_scal_elemwise
def second(a, b): def second(a, b):
"""Create a matrix by filling the shape of a with b""" """Create a matrix by filling the shape of a with b"""
...@@ -3540,8 +3539,8 @@ class Join(Op): ...@@ -3540,8 +3539,8 @@ class Join(Op):
dtypes = [x.type.dtype for x in as_tensor_variable_args] dtypes = [x.type.dtype for x in as_tensor_variable_args]
out_dtype = scal.upcast(*dtypes) out_dtype = scal.upcast(*dtypes)
output_maker = lambda bcastable: tensor(dtype=out_dtype, def output_maker(bcastable):
broadcastable=bcastable) return tensor(dtype=out_dtype, broadcastable=bcastable)
return self._make_node_internal( return self._make_node_internal(
axis, tensors, as_tensor_variable_args, output_maker) axis, tensors, as_tensor_variable_args, output_maker)
...@@ -4361,8 +4360,7 @@ class Tile(Op): ...@@ -4361,8 +4360,7 @@ class Tile(Op):
def make_node(self, x, reps): def make_node(self, x, reps):
warnings.warn(( warnings.warn((
"Tile op is deprecated, use tile function instead."), "Tile op is deprecated, use tile function instead."), stacklevel=3)
stacklevel=3)
x = as_tensor_variable(x) x = as_tensor_variable(x)
reps = as_tensor_variable(reps) reps = as_tensor_variable(reps)
return gof.Apply(self, [x, reps], [tensor(x.type.dtype, [False] * return gof.Apply(self, [x, reps], [tensor(x.type.dtype, [False] *
...@@ -4427,8 +4425,9 @@ def tile(x, reps, ndim=None): ...@@ -4427,8 +4425,9 @@ def tile(x, reps, ndim=None):
except TypeError: except TypeError:
raise ValueError("reps must be iterable") raise ValueError("reps must be iterable")
if not numpy.all([isinstance(r, integer_types) or if not numpy.all([isinstance(r, integer_types) or
(isinstance(r, TensorVariable) and (isinstance(r, TensorVariable) and
r.dtype in ["int8", "int16", "int32", "int64"]) for r in reps]): r.dtype in ["int8", "int16", "int32", "int64"])
for r in reps]):
raise ValueError("elements of reps must be scalars of integer dtype") raise ValueError("elements of reps must be scalars of integer dtype")
elif len(reps) != x.ndim: elif len(reps) != x.ndim:
raise ValueError("len(reps) != x.ndim not currently supported") raise ValueError("len(reps) != x.ndim not currently supported")
...@@ -4442,10 +4441,10 @@ def tile(x, reps, ndim=None): ...@@ -4442,10 +4441,10 @@ def tile(x, reps, ndim=None):
shape = [x.shape[i] for i in xrange(ndim)] shape = [x.shape[i] for i in xrange(ndim)]
alloc_shape = reps + shape alloc_shape = reps + shape
y = alloc(x, *alloc_shape) y = alloc(x, *alloc_shape)
shuffle_ind = numpy.arange(ndim*2).reshape(2, ndim) shuffle_ind = numpy.arange(ndim * 2).reshape(2, ndim)
shuffle_ind = shuffle_ind.transpose().flatten() shuffle_ind = shuffle_ind.transpose().flatten()
y = y.dimshuffle(*shuffle_ind) y = y.dimshuffle(*shuffle_ind)
new_shapes = [sh*reps[i] for i, sh in enumerate(shape)] new_shapes = [sh * reps[i] for i, sh in enumerate(shape)]
y = y.reshape(new_shapes) y = y.reshape(new_shapes)
return y return y
...@@ -4493,12 +4492,12 @@ class ARange(Op): ...@@ -4493,12 +4492,12 @@ class ARange(Op):
def upcast(var): def upcast(var):
if ('int' in var.dtype and if ('int' in var.dtype and
# We do not want to cast uint64 to int64 as this can # We do not want to cast uint64 to int64 as this can
# loose information. If we upcast uint64 with int64, # loose information. If we upcast uint64 with int64,
# this give float64. This is safer then checking for # this give float64. This is safer then checking for
# uint64 in case we support [u]int128 or other in the # uint64 in case we support [u]int128 or other in the
# future. # future.
scal.upcast(var.dtype, 'int64') == 'int64'): scal.upcast(var.dtype, 'int64') == 'int64'):
return cast(var, 'int64') return cast(var, 'int64')
return var return var
...@@ -4512,8 +4511,8 @@ class ARange(Op): ...@@ -4512,8 +4511,8 @@ class ARange(Op):
else: else:
stop = upcast(stop) stop = upcast(stop)
start = upcast(start) start = upcast(start)
return [(maximum(cast(ceil(cast((stop - start), 'float64') return [(maximum(cast(ceil(cast((stop - start), 'float64') / step),
/ step), 'int64'), 0),)] 'int64'), 0),)]
def perform(self, node, inp, out_): def perform(self, node, inp, out_):
start, stop, step = inp start, stop, step = inp
...@@ -4742,8 +4741,8 @@ class PermuteRowElements(Op): ...@@ -4742,8 +4741,8 @@ class PermuteRowElements(Op):
# the gradient over these axes, but keep the dimension (as # the gradient over these axes, but keep the dimension (as
# broadcastable) # broadcastable)
broadcasted_dims = [dim for dim in xrange(gz.type.ndim) broadcasted_dims = [dim for dim in xrange(gz.type.ndim)
if x.type.broadcastable[dim] if x.type.broadcastable[dim] and
and not gz.type.broadcastable[dim]] not gz.type.broadcastable[dim]]
gx = Sum(axis=broadcasted_dims)(gx) gx = Sum(axis=broadcasted_dims)(gx)
# Sum(...) removed the dimensions in broadcasted_dims, # Sum(...) removed the dimensions in broadcasted_dims,
...@@ -4876,17 +4875,17 @@ class Dot(Op): ...@@ -4876,17 +4875,17 @@ class Dot(Op):
xgrad = gz * y xgrad = gz * y
ygrad = gz * x ygrad = gz * x
#x is vector, y is matrix, grad is vector # x is vector, y is matrix, grad is vector
elif xdim == 1 and ydim == 2: elif xdim == 1 and ydim == 2:
xgrad = dot(gz, y.T) xgrad = dot(gz, y.T)
ygrad = outer(x.T, gz) ygrad = outer(x.T, gz)
#x is matrix, y is vector, grad is vector # x is matrix, y is vector, grad is vector
elif xdim == 2 and ydim == 1: elif xdim == 2 and ydim == 1:
xgrad = outer(gz, y.T) xgrad = outer(gz, y.T)
ygrad = dot(x.T, gz) ygrad = dot(x.T, gz)
#x is matrix, y is matrix, grad is matrix # x is matrix, y is matrix, grad is matrix
elif xdim == ydim == 2: elif xdim == ydim == 2:
xgrad = dot(gz, y.T) xgrad = dot(gz, y.T)
ygrad = dot(x.T, gz) ygrad = dot(x.T, gz)
...@@ -4958,8 +4957,8 @@ class Dot(Op): ...@@ -4958,8 +4957,8 @@ class Dot(Op):
if eval_point_values[i] is not None and \ if eval_point_values[i] is not None and \
input_values[i].shape != eval_point_values[i].shape: input_values[i].shape != eval_point_values[i].shape:
raise ValueError( raise ValueError(
'input ' + str(i) + ' and eval_point ' + str(i) 'input ' + str(i) + ' and eval_point ' + str(i) +
+ ' to Dot.R_op should have the same shape, but ' ' to Dot.R_op should have the same shape, but '
'their shapes are %s and %s, respectively' % ( 'their shapes are %s and %s, respectively' % (
str(input_values[i].shape), str(input_values[i].shape),
str(eval_point_values[i].shape))) str(eval_point_values[i].shape)))
...@@ -5230,8 +5229,8 @@ def tensordot(a, b, axes=2): ...@@ -5230,8 +5229,8 @@ def tensordot(a, b, axes=2):
'equal to b.ndim (b.ndim=%i, max(axes[1])=%i).' % 'equal to b.ndim (b.ndim=%i, max(axes[1])=%i).' %
(b.ndim, numpy.max(numpy.array(b_axes)))) (b.ndim, numpy.max(numpy.array(b_axes))))
a_order = (tuple(x for x in tuple(xrange(a.ndim)) if x not in a_axes) a_order = (tuple(x for x in tuple(xrange(a.ndim)) if x not in a_axes) +
+ a_axes) a_axes)
b_order = (b_axes + tuple(x b_order = (b_axes + tuple(x
for x in tuple(xrange(b.ndim)) for x in tuple(xrange(b.ndim))
if x not in b_axes)) if x not in b_axes))
...@@ -5528,8 +5527,8 @@ class Choose(Op): ...@@ -5528,8 +5527,8 @@ class Choose(Op):
# dimensions for the output # dimensions for the output
l = [] l = []
for sh1, sh2, b1 in zip(shapes[0], for sh1, sh2, b1 in zip(shapes[0],
shapes[1][1:], shapes[1][1:],
node.inputs[0].broadcastable): node.inputs[0].broadcastable):
if b1: if b1:
l.append(sh2) l.append(sh2)
else: else:
...@@ -5635,7 +5634,7 @@ class AllocEmpty(gof.Op): ...@@ -5635,7 +5634,7 @@ class AllocEmpty(gof.Op):
out[0] = numpy.empty(sh, dtype=self.dtype) out[0] = numpy.empty(sh, dtype=self.dtype)
def c_code(self, node, name, inputs, out_, sub): def c_code(self, node, name, inputs, out_, sub):
dtype = "NPY_"+self.dtype.upper() dtype = "NPY_" + self.dtype.upper()
out, = out_ out, = out_
fail = sub['fail'] fail = sub['fail']
shps = inputs shps = inputs
......
theano/tensor/blas.py
浏览文件 @ bd11e130
...@@ -266,7 +266,7 @@ SOMEPATH/Canopy_64bit/User/lib/python2.7/site-packages/numpy/distutils/system_in ...@@ -266,7 +266,7 @@ SOMEPATH/Canopy_64bit/User/lib/python2.7/site-packages/numpy/distutils/system_in
# Using "conda install mkl" will install both, as well as # Using "conda install mkl" will install both, as well as
# optimized versions of numpy and scipy. # optimized versions of numpy and scipy.
try: try:
import mkl import mkl #noqa
except ImportError as e: except ImportError as e:
_logger.info('Conda mkl is not available: %s', e) _logger.info('Conda mkl is not available: %s', e)
else: else:
...@@ -1599,11 +1599,11 @@ class GemmOptimizer(Optimizer): ...@@ -1599,11 +1599,11 @@ class GemmOptimizer(Optimizer):
) )
did_something = True did_something = True
nb_replacement += 1 nb_replacement += 1
except InconsistencyError as e: except InconsistencyError:
# TODO: retry other applications of gemm (see comment # TODO: retry other applications of gemm (see comment
# in _gemm_from_node) # in _gemm_from_node)
nb_inconsistency_replace += 1 nb_inconsistency_replace += 1
except ReplacementDidntRemovedError as e: except ReplacementDidntRemovedError:
nb_replacement_didn_t_remove += 1 nb_replacement_didn_t_remove += 1
self.warned = True self.warned = True
fgraph.remove_feature(u) fgraph.remove_feature(u)
......
theano/tensor/elemwise.py
浏览文件 @ bd11e130
...@@ -11,7 +11,7 @@ from six import iteritems ...@@ -11,7 +11,7 @@ from six import iteritems
from six.moves import xrange from six.moves import xrange
from theano.gof import Apply, Op, OpenMPOp from theano.gof import Apply, Op, OpenMPOp
from theano import scalar from theano import scalar
from theano.scalar import Scalar, get_scalar_type from theano.scalar import get_scalar_type
from theano.printing import pprint from theano.printing import pprint
from theano.tensor.utils import hash_from_dict from theano.tensor.utils import hash_from_dict
from theano.gradient import DisconnectedType from theano.gradient import DisconnectedType
...@@ -50,7 +50,7 @@ def TensorConstant(*inputs, **kwargs): ...@@ -50,7 +50,7 @@ def TensorConstant(*inputs, **kwargs):
################## ##################
### DimShuffle ### # DimShuffle #
################## ##################
class DimShuffle(Op): class DimShuffle(Op):
...@@ -139,8 +139,8 @@ class DimShuffle(Op): ...@@ -139,8 +139,8 @@ class DimShuffle(Op):
raise TypeError("DimShuffle indices must be python ints.") raise TypeError("DimShuffle indices must be python ints.")
if j >= len(input_broadcastable): if j >= len(input_broadcastable):
raise ValueError(("new_order[%d] is %d, but the input " raise ValueError(("new_order[%d] is %d, but the input "
"only has %d axes.") % "only has %d axes.") %
(i, j, len(input_broadcastable))) (i, j, len(input_broadcastable)))
if j in new_order[(i + 1):]: if j in new_order[(i + 1):]:
raise ValueError("The same input dimension may not appear " raise ValueError("The same input dimension may not appear "
"twice in the list of output dimensions", "twice in the list of output dimensions",
...@@ -207,7 +207,7 @@ class DimShuffle(Op): ...@@ -207,7 +207,7 @@ class DimShuffle(Op):
ob.append(ib[value]) ob.append(ib[value])
output = TensorType(dtype=input.type.dtype, output = TensorType(dtype=input.type.dtype,
broadcastable=ob).make_variable() broadcastable=ob).make_variable()
return Apply(self, [input], [output]) return Apply(self, [input], [output])
...@@ -219,12 +219,11 @@ class DimShuffle(Op): ...@@ -219,12 +219,11 @@ class DimShuffle(Op):
and self.input_broadcastable == other.input_broadcastable and self.input_broadcastable == other.input_broadcastable
def _rehash(self): def _rehash(self):
self._hashval = ( self._hashval = (hash(type(self).__name__) ^
hash(type(self).__name__) hash(type(self).__module__) ^
^ hash(type(self).__module__) hash(self.inplace) ^
^ hash(self.inplace) hash(self.new_order) ^
^ hash(self.new_order) hash(self.input_broadcastable))
^ hash(self.input_broadcastable))
def __hash__(self): def __hash__(self):
return self._hashval return self._hashval
...@@ -232,7 +231,7 @@ class DimShuffle(Op): ...@@ -232,7 +231,7 @@ class DimShuffle(Op):
def __str__(self): def __str__(self):
if self.inplace: if self.inplace:
return "InplaceDimShuffle{%s}" % ",".join(str(x) return "InplaceDimShuffle{%s}" % ",".join(str(x)
for x in self.new_order) for x in self.new_order)
else: else:
return "DimShuffle{%s}" % ",".join(str(x) for x in self.new_order) return "DimShuffle{%s}" % ",".join(str(x) for x in self.new_order)
...@@ -286,7 +285,8 @@ class DimShuffle(Op): ...@@ -286,7 +285,8 @@ class DimShuffle(Op):
nd_out = len(self.new_order) nd_out = len(self.new_order)
check_input_nd = [('if (PyArray_NDIM(%(input)s) != ' + str(nd_in) + ')' check_input_nd = [('if (PyArray_NDIM(%(input)s) != ' + str(nd_in) + ')'
'{PyErr_SetString(PyExc_NotImplementedError, "input nd"); %(fail)s;}')] '{PyErr_SetString(PyExc_NotImplementedError, '
'"input nd"); %(fail)s;}')]
clear_output = ['if (%(res)s) {Py_XDECREF(%(res)s);}'] clear_output = ['if (%(res)s) {Py_XDECREF(%(res)s);}']
...@@ -296,8 +296,10 @@ class DimShuffle(Op): ...@@ -296,8 +296,10 @@ class DimShuffle(Op):
get_base = [ get_base = [
'{ PyArrayObject * %(basename)s = %(input)s', 'Py_INCREF((PyObject*)%(basename)s)'] '{ PyArrayObject * %(basename)s = %(input)s', 'Py_INCREF((PyObject*)%(basename)s)']
else: else:
get_base = [('{ PyArrayObject * %(basename)s = (PyArrayObject*)PyArray_FromAny((PyObject*)%(input)s, NULL,' get_base = [('{ PyArrayObject * %(basename)s = '
'0, 0, NPY_ARRAY_ALIGNED|NPY_ARRAY_ENSURECOPY, NULL)')] '(PyArrayObject*)PyArray_FromAny((PyObject*)%(input)s,'
' NULL, 0, 0, NPY_ARRAY_ALIGNED|NPY_ARRAY_ENSURECOPY,'
' NULL)')]
shape_statements = ['npy_intp dimensions[%i]' % nd_out] shape_statements = ['npy_intp dimensions[%i]' % nd_out]
for i, o in enumerate(self.new_order): for i, o in enumerate(self.new_order):
...@@ -312,9 +314,12 @@ class DimShuffle(Op): ...@@ -312,9 +314,12 @@ class DimShuffle(Op):
# set the strides of the non-broadcasted dimensions # set the strides of the non-broadcasted dimensions
for i, o in enumerate(self.new_order): for i, o in enumerate(self.new_order):
if o != 'x': if o != 'x':
strides_statements += [('strides[' + str(i) strides_statements += [('strides[' + str(i) +
+ '] = PyArray_DIMS(%(basename)s)[' + str(o) '] = PyArray_DIMS(%(basename)s)[' +
+ '] == 1? 0 : PyArray_STRIDES(%(basename)s)[' + str(o) + ']')] str(o) +
'] == 1? 0 : '
'PyArray_STRIDES(%(basename)s)[' +
str(o) + ']')]
else: else:
strides_statements += [('strides[' + str(i) + '] = 0')] strides_statements += [('strides[' + str(i) + '] = 0')]
...@@ -360,12 +365,12 @@ PyArray_SetBaseObject(%(res)s, (PyObject*)%(basename)s); ...@@ -360,12 +365,12 @@ PyArray_SetBaseObject(%(res)s, (PyObject*)%(basename)s);
""" """
'}'] '}']
full_code = statements(check_input_nd full_code = statements(check_input_nd +
+ clear_output clear_output +
+ get_base get_base +
+ shape_statements shape_statements +
+ strides_statements strides_statements +
+ close_bracket) close_bracket)
if 0: if 0:
print('C_CODE') print('C_CODE')
...@@ -408,7 +413,7 @@ PyArray_SetBaseObject(%(res)s, (PyObject*)%(basename)s); ...@@ -408,7 +413,7 @@ PyArray_SetBaseObject(%(res)s, (PyObject*)%(basename)s);
class DimShufflePrinter: class DimShufflePrinter:
def __p(self, new_order, pstate, r): def __p(self, new_order, pstate, r):
if new_order != () and new_order[0] == 'x': if new_order != () and new_order[0] == 'x':
return "%s" % self.__p(new_order[1:], pstate, r) return "%s" % self.__p(new_order[1:], pstate, r)
# return "[%s]" % self.__p(new_order[1:], pstate, r) # return "[%s]" % self.__p(new_order[1:], pstate, r)
if list(new_order) == list(range(r.type.ndim)): if list(new_order) == list(range(r.type.ndim)):
...@@ -416,7 +421,7 @@ class DimShufflePrinter: ...@@ -416,7 +421,7 @@ class DimShufflePrinter:
if list(new_order) == list(reversed(range(r.type.ndim))): if list(new_order) == list(reversed(range(r.type.ndim))):
return "%s.T" % pstate.pprinter.process(r) return "%s.T" % pstate.pprinter.process(r)
return "DimShuffle{%s}(%s)" % (", ".join(map(str, new_order)), return "DimShuffle{%s}(%s)" % (", ".join(map(str, new_order)),
pstate.pprinter.process(r)) pstate.pprinter.process(r))
def process(self, r, pstate): def process(self, r, pstate):
if r.owner is None: if r.owner is None:
...@@ -428,11 +433,11 @@ class DimShufflePrinter: ...@@ -428,11 +433,11 @@ class DimShufflePrinter:
raise TypeError("Can only print DimShuffle.") raise TypeError("Can only print DimShuffle.")
pprint.assign(lambda pstate, r: r.owner and isinstance(r.owner.op, DimShuffle), pprint.assign(lambda pstate, r: r.owner and isinstance(r.owner.op, DimShuffle),
DimShufflePrinter()) DimShufflePrinter())
################ ################
### Elemwise ### # Elemwise #
################ ################
class Elemwise(OpenMPOp): class Elemwise(OpenMPOp):
...@@ -496,7 +501,7 @@ class Elemwise(OpenMPOp): ...@@ -496,7 +501,7 @@ class Elemwise(OpenMPOp):
self.nfunc = getattr(numpy, nfunc_spec[0]) self.nfunc = getattr(numpy, nfunc_spec[0])
elif scalar_op.nin > 0: elif scalar_op.nin > 0:
self.ufunc = numpy.frompyfunc(scalar_op.impl, scalar_op.nin, self.ufunc = numpy.frompyfunc(scalar_op.impl, scalar_op.nin,
scalar_op.nout) scalar_op.nout)
# precompute the hash of this node # precompute the hash of this node
self._rehash() self._rehash()
...@@ -518,7 +523,8 @@ class Elemwise(OpenMPOp): ...@@ -518,7 +523,8 @@ class Elemwise(OpenMPOp):
self.nfunc = getattr(numpy, self.nfunc_spec[0]) self.nfunc = getattr(numpy, self.nfunc_spec[0])
elif self.scalar_op.nin > 0: elif self.scalar_op.nin > 0:
self.ufunc = numpy.frompyfunc(self.scalar_op.impl, self.ufunc = numpy.frompyfunc(self.scalar_op.impl,
self.scalar_op.nin, self.scalar_op.nout) self.scalar_op.nin,
self.scalar_op.nout)
self._rehash() self._rehash()
def make_node(self, *inputs): def make_node(self, *inputs):
...@@ -557,15 +563,16 @@ class Elemwise(OpenMPOp): ...@@ -557,15 +563,16 @@ class Elemwise(OpenMPOp):
# it is multiplied by nout because Elemwise supports multiple outputs # it is multiplied by nout because Elemwise supports multiple outputs
# (nout of them) # (nout of them)
out_broadcastables = [[all(bcast) out_broadcastables = [[all(bcast)
for bcast in izip(*[input.type.broadcastable for bcast in
for input in inputs])]] * shadow.nout izip(*[input.type.broadcastable
for input in inputs])]] * shadow.nout
# inplace_pattern maps output idx -> input idx # inplace_pattern maps output idx -> input idx
inplace_pattern = self.inplace_pattern inplace_pattern = self.inplace_pattern
if inplace_pattern: if inplace_pattern:
for overwriter, overwritten in iteritems(inplace_pattern): for overwriter, overwritten in iteritems(inplace_pattern):
for ob, ib in izip(out_broadcastables[overwriter], for ob, ib in izip(out_broadcastables[overwriter],
inputs[overwritten].type.broadcastable): inputs[overwritten].type.broadcastable):
if ib and not ob: if ib and not ob:
raise ValueError( raise ValueError(
"Operation cannot be done inplace on an input " "Operation cannot be done inplace on an input "
...@@ -579,8 +586,8 @@ class Elemwise(OpenMPOp): ...@@ -579,8 +586,8 @@ class Elemwise(OpenMPOp):
([i.type.dtype for i in inputs], out_dtypes, inplace_pattern))) ([i.type.dtype for i in inputs], out_dtypes, inplace_pattern)))
outputs = [TensorType(dtype=dtype, broadcastable=broadcastable)() outputs = [TensorType(dtype=dtype, broadcastable=broadcastable)()
for dtype, broadcastable in izip(out_dtypes, out_broadcastables) for dtype, broadcastable in izip(out_dtypes,
] out_broadcastables)]
return Apply(self, inputs, outputs) return Apply(self, inputs, outputs)
def __eq__(self, other): def __eq__(self, other):
...@@ -589,8 +596,8 @@ class Elemwise(OpenMPOp): ...@@ -589,8 +596,8 @@ class Elemwise(OpenMPOp):
other_items = list(other.inplace_pattern.items()) other_items = list(other.inplace_pattern.items())
items.sort() items.sort()
other_items.sort() other_items.sort()
rval = ((self.scalar_op == other.scalar_op) rval = ((self.scalar_op == other.scalar_op) and
and (items == other_items)) (items == other_items))
return rval return rval
return False return False
...@@ -628,7 +635,7 @@ class Elemwise(OpenMPOp): ...@@ -628,7 +635,7 @@ class Elemwise(OpenMPOp):
rop_out = None rop_out = None
for jdx, (inp, eval_point) in enumerate(izip(inputs, for jdx, (inp, eval_point) in enumerate(izip(inputs,
eval_points)): eval_points)):
# if None, then we can just ignore this branch .. # if None, then we can just ignore this branch ..
# what we do is to assume that for any non-differentiable # what we do is to assume that for any non-differentiable
# branch, the gradient is actually 0, which I think is not # branch, the gradient is actually 0, which I think is not
...@@ -668,7 +675,7 @@ class Elemwise(OpenMPOp): ...@@ -668,7 +675,7 @@ class Elemwise(OpenMPOp):
# to the gradient.grad method when the outputs have # to the gradient.grad method when the outputs have
# some integer and some floating point outputs # some integer and some floating point outputs
if False in [str(out.type.dtype).find('int') == -1 if False in [str(out.type.dtype).find('int') == -1
for out in outs]: for out in outs]:
# For integer output, return value may # For integer output, return value may
# only be zero or undefined # only be zero or undefined
# We don't bother with trying to check # We don't bother with trying to check
...@@ -699,7 +706,7 @@ class Elemwise(OpenMPOp): ...@@ -699,7 +706,7 @@ class Elemwise(OpenMPOp):
# we can sum over them # we can sum over them
# todo: only count dimensions that were effectively broadcasted # todo: only count dimensions that were effectively broadcasted
to_sum = [j for j, bcast in enumerate(ipt.type.broadcastable) to_sum = [j for j, bcast in enumerate(ipt.type.broadcastable)
if bcast] if bcast]
if to_sum: if to_sum:
shuffle = [] shuffle = []
...@@ -714,7 +721,7 @@ class Elemwise(OpenMPOp): ...@@ -714,7 +721,7 @@ class Elemwise(OpenMPOp):
# close for # close for
sr = Sum(axis=to_sum)(rval[i]) sr = Sum(axis=to_sum)(rval[i])
sr = sr.dimshuffle(shuffle) sr = sr.dimshuffle(shuffle)
#sr = DimShuffle(sr.type.broadcastable, shuffle)(sr) # sr = DimShuffle(sr.type.broadcastable, shuffle)(sr)
rval[i] = sr rval[i] = sr
# close if # close if
# close for # close for
...@@ -747,7 +754,7 @@ class Elemwise(OpenMPOp): ...@@ -747,7 +754,7 @@ class Elemwise(OpenMPOp):
if not isinstance(scalar_igrads, (list, tuple)): if not isinstance(scalar_igrads, (list, tuple)):
raise TypeError('%s.grad returned %s instead of list or tuple' % raise TypeError('%s.grad returned %s instead of list or tuple' %
(str(self.scalar_op), str(type(scalar_igrads)))) (str(self.scalar_op), str(type(scalar_igrads))))
nd = len(inputs[0].type.broadcastable) # this is the same for everyone nd = len(inputs[0].type.broadcastable) # this is the same for everyone
...@@ -787,9 +794,8 @@ class Elemwise(OpenMPOp): ...@@ -787,9 +794,8 @@ class Elemwise(OpenMPOp):
# should be disabled. # should be disabled.
super(Elemwise, self).perform(node, inputs, output_storage) super(Elemwise, self).perform(node, inputs, output_storage)
maxsize = max(len(input.shape) for input in inputs)
for dims in izip(*[list(zip(input.shape, sinput.type.broadcastable)) for dims in izip(*[list(zip(input.shape, sinput.type.broadcastable))
for input, sinput in zip(inputs, node.inputs)]): for input, sinput in zip(inputs, node.inputs)]):
if max(d for d, b in dims) != 1 and (1, False) in dims: if max(d for d, b in dims) != 1 and (1, False) in dims:
# yes there may be more compact ways to write this code, # yes there may be more compact ways to write this code,
# but please maintain python 2.4 compatibility # but please maintain python 2.4 compatibility
...@@ -1115,7 +1121,7 @@ class Elemwise(OpenMPOp): ...@@ -1115,7 +1121,7 @@ class Elemwise(OpenMPOp):
# use it! The scalar_op need to check the broadcast flag himself. # use it! The scalar_op need to check the broadcast flag himself.
if (all([o.ndim >= 1 for o in node.outputs]) and if (all([o.ndim >= 1 for o in node.outputs]) and
# Don't use the contig code for broadcasted scalar. # Don't use the contig code for broadcasted scalar.
not all(node.outputs[0].broadcastable)): not all(node.outputs[0].broadcastable)):
contig = None contig = None
try: try:
contig = self.scalar_op.c_code_contiguous( contig = self.scalar_op.c_code_contiguous(
...@@ -1192,19 +1198,20 @@ class Elemwise(OpenMPOp): ...@@ -1192,19 +1198,20 @@ class Elemwise(OpenMPOp):
return self.scalar_op.c_support_code() return self.scalar_op.c_support_code()
def c_support_code_apply(self, node, nodename): def c_support_code_apply(self, node, nodename):
support_code = self.scalar_op.c_support_code_apply(node, support_code = self.scalar_op.c_support_code_apply(node, nodename +
nodename + '_scalar_') '_scalar_')
return support_code return support_code
def c_code_cache_version_apply(self, node): def c_code_cache_version_apply(self, node):
version = [12] # the version corresponding to the c code in this Op version = [12] # the version corresponding to the c code in this Op
# now we insert versions for the ops on which we depend... # now we insert versions for the ops on which we depend...
scalar_node = Apply(self.scalar_op, scalar_node = Apply(
[get_scalar_type(dtype=input.type.dtype).make_variable() self.scalar_op,
for input in node.inputs], [get_scalar_type(dtype=input.type.dtype).make_variable()
[get_scalar_type(dtype=output.type.dtype).make_variable() for input in node.inputs],
for output in node.outputs]) [get_scalar_type(dtype=output.type.dtype).make_variable()
for output in node.outputs])
version.append(self.scalar_op.c_code_cache_version_apply(scalar_node)) version.append(self.scalar_op.c_code_cache_version_apply(scalar_node))
for i in node.inputs + node.outputs: for i in node.inputs + node.outputs:
version.append(get_scalar_type(dtype=i.type.dtype).c_code_cache_version()) version.append(get_scalar_type(dtype=i.type.dtype).c_code_cache_version())
...@@ -1233,7 +1240,7 @@ class Elemwise(OpenMPOp): ...@@ -1233,7 +1240,7 @@ class Elemwise(OpenMPOp):
################ ################
### CAReduce ### # CAReduce #
################ ################
class CAReduce(Op): class CAReduce(Op):
...@@ -1325,8 +1332,8 @@ class CAReduce(Op): ...@@ -1325,8 +1332,8 @@ class CAReduce(Op):
if self.axis is not None: if self.axis is not None:
for axis in self.axis: for axis in self.axis:
if (axis >= input.type.ndim if (axis >= input.type.ndim or
or (axis < 0 and abs(axis) > input.type.ndim)): (axis < 0 and abs(axis) > input.type.ndim)):
raise ValueError(( raise ValueError((
'Not enough dimensions on %s to reduce on axis %s' 'Not enough dimensions on %s to reduce on axis %s'
% (input, axis))) % (input, axis)))
...@@ -1366,9 +1373,9 @@ class CAReduce(Op): ...@@ -1366,9 +1373,9 @@ class CAReduce(Op):
self.set_ufunc(self.scalar_op) self.set_ufunc(self.scalar_op)
def __eq__(self, other): def __eq__(self, other):
return (type(self) == type(other) return (type(self) == type(other) and
and self.scalar_op == other.scalar_op self.scalar_op == other.scalar_op and
and self.axis == other.axis) self.axis == other.axis)
def __hash__(self): def __hash__(self):
if self.axis is None: if self.axis is None:
...@@ -1420,13 +1427,13 @@ class CAReduce(Op): ...@@ -1420,13 +1427,13 @@ class CAReduce(Op):
# was built with "frompyfunc". We need to find out if we # was built with "frompyfunc". We need to find out if we
# are in one of these cases (only "object" is supported in # are in one of these cases (only "object" is supported in
# the output). # the output).
if ((self.ufunc.ntypes == 1) if ((self.ufunc.ntypes == 1) and
and (self.ufunc.types[0][-1] == 'O')): (self.ufunc.types[0][-1] == 'O')):
variable = self.ufunc.reduce(variable, dimension, variable = self.ufunc.reduce(variable, dimension,
dtype='object') dtype='object')
else: else:
variable = self.ufunc.reduce(variable, dimension, variable = self.ufunc.reduce(variable, dimension,
dtype=acc_dtype) dtype=acc_dtype)
variable = numpy.asarray(variable) variable = numpy.asarray(variable)
if numpy.may_share_memory(variable, input): if numpy.may_share_memory(variable, input):
...@@ -1434,7 +1441,7 @@ class CAReduce(Op): ...@@ -1434,7 +1441,7 @@ class CAReduce(Op):
# We don't want this. # We don't want this.
variable = variable.copy() variable = variable.copy()
output[0] = theano._asarray(variable, output[0] = theano._asarray(variable,
dtype=node.outputs[0].type.dtype) dtype=node.outputs[0].type.dtype)
else: else:
# Force a copy # Force a copy
output[0] = numpy.array(variable, copy=True, output[0] = numpy.array(variable, copy=True,
...@@ -1568,27 +1575,25 @@ for(int i=0;i<PyArray_NDIM(%(iname)s);i++){ ...@@ -1568,27 +1575,25 @@ for(int i=0;i<PyArray_NDIM(%(iname)s);i++){
""" % locals() """ % locals()
else: else:
raise TypeError( raise TypeError(
"The CAReduce.scalar_op must have an identity field.") "The CAReduce.scalar_op must have an identity field.")
task0_decl = ( task0_decl = ("%(dtype)s& %(name)s_i = *%(name)s_iter;\n"
"%(dtype)s& %(name)s_i = *%(name)s_iter;\n" "%(name)s_i = %(identity)s;"
"%(name)s_i = %(identity)s;" % dict(dtype=adtype, name=aname, identity=identity))
% dict(dtype=adtype, name=aname, identity=identity))
task1_decl = ("%(dtype)s& %(name)s_i = *%(name)s_iter;\n" task1_decl = ("%(dtype)s& %(name)s_i = *%(name)s_iter;\n"
% dict(dtype=idtype, name=inames[0])) % dict(dtype=idtype, name=inames[0]))
task1_code = self.scalar_op.c_code( task1_code = self.scalar_op.c_code(
Apply( Apply(self.scalar_op,
self.scalar_op, [get_scalar_type(dtype=input.type.dtype).make_variable()
[get_scalar_type(dtype=input.type.dtype).make_variable() for input in (node.inputs * 2)],
for input in (node.inputs * 2)], [get_scalar_type(dtype=output.type.dtype).make_variable()
[get_scalar_type(dtype=output.type.dtype).make_variable() for input in node.outputs]),
for input in node.outputs]), None,
None, ["%s_i" % aname, "%s_i" % inames[0]],
["%s_i" % aname, "%s_i" % inames[0]], ["%s_i" % aname],
["%s_i" % aname], sub)
sub)
code1 = """ code1 = """
{ {
%(task1_decl)s %(task1_decl)s
...@@ -1600,11 +1605,10 @@ for(int i=0;i<PyArray_NDIM(%(iname)s);i++){ ...@@ -1600,11 +1605,10 @@ for(int i=0;i<PyArray_NDIM(%(iname)s);i++){
if len(axis) == 1: if len(axis) == 1:
all_code = [("", "")] * nnested + [(task0_decl, code1), ""] all_code = [("", "")] * nnested + [(task0_decl, code1), ""]
else: else:
all_code = ( all_code = ([("", "")] * nnested +
[("", "")] * nnested [(task0_decl, "")] +
+ [(task0_decl, "")] [("", "")] * (len(axis) - 2) +
+ [("", "")] * (len(axis) - 2) [("", code1), ""])
+ [("", code1), ""])
else: else:
all_code = [task0_decl + code1] all_code = [task0_decl + code1]
loop = cgen.make_loop_careduce( loop = cgen.make_loop_careduce(
...@@ -1632,11 +1636,12 @@ for(int i=0;i<PyArray_NDIM(%(iname)s);i++){ ...@@ -1632,11 +1636,12 @@ for(int i=0;i<PyArray_NDIM(%(iname)s);i++){
version = [5] # the version corresponding to the c code in this Op version = [5] # the version corresponding to the c code in this Op
# now we insert versions for the ops on which we depend... # now we insert versions for the ops on which we depend...
scalar_node = Apply(self.scalar_op, scalar_node = Apply(
[get_scalar_type(dtype=input.type.dtype).make_variable() self.scalar_op,
for input in node.inputs], [get_scalar_type(dtype=input.type.dtype).make_variable()
[get_scalar_type(dtype=output.type.dtype).make_variable() for input in node.inputs],
for output in node.outputs]) [get_scalar_type(dtype=output.type.dtype).make_variable()
for output in node.outputs])
version.append(self.scalar_op.c_code_cache_version_apply(scalar_node)) version.append(self.scalar_op.c_code_cache_version_apply(scalar_node))
for i in node.inputs + node.outputs: for i in node.inputs + node.outputs:
version.append(get_scalar_type(dtype=i.type.dtype).c_code_cache_version()) version.append(get_scalar_type(dtype=i.type.dtype).c_code_cache_version())
...@@ -1760,9 +1765,9 @@ class CAReduceDtype(CAReduce): ...@@ -1760,9 +1765,9 @@ class CAReduceDtype(CAReduce):
self.acc_dtype = acc_dtype self.acc_dtype = acc_dtype
def __eq__(self, other): def __eq__(self, other):
return (CAReduce.__eq__(self, other) return (CAReduce.__eq__(self, other) and
and self.dtype == other.dtype self.dtype == other.dtype and
and self.acc_dtype == other.acc_dtype) self.acc_dtype == other.acc_dtype)
def __hash__(self): def __hash__(self):
return CAReduce.__hash__(self) ^ hash((self.dtype, self.acc_dtype)) return CAReduce.__hash__(self) ^ hash((self.dtype, self.acc_dtype))
...@@ -1968,8 +1973,8 @@ class Prod(CAReduceDtype): ...@@ -1968,8 +1973,8 @@ class Prod(CAReduceDtype):
self.no_zeros_in_input = False self.no_zeros_in_input = False
def __eq__(self, other): def __eq__(self, other):
return (CAReduceDtype.__eq__(self, other) return (CAReduceDtype.__eq__(self, other) and
and self.no_zeros_in_input == other.no_zeros_in_input) self.no_zeros_in_input == other.no_zeros_in_input)
def __hash__(self): def __hash__(self):
return (CAReduceDtype.__hash__(self) ^ return (CAReduceDtype.__hash__(self) ^
...@@ -2124,25 +2129,26 @@ class MulWithoutZeros(scalar.BinaryScalarOp): ...@@ -2124,25 +2129,26 @@ class MulWithoutZeros(scalar.BinaryScalarOp):
def c_code(self, node, name, inp, out, sub): def c_code(self, node, name, inp, out, sub):
x, y = inp x, y = inp
z, = out z, = out
return (("%(z)s = ((%(x)s == 0) ? (%(y)s) : " return (("%(z)s = ((%(x)s == 0) ? (%(y)s) : " +
+ "((%(y)s == 0) ? (%(x)s) : ((%(y)s)*(%(x)s))) );") "((%(y)s == 0) ? (%(x)s) : ((%(y)s)*(%(x)s))) );")
% locals()) % locals())
def c_code_cache_version(self): def c_code_cache_version(self):
return (1,) return (1,)
mul_without_zeros = MulWithoutZeros(scalar.upcast_out, mul_without_zeros = MulWithoutZeros(scalar.upcast_out, name='mul_without_zeros')
name='mul_without_zeros')
class ProdWithoutZeros(CAReduceDtype): class ProdWithoutZeros(CAReduceDtype):
def __init__(self, axis=None, dtype=None, acc_dtype=None): def __init__(self, axis=None, dtype=None, acc_dtype=None):
CAReduceDtype.__init__(self, mul_without_zeros, axis=axis, CAReduceDtype.__init__(self, mul_without_zeros, axis=axis,
dtype=dtype, acc_dtype=acc_dtype) dtype=dtype, acc_dtype=acc_dtype)
def grad(self, inp, grads): def grad(self, inp, grads):
a, = inp a, = inp
a_grad = theano.gradient.grad_not_implemented(self, 0, a, a_grad = theano.gradient.grad_not_implemented(
"2nd derivatives of `product(a)` is not currently supported." self, 0, a,
"If `a` is guarenteed to contains no zeros, use `product(a, no_zeros_in_input=True)`." "2nd derivatives of `product(a)` is not currently supported."
) "If `a` is guarenteed to contains no zeros, use "
"`product(a, no_zeros_in_input=True)`.")
return [a_grad] return [a_grad]
theano/tensor/inplace.py
浏览文件 @ bd11e130
...@@ -28,7 +28,6 @@ def _scal_inplace(symbol): ...@@ -28,7 +28,6 @@ def _scal_inplace(symbol):
def chk(pstate, r): def chk(pstate, r):
if not r.owner: if not r.owner:
return False return False
op = r.owner.op
return r.owner.op == rval return r.owner.op == rval
pprint.assign(chk, printing.FunctionPrinter(symbolname.replace('_inplace', '='))) pprint.assign(chk, printing.FunctionPrinter(symbolname.replace('_inplace', '=')))
......
theano/tensor/opt.py
浏览文件 @ bd11e130
...@@ -6,8 +6,6 @@ from __future__ import print_function ...@@ -6,8 +6,6 @@ from __future__ import print_function
# TODO: 0*x -> 0 # TODO: 0*x -> 0
import logging import logging
_logger = logging.getLogger('theano.tensor.opt')
import itertools import itertools
import operator import operator
import sys import sys
...@@ -34,12 +32,10 @@ from theano.tensor.subtensor import (get_idx_list, get_canonical_form_slice, ...@@ -34,12 +32,10 @@ from theano.tensor.subtensor import (get_idx_list, get_canonical_form_slice,
Subtensor, IncSubtensor, make_constant, Subtensor, IncSubtensor, make_constant,
AdvancedIncSubtensor1, AdvancedIncSubtensor1,
AdvancedIncSubtensor, AdvancedIncSubtensor,
AdvancedSubtensor,
AdvancedSubtensor1, AdvancedSubtensor1,
advanced_subtensor, advanced_subtensor,
advanced_subtensor1, advanced_subtensor1,
advanced_inc_subtensor1, advanced_inc_subtensor1)
inc_subtensor)
from theano import scalar from theano import scalar
from theano.scalar import basic from theano.scalar import basic
from theano.tensor import basic as T from theano.tensor import basic as T
...@@ -56,6 +52,8 @@ from theano.gof import toolbox ...@@ -56,6 +52,8 @@ from theano.gof import toolbox
from theano.tensor.basic import get_scalar_constant_value, ShapeError, NotScalarConstantError from theano.tensor.basic import get_scalar_constant_value, ShapeError, NotScalarConstantError
from six import StringIO from six import StringIO
_logger = logging.getLogger('theano.tensor.opt')
theano.configparser.AddConfigVar('on_shape_error', theano.configparser.AddConfigVar('on_shape_error',
"warn: print a warning and use the default" "warn: print a warning and use the default"
" value. raise: raise an error", " value. raise: raise an error",
...@@ -165,23 +163,24 @@ def broadcast_like(value, template, fgraph, dtype=None): ...@@ -165,23 +163,24 @@ def broadcast_like(value, template, fgraph, dtype=None):
# the template may have 1s in its shape without being broadcastable # the template may have 1s in its shape without being broadcastable
if rval.broadcastable != template.broadcastable: if rval.broadcastable != template.broadcastable:
rval = T.unbroadcast(rval, *[i for i in xrange(rval.ndim) rval = T.unbroadcast(rval, *[i for i in xrange(rval.ndim)
if rval.broadcastable[i] if rval.broadcastable[i] and
and not template.broadcastable[i]]) not template.broadcastable[i]])
assert rval.type.dtype == dtype assert rval.type.dtype == dtype
if rval.type.broadcastable != template.broadcastable: if rval.type.broadcastable != template.broadcastable:
raise AssertionError("rval.type.broadcastable is " + raise AssertionError("rval.type.broadcastable is " +
str(rval.type.broadcastable) + str(rval.type.broadcastable) +
" but template.broadcastable is" + " but template.broadcastable is" +
str(template.broadcastable)) str(template.broadcastable))
return rval return rval
theano.configparser.AddConfigVar('tensor.insert_inplace_optimizer_validate_nb', theano.configparser.AddConfigVar(
"-1: auto, if graph have less then 500 nodes 1, else 10", 'tensor.insert_inplace_optimizer_validate_nb',
theano.configparser.IntParam(-1), "-1: auto, if graph have less then 500 nodes 1, else 10",
in_c_key=False) theano.configparser.IntParam(-1),
in_c_key=False)
def inplace_elemwise_optimizer_op(OP): def inplace_elemwise_optimizer_op(OP):
...@@ -251,11 +250,10 @@ def inplace_elemwise_optimizer_op(OP): ...@@ -251,11 +250,10 @@ def inplace_elemwise_optimizer_op(OP):
# target. # target.
# Remove here as faster. # Remove here as faster.
candidate_inputs = [i for i in xrange(len(node.inputs)) candidate_inputs = [i for i in xrange(len(node.inputs))
if i not in baseline.values() \ if i not in baseline.values() and
and not isinstance(node.inputs[i], not isinstance(node.inputs[i], Constant) and
Constant)\ not fgraph.destroyers(node.inputs[i]) and
and not fgraph.destroyers(node.inputs[i])\ node.inputs[i] not in protected_inputs]
and node.inputs[i] not in protected_inputs]
verbose = False verbose = False
...@@ -265,7 +263,7 @@ def inplace_elemwise_optimizer_op(OP): ...@@ -265,7 +263,7 @@ def inplace_elemwise_optimizer_op(OP):
for candidate_input in candidate_inputs: for candidate_input in candidate_inputs:
# remove inputs that don't have the same dtype as the output # remove inputs that don't have the same dtype as the output
if node.inputs[candidate_input].type != node.outputs[ if node.inputs[candidate_input].type != node.outputs[
candidate_output].type: candidate_output].type:
continue continue
inplace_pattern = dict(baseline) inplace_pattern = dict(baseline)
...@@ -274,20 +272,20 @@ def inplace_elemwise_optimizer_op(OP): ...@@ -274,20 +272,20 @@ def inplace_elemwise_optimizer_op(OP):
if hasattr(op.scalar_op, "make_new_inplace"): if hasattr(op.scalar_op, "make_new_inplace"):
new_scal = op.scalar_op.make_new_inplace( new_scal = op.scalar_op.make_new_inplace(
scalar.transfer_type( scalar.transfer_type(
*[inplace_pattern.get(i, None) \ *[inplace_pattern.get(i, None)
for i in xrange(len(node.outputs))])) for i in xrange(len(node.outputs))]))
else: else:
new_scal = op.scalar_op.__class__( new_scal = op.scalar_op.__class__(
scalar.transfer_type( scalar.transfer_type(
*[inplace_pattern.get(i, None) \ *[inplace_pattern.get(i, None)
for i in xrange(len(node.outputs))])) for i in xrange(len(node.outputs))]))
new_outputs = OP(new_scal, inplace_pattern)( new_outputs = OP(new_scal, inplace_pattern)(
*node.inputs, **dict(return_list=True)) *node.inputs, **dict(return_list=True))
new_node = new_outputs[0].owner new_node = new_outputs[0].owner
for r, new_r in zip(node.outputs, new_outputs): for r, new_r in zip(node.outputs, new_outputs):
fgraph.replace(r, new_r, fgraph.replace(r, new_r,
reason="inplace_elemwise_optimizer") reason="inplace_elemwise_optimizer")
nb_change_no_validate += 1 nb_change_no_validate += 1
if nb_change_no_validate >= check_each_change: if nb_change_no_validate >= check_each_change:
fgraph.validate() fgraph.validate()
...@@ -295,9 +293,9 @@ def inplace_elemwise_optimizer_op(OP): ...@@ -295,9 +293,9 @@ def inplace_elemwise_optimizer_op(OP):
nb_change_no_validate = 0 nb_change_no_validate = 0
except (ValueError, TypeError, InconsistencyError) as e: except (ValueError, TypeError, InconsistencyError) as e:
if check_each_change != 1 and not raised_warning: if check_each_change != 1 and not raised_warning:
print(( print(("Some inplace optimization was not "
"Some inplace optimization was not " "performed due to unexpected error:"),
"performed due to unexpected error:"), file=sys.stderr) file=sys.stderr)
print(e, file=sys.stderr) print(e, file=sys.stderr)
raised_warning = True raised_warning = True
fgraph.revert(chk) fgraph.revert(chk)
...@@ -313,7 +311,8 @@ def inplace_elemwise_optimizer_op(OP): ...@@ -313,7 +311,8 @@ def inplace_elemwise_optimizer_op(OP):
except Exception: except Exception:
if not raised_warning: if not raised_warning:
print(("Some inplace optimization was not " print(("Some inplace optimization was not "
"performed due to unexpected error"), file=sys.stderr) "performed due to unexpected error"),
file=sys.stderr)
fgraph.revert(chk) fgraph.revert(chk)
return inplace_elemwise_optimizer return inplace_elemwise_optimizer
...@@ -381,8 +380,8 @@ def register_specialize_device(lopt, *tags, **kwargs): ...@@ -381,8 +380,8 @@ def register_specialize_device(lopt, *tags, **kwargs):
# Register merge_optimizer as a global opt during canonicalize # Register merge_optimizer as a global opt during canonicalize
compile.optdb['canonicalize'].register( compile.optdb['canonicalize'].register('canon_merge', merge_optimizer,
'canon_merge', merge_optimizer, 'fast_run', final_opt=True) 'fast_run', final_opt=True)
##################### #####################
...@@ -512,11 +511,10 @@ def local_lift_transpose_through_dot(node): ...@@ -512,11 +511,10 @@ def local_lift_transpose_through_dot(node):
inplace. The newly-introduced transpositions are not inplace, this will inplace. The newly-introduced transpositions are not inplace, this will
be taken care of in a later optimization phase. be taken care of in a later optimization phase.
""" """
if not (isinstance(node.op, T.DimShuffle) if not (isinstance(node.op, T.DimShuffle) and node.op.new_order == (1, 0)):
and node.op.new_order == (1, 0)):
return False return False
if not (node.inputs[0].owner if not (node.inputs[0].owner and
and isinstance(node.inputs[0].owner.op, T.Dot)): isinstance(node.inputs[0].owner.op, T.Dot)):
return False return False
x, y = node.inputs[0].owner.inputs x, y = node.inputs[0].owner.inputs
...@@ -601,22 +599,19 @@ class MakeVector(T.Op): ...@@ -601,22 +599,19 @@ class MakeVector(T.Op):
def make_node(self, *inputs): def make_node(self, *inputs):
inputs = list(map(T.as_tensor_variable, inputs)) inputs = list(map(T.as_tensor_variable, inputs))
if not all(a.type == inputs[0].type for a in inputs) or ( if (not all(a.type == inputs[0].type for a in inputs) or
len(inputs) > 0 and inputs[0].dtype != self.dtype): (len(inputs) > 0 and inputs[0].dtype != self.dtype)):
dtype = theano.scalar.upcast(self.dtype, dtype = theano.scalar.upcast(self.dtype, *[i.dtype for i in inputs])
*[i.dtype for i in inputs])
# upcast the input to the determined dtype, # upcast the input to the determined dtype,
# but don't downcast anything # but don't downcast anything
assert dtype == self.dtype, ( assert dtype == self.dtype, (
"The upcast of the inputs to MakeVector should match the " "The upcast of the inputs to MakeVector should match the "
"dtype given in __init__.") "dtype given in __init__.")
if not all(self.dtype == T.cast(i, dtype=dtype).dtype if not all(self.dtype == T.cast(i, dtype=dtype).dtype
for i in inputs): for i in inputs):
raise TypeError("MakeVector.make_node expected inputs" raise TypeError("MakeVector.make_node expected inputs"
" upcastable to %s. got %s" % ( " upcastable to %s. got %s" %
self.dtype, (self.dtype, str([i.dtype for i in inputs])))
str([i.dtype for i in inputs])
))
inputs = [T.cast(i, dtype=dtype) for i in inputs] inputs = [T.cast(i, dtype=dtype) for i in inputs]
assert all(self.dtype == a.dtype for a in inputs) assert all(self.dtype == a.dtype for a in inputs)
assert all(a.ndim == 0 for a in inputs) assert all(a.ndim == 0 for a in inputs)
...@@ -625,11 +620,9 @@ class MakeVector(T.Op): ...@@ -625,11 +620,9 @@ class MakeVector(T.Op):
dtype = inputs[0].type.dtype dtype = inputs[0].type.dtype
else: else:
dtype = self.dtype dtype = self.dtype
#bcastable = (len(inputs) == 1) # bcastable = (len(inputs) == 1)
bcastable = False bcastable = False
otype = T.TensorType( otype = T.TensorType(broadcastable=(bcastable,), dtype=dtype)
broadcastable=(bcastable,),
dtype=dtype)
return T.Apply(self, inputs, [otype()]) return T.Apply(self, inputs, [otype()])
def __str__(self): def __str__(self):
...@@ -700,13 +693,14 @@ class MakeVectorPrinter: ...@@ -700,13 +693,14 @@ class MakeVectorPrinter:
if r.owner is None: if r.owner is None:
raise TypeError("Can only print make_vector.") raise TypeError("Can only print make_vector.")
elif isinstance(r.owner.op, MakeVector): elif isinstance(r.owner.op, MakeVector):
return "[%s]" % ", ".join(pstate.pprinter.process( return "[%s]" % ", ".join(
input, pstate.clone(precedence=1000)) for input pstate.pprinter.process(input, pstate.clone(precedence=1000))
in r.owner.inputs) for input in r.owner.inputs)
else: else:
raise TypeError("Can only print make_vector.") raise TypeError("Can only print make_vector.")
T.pprint.assign(lambda pstate, r: r.owner and isinstance(
r.owner.op, MakeVector), MakeVectorPrinter()) T.pprint.assign(lambda pstate, r: r.owner and
isinstance(r.owner.op, MakeVector), MakeVectorPrinter())
class ShapeFeature(object): class ShapeFeature(object):
...@@ -843,8 +837,8 @@ class ShapeFeature(object): ...@@ -843,8 +837,8 @@ class ShapeFeature(object):
# by always returning the same object to represent 1 # by always returning the same object to represent 1
return self.lscalar_one return self.lscalar_one
if (type(s_i) in integer_types or if (type(s_i) in integer_types or
isinstance(s_i, numpy.integer) or isinstance(s_i, numpy.integer) or
(isinstance(s_i, numpy.ndarray) and s_i.ndim == 0)): (isinstance(s_i, numpy.ndarray) and s_i.ndim == 0)):
# this shape is a constant # this shape is a constant
assert s_i >= 0 assert s_i >= 0
return T.constant(s_i, dtype='int64') return T.constant(s_i, dtype='int64')
...@@ -859,9 +853,9 @@ class ShapeFeature(object): ...@@ -859,9 +853,9 @@ class ShapeFeature(object):
# s_i is x.shape[i], we change it to Shape_i. # s_i is x.shape[i], we change it to Shape_i.
if (s_i.owner and if (s_i.owner and
isinstance(s_i.owner.op, Subtensor) and isinstance(s_i.owner.op, Subtensor) and
s_i.owner.inputs[0].owner and s_i.owner.inputs[0].owner and
isinstance(s_i.owner.inputs[0].owner.op, T.Shape)): isinstance(s_i.owner.inputs[0].owner.op, T.Shape)):
assert s_i.ndim == 0 assert s_i.ndim == 0
assert len(s_i.owner.op.idx_list) == 1 assert len(s_i.owner.op.idx_list) == 1
...@@ -883,7 +877,7 @@ class ShapeFeature(object): ...@@ -883,7 +877,7 @@ class ShapeFeature(object):
return s_i return s_i
else: else:
raise TypeError('Unsupported shape element', raise TypeError('Unsupported shape element',
s_i, type(s_i), getattr(s_i, 'type', None)) s_i, type(s_i), getattr(s_i, 'type', None))
def set_shape(self, r, s): def set_shape(self, r, s):
"""Assign the shape `s` to previously un-shaped variable `r`. """Assign the shape `s` to previously un-shaped variable `r`.
...@@ -910,7 +904,7 @@ class ShapeFeature(object): ...@@ -910,7 +904,7 @@ class ShapeFeature(object):
shape_vars = [] shape_vars = []
for i in xrange(r.ndim): for i in xrange(r.ndim):
if (hasattr(r.type, 'broadcastable') and if (hasattr(r.type, 'broadcastable') and
r.type.broadcastable[i]): r.type.broadcastable[i]):
shape_vars.append(self.lscalar_one) shape_vars.append(self.lscalar_one)
else: else:
shape_vars.append(self.unpack(s[i])) shape_vars.append(self.unpack(s[i]))
...@@ -947,8 +941,8 @@ class ShapeFeature(object): ...@@ -947,8 +941,8 @@ class ShapeFeature(object):
self.set_shape(r, other_shape) self.set_shape(r, other_shape)
return return
if (other_r.owner and r.owner and if (other_r.owner and r.owner and
other_r.owner.inputs == r.owner.inputs and other_r.owner.inputs == r.owner.inputs and
other_r.owner.op == r.owner.op): other_r.owner.op == r.owner.op):
# We are doing a merge. So the 2 shapes graph will be the # We are doing a merge. So the 2 shapes graph will be the
# same. This is only a speed optimization to call # same. This is only a speed optimization to call
# ancestors() less frequently. # ancestors() less frequently.
...@@ -957,10 +951,10 @@ class ShapeFeature(object): ...@@ -957,10 +951,10 @@ class ShapeFeature(object):
# Merge other_shape with r_shape, giving the priority to other_shape # Merge other_shape with r_shape, giving the priority to other_shape
merged_shape = [] merged_shape = []
for i, ps in enumerate(other_shape): for i, ps in enumerate(other_shape):
if (ps.owner if (ps.owner and
and isinstance(getattr(ps.owner, 'op', None), Shape_i) isinstance(getattr(ps.owner, 'op', None), Shape_i) and
and ps.owner.op.i == i ps.owner.op.i == i and
and ps.owner.inputs[0] in (r, other_r)): ps.owner.inputs[0] in (r, other_r)):
# If other_shape[i] is uninformative, use r_shape[i]. # If other_shape[i] is uninformative, use r_shape[i].
# For now, we consider 2 cases of uninformative other_shape[i]: # For now, we consider 2 cases of uninformative other_shape[i]:
# - Shape_i(i)(other_r); # - Shape_i(i)(other_r);
...@@ -1084,11 +1078,11 @@ class ShapeFeature(object): ...@@ -1084,11 +1078,11 @@ class ShapeFeature(object):
r in node.inputs]) r in node.inputs])
except NotImplementedError as e: except NotImplementedError as e:
raise NotImplementedError( raise NotImplementedError(
'Code called by infer_shape failed raising a ' 'Code called by infer_shape failed raising a '
'NotImplementedError. Raising NotImplementedError to ' 'NotImplementedError. Raising NotImplementedError to '
'indicate that a shape cannot be computed is no longer ' 'indicate that a shape cannot be computed is no longer '
'supported, and one should now use tensor.ShapeError ' 'supported, and one should now use tensor.ShapeError '
'instead. The original exception message is: %s' % e) 'instead. The original exception message is: %s' % e)
except Exception as e: except Exception as e:
msg = ('Failed to infer_shape from Op %s.\nInput shapes: ' msg = ('Failed to infer_shape from Op %s.\nInput shapes: '
'%s\nException encountered during infer_shape: ' '%s\nException encountered during infer_shape: '
...@@ -1108,10 +1102,10 @@ class ShapeFeature(object): ...@@ -1108,10 +1102,10 @@ class ShapeFeature(object):
if len(o_shapes) != len(node.outputs): if len(o_shapes) != len(node.outputs):
raise Exception( raise Exception(
('The infer_shape method for the Op "%s" returned a list ' + ('The infer_shape method for the Op "%s" returned a list ' +
'with the wrong number of element: len(o_shapes) = %d ' + 'with the wrong number of element: len(o_shapes) = %d ' +
' != len(node.outputs) = %d') % (str(node.op), ' != len(node.outputs) = %d') % (str(node.op),
len(o_shapes), len(o_shapes),
len(node.outputs))) len(node.outputs)))
# Ensure shapes are in 'int64'. This is to make sure the assert # Ensure shapes are in 'int64'. This is to make sure the assert
# found in the `local_useless_subtensor` optimization does not fail. # found in the `local_useless_subtensor` optimization does not fail.
...@@ -1173,9 +1167,9 @@ class ShapeFeature(object): ...@@ -1173,9 +1167,9 @@ class ShapeFeature(object):
# with the InputToGpuOptimizer optimizer. # with the InputToGpuOptimizer optimizer.
continue continue
if (repl.owner and if (repl.owner and
repl.owner.inputs[0] is shpnode.inputs[0] and repl.owner.inputs[0] is shpnode.inputs[0] and
isinstance(repl.owner.op, Shape_i) and isinstance(repl.owner.op, Shape_i) and
repl.owner.op.i == shpnode.op.i): repl.owner.op.i == shpnode.op.i):
# The replacement is a shape_i of the same # The replacement is a shape_i of the same
# input. So no need to do this equivalent # input. So no need to do this equivalent
# replacement. # replacement.
...@@ -1239,7 +1233,7 @@ class ShapeFeature(object): ...@@ -1239,7 +1233,7 @@ class ShapeFeature(object):
if not dx.owner or not dy.owner: if not dx.owner or not dy.owner:
return False return False
if (not isinstance(dx.owner.op, Shape_i) or if (not isinstance(dx.owner.op, Shape_i) or
not isinstance(dy.owner.op, Shape_i)): not isinstance(dy.owner.op, Shape_i)):
return False return False
opx = dx.owner.op opx = dx.owner.op
opy = dy.owner.op opy = dy.owner.op
...@@ -1310,10 +1304,9 @@ def local_fill_to_alloc(node): ...@@ -1310,10 +1304,9 @@ def local_fill_to_alloc(node):
return return
# TODO: cut out un-necessary dimshuffles of v # TODO: cut out un-necessary dimshuffles of v
assert rval[0].type == node.outputs[0].type, ('rval', rval[0].type, assert rval[0].type == node.outputs[0].type, (
'orig', node.outputs[0].type, 'rval', rval[0].type, 'orig', node.outputs[0].type, 'node',
'node', node, node,) # theano.printing.debugprint(node.outputs[0], file='str'))
) # theano.printing.debugprint(node.outputs[0], file='str'))
return rval return rval
...@@ -1404,7 +1397,7 @@ def local_subtensor_make_vector(node): ...@@ -1404,7 +1397,7 @@ def local_subtensor_make_vector(node):
try: try:
idx, = node.op.idx_list idx, = node.op.idx_list
except Exception: except Exception:
#'how can you have multiple indexes into a shape?' # 'how can you have multiple indexes into a shape?'
raise raise
if isinstance(idx, (scalar.Scalar, T.TensorType)): if isinstance(idx, (scalar.Scalar, T.TensorType)):
...@@ -1467,13 +1460,13 @@ def local_useless_elemwise(node): ...@@ -1467,13 +1460,13 @@ def local_useless_elemwise(node):
if isinstance(node.op, T.Elemwise): if isinstance(node.op, T.Elemwise):
if node.op.scalar_op == theano.scalar.eq and len(node.inputs) == 2: if node.op.scalar_op == theano.scalar.eq and len(node.inputs) == 2:
if node.inputs[0] == node.inputs[1]: if node.inputs[0] == node.inputs[1]:
# it is the same var in the graph. That will always be true # it is the same var in the graph. That will always be true
return [T.fill(node.inputs[0], return [T.fill(node.inputs[0],
T.constant(1.0, T.constant(1.0,
dtype=node.outputs[0].type.dtype))] dtype=node.outputs[0].type.dtype))]
elif node.op.scalar_op == theano.scalar.neq and len(node.inputs) == 2: elif node.op.scalar_op == theano.scalar.neq and len(node.inputs) == 2:
if node.inputs[0] == node.inputs[1]: if node.inputs[0] == node.inputs[1]:
# it is the same var in the graph. That will always be false # it is the same var in the graph. That will always be false
return [T.fill(node.inputs[0], return [T.fill(node.inputs[0],
T.constant(0.0, T.constant(0.0,
dtype=node.outputs[0].type.dtype))] dtype=node.outputs[0].type.dtype))]
...@@ -1482,8 +1475,8 @@ def local_useless_elemwise(node): ...@@ -1482,8 +1475,8 @@ def local_useless_elemwise(node):
elif node.op.scalar_op == theano.scalar.add and len(node.inputs) == 1: elif node.op.scalar_op == theano.scalar.add and len(node.inputs) == 1:
return [node.inputs[0]] return [node.inputs[0]]
elif (node.op.scalar_op == theano.scalar.identity elif (node.op.scalar_op == theano.scalar.identity and
and len(node.inputs) == 1): len(node.inputs) == 1):
return [node.inputs[0]] return [node.inputs[0]]
...@@ -1513,12 +1506,12 @@ def local_cast_cast(node): ...@@ -1513,12 +1506,12 @@ def local_cast_cast(node):
and the first cast cause an upcast. and the first cast cause an upcast.
""" """
if (not isinstance(node.op, T.Elemwise) or if (not isinstance(node.op, T.Elemwise) or
not isinstance(node.op.scalar_op, scalar.Cast)): not isinstance(node.op.scalar_op, scalar.Cast)):
return return
x = node.inputs[0] x = node.inputs[0]
if (not x.owner or if (not x.owner or
not isinstance(x.owner.op, T.Elemwise) or not isinstance(x.owner.op, T.Elemwise) or
not isinstance(x.owner.op.scalar_op, scalar.Cast)): not isinstance(x.owner.op.scalar_op, scalar.Cast)):
return return
if node.op.scalar_op.o_type == x.owner.op.scalar_op.o_type: if node.op.scalar_op.o_type == x.owner.op.scalar_op.o_type:
return [x] return [x]
...@@ -1738,7 +1731,7 @@ def local_elemwise_alloc_op(ElemwiseOP, AllocOP, DimShuffleOP): ...@@ -1738,7 +1731,7 @@ def local_elemwise_alloc_op(ElemwiseOP, AllocOP, DimShuffleOP):
# The broadcast pattern of the ouptut must match the broadcast # The broadcast pattern of the ouptut must match the broadcast
# pattern of at least one of the inputs. # pattern of at least one of the inputs.
if not any([i.type.broadcastable == if not any([i.type.broadcastable ==
node.outputs[0].type.broadcastable for i in node.inputs]): node.outputs[0].type.broadcastable for i in node.inputs]):
return False return False
def dimshuffled_alloc(i): def dimshuffled_alloc(i):
...@@ -1749,10 +1742,8 @@ def local_elemwise_alloc_op(ElemwiseOP, AllocOP, DimShuffleOP): ...@@ -1749,10 +1742,8 @@ def local_elemwise_alloc_op(ElemwiseOP, AllocOP, DimShuffleOP):
# At least one input must have an owner that is either a AllocOP or a # At least one input must have an owner that is either a AllocOP or a
# DimShuffleOP with an owner that is a AllocOP -- otherwise there is # DimShuffleOP with an owner that is a AllocOP -- otherwise there is
# nothing to optimize. # nothing to optimize.
if not any([i.owner if not any([i.owner and (isinstance(i.owner.op, AllocOP) or
and (isinstance(i.owner.op, AllocOP) or dimshuffled_alloc(i)) for i in node.inputs]):
dimshuffled_alloc(i))
for i in node.inputs]):
return False return False
# Search for input that we can use as a baseline for the dimensions. # Search for input that we can use as a baseline for the dimensions.
...@@ -1761,9 +1752,8 @@ def local_elemwise_alloc_op(ElemwiseOP, AllocOP, DimShuffleOP): ...@@ -1761,9 +1752,8 @@ def local_elemwise_alloc_op(ElemwiseOP, AllocOP, DimShuffleOP):
if i.type.broadcastable == node.outputs[0].type.broadcastable: if i.type.broadcastable == node.outputs[0].type.broadcastable:
# Prefer an input that is not a AllocOP nor a DimShuffleOP of a # Prefer an input that is not a AllocOP nor a DimShuffleOP of a
# AllocOP so that all allocs can be optimized. # AllocOP so that all allocs can be optimized.
if not (i.owner if not (i.owner and (isinstance(i.owner.op, AllocOP) or
and (isinstance(i.owner.op, AllocOP) dimshuffled_alloc(i))):
or dimshuffled_alloc(i))):
assert_op_idx = idx assert_op_idx = idx
break break
...@@ -1773,8 +1763,8 @@ def local_elemwise_alloc_op(ElemwiseOP, AllocOP, DimShuffleOP): ...@@ -1773,8 +1763,8 @@ def local_elemwise_alloc_op(ElemwiseOP, AllocOP, DimShuffleOP):
# there is more than one then do all but one. number of # there is more than one then do all but one. number of
# inputs with alloc or dimshuffle alloc # inputs with alloc or dimshuffle alloc
l2 = [i for i in node.inputs l2 = [i for i in node.inputs
if (i.owner and (isinstance(i.owner.op, AllocOP) if (i.owner and (isinstance(i.owner.op, AllocOP) or
or dimshuffled_alloc(i)))] dimshuffled_alloc(i)))]
# If only 1 alloc or dimshuffle alloc, it is the one we # If only 1 alloc or dimshuffle alloc, it is the one we
# will use for the shape. So no alloc would be removed. # will use for the shape. So no alloc would be removed.
if len(l2) > 1: if len(l2) > 1:
...@@ -1794,14 +1784,13 @@ def local_elemwise_alloc_op(ElemwiseOP, AllocOP, DimShuffleOP): ...@@ -1794,14 +1784,13 @@ def local_elemwise_alloc_op(ElemwiseOP, AllocOP, DimShuffleOP):
same_shape = node.fgraph.shape_feature.same_shape same_shape = node.fgraph.shape_feature.same_shape
for i in node.inputs: for i in node.inputs:
# Remove alloc # Remove alloc
if (i.owner and isinstance(i.owner.op, AllocOP) if (i.owner and isinstance(i.owner.op, AllocOP) and
and i.owner.inputs[0].type != i.owner.outputs[0].type): i.owner.inputs[0].type != i.owner.outputs[0].type):
# when i.owner.inputs[0].type == i.owner.outputs[0].type we # when i.owner.inputs[0].type == i.owner.outputs[0].type we
# will remove that alloc later # will remove that alloc later
assert i.type.ndim == cmp_op.ndim assert i.type.ndim == cmp_op.ndim
if (theano.config.experimental.local_alloc_elemwise_assert if (theano.config.experimental.local_alloc_elemwise_assert and
and not same_shape(i, cmp_op)): not same_shape(i, cmp_op)):
assert_op = assert_(assert_op, assert_op = assert_(assert_op,
*[T.eq(i.shape[idx], cmp_op.shape[idx]) *[T.eq(i.shape[idx], cmp_op.shape[idx])
for idx in xrange(i.type.ndim) for idx in xrange(i.type.ndim)
...@@ -1891,7 +1880,7 @@ def local_upcast_elemwise_constant_inputs(node): ...@@ -1891,7 +1880,7 @@ def local_upcast_elemwise_constant_inputs(node):
scalar_op = node.op.scalar_op scalar_op = node.op.scalar_op
# print "aa", scalar_op.output_types_preference # print "aa", scalar_op.output_types_preference
if (getattr(scalar_op, 'output_types_preference', None) if (getattr(scalar_op, 'output_types_preference', None)
in (T.scal.upgrade_to_float, T.scal.upcast_out)): in (T.scal.upgrade_to_float, T.scal.upcast_out)):
# this is the kind of op that we can screw with the input # this is the kind of op that we can screw with the input
# dtypes by upcasting explicitly # dtypes by upcasting explicitly
output_dtype = node.outputs[0].type.dtype output_dtype = node.outputs[0].type.dtype
...@@ -1910,11 +1899,12 @@ def local_upcast_elemwise_constant_inputs(node): ...@@ -1910,11 +1899,12 @@ def local_upcast_elemwise_constant_inputs(node):
else: else:
if shape_i is None: if shape_i is None:
return return
new_inputs.append(T.alloc(T.cast(cval_i, new_inputs.append(
output_dtype), T.alloc(T.cast(cval_i, output_dtype),
*[shape_i(d)(i) for d in xrange(i.ndim)])) *[shape_i(d)(i)
#print >> sys.stderr, "AAA", for d in xrange(i.ndim)]))
#*[Shape_i(d)(i) for d in xrange(i.ndim)] # print >> sys.stderr, "AAA",
# *[Shape_i(d)(i) for d in xrange(i.ndim)]
except NotScalarConstantError: except NotScalarConstantError:
# for the case of a non-scalar # for the case of a non-scalar
if isinstance(i, T.TensorConstant): if isinstance(i, T.TensorConstant):
...@@ -1958,7 +1948,7 @@ def local_useless_inc_subtensor(node): ...@@ -1958,7 +1948,7 @@ def local_useless_inc_subtensor(node):
except NotScalarConstantError: except NotScalarConstantError:
return return
if (node.inputs[0].ndim != node.inputs[1].ndim or if (node.inputs[0].ndim != node.inputs[1].ndim or
node.inputs[0].broadcastable != node.inputs[1].broadcastable): node.inputs[0].broadcastable != node.inputs[1].broadcastable):
# FB: I didn't check if this case can happen, but this opt # FB: I didn't check if this case can happen, but this opt
# don't support it. # don't support it.
return return
...@@ -1994,16 +1984,16 @@ def local_set_to_inc_subtensor(node): ...@@ -1994,16 +1984,16 @@ def local_set_to_inc_subtensor(node):
AdvancedIncSubtensor1(x, other, ilist, set_instead_of_inc=False) AdvancedIncSubtensor1(x, other, ilist, set_instead_of_inc=False)
""" """
if (isinstance(node.op, AdvancedIncSubtensor1) and if (isinstance(node.op, AdvancedIncSubtensor1) and
node.op.set_instead_of_inc == True and node.op.set_instead_of_inc and
node.inputs[1].owner and node.inputs[1].owner and
isinstance(node.inputs[1].owner.op, Elemwise) and isinstance(node.inputs[1].owner.op, Elemwise) and
isinstance(node.inputs[1].owner.op.scalar_op, scalar.Add)): isinstance(node.inputs[1].owner.op.scalar_op, scalar.Add)):
addn = node.inputs[1].owner addn = node.inputs[1].owner
subn = None subn = None
other = None other = None
if (addn.inputs[0].owner and if (addn.inputs[0].owner and
isinstance(addn.inputs[0].owner.op, AdvancedSubtensor1)): isinstance(addn.inputs[0].owner.op, AdvancedSubtensor1)):
subn = addn.inputs[0].owner subn = addn.inputs[0].owner
other = addn.inputs[1] other = addn.inputs[1]
elif (addn.inputs[1].owner and elif (addn.inputs[1].owner and
...@@ -2013,7 +2003,7 @@ def local_set_to_inc_subtensor(node): ...@@ -2013,7 +2003,7 @@ def local_set_to_inc_subtensor(node):
else: else:
return return
if (subn.inputs[1] != node.inputs[2] or if (subn.inputs[1] != node.inputs[2] or
subn.inputs[0] != node.inputs[0]): subn.inputs[0] != node.inputs[0]):
return return
return [advanced_inc_subtensor1(node.inputs[0], other, node.inputs[2])] return [advanced_inc_subtensor1(node.inputs[0], other, node.inputs[2])]
...@@ -2030,9 +2020,9 @@ def local_useless_slice(node): ...@@ -2030,9 +2020,9 @@ def local_useless_slice(node):
last_slice = len(slices) last_slice = len(slices)
for s in slices[::-1]: for s in slices[::-1]:
# check if slice and then check slice indices # check if slice and then check slice indices
if (isinstance(s, slice) and s.start is None and s.stop is None if (isinstance(s, slice) and s.start is None and s.stop is None and
and (s.step is None or T.extract_constant(s.step) == 1)): (s.step is None or T.extract_constant(s.step) == 1)):
last_slice -= 1 last_slice -= 1
else: else:
break break
# check if we removed something # check if we removed something
...@@ -2098,11 +2088,10 @@ def local_useless_subtensor(node): ...@@ -2098,11 +2088,10 @@ def local_useless_subtensor(node):
# the same underlying variable. # the same underlying variable.
if (length_pos_shape_i.owner and if (length_pos_shape_i.owner and
isinstance(length_pos_shape_i.owner.op, isinstance(length_pos_shape_i.owner.op,
T.ScalarFromTensor)): T.ScalarFromTensor)):
length_pos_shape_i = length_pos_shape_i.owner.inputs[0] length_pos_shape_i = length_pos_shape_i.owner.inputs[0]
elif (length_pos.owner and elif (length_pos.owner and
isinstance(length_pos.owner.op, isinstance(length_pos.owner.op, T.TensorFromScalar)):
T.TensorFromScalar)):
length_pos = length_pos.owner.inputs[0] length_pos = length_pos.owner.inputs[0]
else: else:
# We did not find underlying variables of the same type # We did not find underlying variables of the same type
...@@ -2322,8 +2311,8 @@ def merge_two_slices(slice1, len1, slice2, len2): ...@@ -2322,8 +2311,8 @@ def merge_two_slices(slice1, len1, slice2, len2):
pn_stop = sl1.start + (sl2.start - 1) * sl1.step pn_stop = sl1.start + (sl2.start - 1) * sl1.step
pn_stop = T.switch(T.and_(T.lt(pn_stop, 0), pn_stop = T.switch(T.and_(T.lt(pn_stop, 0),
T.gt(flen, 0)), T.gt(flen, 0)),
-len1 - 1, -len1 - 1,
T.minimum(pn_stop, sl1.stop)) T.minimum(pn_stop, sl1.stop))
pn_start = sl1.start + (sl2.stop - 1) * sl1.step pn_start = sl1.start + (sl2.stop - 1) * sl1.step
pn_start = T.minimum(pn_start, sl1.stop) pn_start = T.minimum(pn_start, sl1.stop)
pn_start = T.maximum(pn_start, 0) pn_start = T.maximum(pn_start, 0)
...@@ -2345,9 +2334,8 @@ def merge_two_slices(slice1, len1, slice2, len2): ...@@ -2345,9 +2334,8 @@ def merge_two_slices(slice1, len1, slice2, len2):
pp_start)) pp_start))
stop = T.switch(T.lt(reverse2 * reverse1, 0), stop = T.switch(T.lt(reverse2 * reverse1, 0),
T.switch(T.lt(reverse1, 0), np_stop, pn_stop), T.switch(T.lt(reverse1, 0), np_stop, pn_stop),
T.switch(T.lt(reverse1, 0), nn_stop, pp_stop T.switch(T.lt(reverse1, 0), nn_stop, pp_stop))
))
step = T.switch(T.lt(reverse2 * reverse1, 0), n_step, p_step) step = T.switch(T.lt(reverse2 * reverse1, 0), n_step, p_step)
start = T.switch(T.le(flen, 0), 0, start) start = T.switch(T.le(flen, 0), 0, start)
...@@ -2463,7 +2451,7 @@ def local_subtensor_of_alloc(node): ...@@ -2463,7 +2451,7 @@ def local_subtensor_of_alloc(node):
# We check that the corresponding val dimensions was # We check that the corresponding val dimensions was
# not a broadcasted dimensions. # not a broadcasted dimensions.
if (val.type.ndim > (i - n_added_dims) and if (val.type.ndim > (i - n_added_dims) and
val.type.broadcastable[i - n_added_dims]): val.type.broadcastable[i - n_added_dims]):
val_slices.append(slice(None)) val_slices.append(slice(None))
else: else:
val_slices.append(sl) val_slices.append(sl)
...@@ -2496,8 +2484,8 @@ def local_subtensor_of_alloc(node): ...@@ -2496,8 +2484,8 @@ def local_subtensor_of_alloc(node):
rval[0] = theano.tensor.unbroadcast( rval[0] = theano.tensor.unbroadcast(
rval[0], rval[0],
*[i for i, (b1, b2) in enumerate(zip(rval[0].broadcastable, *[i for i, (b1, b2) in enumerate(zip(rval[0].broadcastable,
node.outputs[0].broadcastable)) node.outputs[0].broadcastable))
if b1 and not b2]) if b1 and not b2])
return rval return rval
...@@ -2518,7 +2506,7 @@ def local_subtensor_of_dot(node): ...@@ -2518,7 +2506,7 @@ def local_subtensor_of_dot(node):
if not isinstance(node.op, Subtensor): if not isinstance(node.op, Subtensor):
return return
if (not node.inputs[0].owner or if (not node.inputs[0].owner or
not isinstance(node.inputs[0].owner.op, T.Dot)): not isinstance(node.inputs[0].owner.op, T.Dot)):
return return
# If there is other node that use the outputs of the dot # If there is other node that use the outputs of the dot
# We don't want to compute twice the sub part. # We don't want to compute twice the sub part.
...@@ -2540,7 +2528,8 @@ def local_subtensor_of_dot(node): ...@@ -2540,7 +2528,8 @@ def local_subtensor_of_dot(node):
# We skip this if b.ndim = 1, since then we just want b_sub = b, not b_sub = b[:] # We skip this if b.ndim = 1, since then we just want b_sub = b, not b_sub = b[:]
# (dot also handles b.ndim < 2 as a special case) # (dot also handles b.ndim < 2 as a special case)
if b.ndim > 1 and len(b_indices) >= b.ndim - 1: if b.ndim > 1 and len(b_indices) >= b.ndim - 1:
b_indices = b_indices[:b.ndim-2] + (slice(None, None, None),) + b_indices[b.ndim-2:] b_indices = (b_indices[:b.ndim - 2] +
(slice(None, None, None),) + b_indices[b.ndim - 2:])
a_sub = a.__getitem__(tuple(a_indices)) a_sub = a.__getitem__(tuple(a_indices))
b_sub = b.__getitem__(tuple(b_indices)) if b_indices else b b_sub = b.__getitem__(tuple(b_indices)) if b_indices else b
...@@ -2583,14 +2572,13 @@ def local_IncSubtensor_serialize(node): ...@@ -2583,14 +2572,13 @@ def local_IncSubtensor_serialize(node):
""" """
def movable(i): def movable(i):
# Return True iff this is a incsubtensor that we can move # Return True iff this is a incsubtensor that we can move
return i.owner \ return (i.owner and
and isinstance(i.owner.op, (IncSubtensor, isinstance(i.owner.op, (IncSubtensor,
AdvancedIncSubtensor1, AdvancedIncSubtensor1,
AdvancedIncSubtensor, AdvancedIncSubtensor,)) and
)) \ i.type == o_type and
and i.type == o_type \ len(i.clients) == 1 and
and len(i.clients) == 1 \ not i.owner.op.set_instead_of_inc)
and not i.owner.op.set_instead_of_inc
if node.op == T.add: if node.op == T.add:
o_type = node.outputs[0].type o_type = node.outputs[0].type
...@@ -2598,8 +2586,8 @@ def local_IncSubtensor_serialize(node): ...@@ -2598,8 +2586,8 @@ def local_IncSubtensor_serialize(node):
movable_inputs = [i for i in node.inputs if movable(i)] movable_inputs = [i for i in node.inputs if movable(i)]
if movable_inputs: if movable_inputs:
new_inputs = [i for i in node.inputs if not movable(i)] \ new_inputs = ([i for i in node.inputs if not movable(i)] +
+ [mi.owner.inputs[0] for mi in movable_inputs] [mi.owner.inputs[0] for mi in movable_inputs])
new_add = T.add(*new_inputs) new_add = T.add(*new_inputs)
# stack up the new incsubtensors # stack up the new incsubtensors
...@@ -2638,9 +2626,10 @@ def local_inplace_setsubtensor(node): ...@@ -2638,9 +2626,10 @@ def local_inplace_setsubtensor(node):
return [new_node] return [new_node]
return False return False
compile.optdb.register('local_inplace_setsubtensor', compile.optdb.register('local_inplace_setsubtensor',
TopoOptimizer(local_inplace_setsubtensor, TopoOptimizer(
failure_callback=TopoOptimizer.warn_inplace), 60, local_inplace_setsubtensor,
'fast_run', 'inplace') # DEBUG failure_callback=TopoOptimizer.warn_inplace),
60, 'fast_run', 'inplace') # DEBUG
@gof.local_optimizer([AdvancedIncSubtensor1], inplace=True) @gof.local_optimizer([AdvancedIncSubtensor1], inplace=True)
...@@ -2653,8 +2642,8 @@ def local_inplace_incsubtensor1(node): ...@@ -2653,8 +2642,8 @@ def local_inplace_incsubtensor1(node):
return False return False
compile.optdb.register('local_inplace_incsubtensor1', compile.optdb.register('local_inplace_incsubtensor1',
TopoOptimizer( TopoOptimizer(
local_inplace_incsubtensor1, local_inplace_incsubtensor1,
failure_callback=TopoOptimizer.warn_inplace), failure_callback=TopoOptimizer.warn_inplace),
60, 'fast_run', 'inplace') # DEBUG 60, 'fast_run', 'inplace') # DEBUG
...@@ -2671,7 +2660,7 @@ def local_incsubtensor_of_zeros(node): ...@@ -2671,7 +2660,7 @@ def local_incsubtensor_of_zeros(node):
if (isinstance(node.op, (IncSubtensor, if (isinstance(node.op, (IncSubtensor,
AdvancedIncSubtensor, AdvancedIncSubtensor,
AdvancedIncSubtensor1)) and AdvancedIncSubtensor1)) and
not node.op.set_instead_of_inc): not node.op.set_instead_of_inc):
x = node.inputs[0] x = node.inputs[0]
y = node.inputs[1] y = node.inputs[1]
replace = False replace = False
...@@ -2713,8 +2702,8 @@ def local_setsubtensor_of_constants(node): ...@@ -2713,8 +2702,8 @@ def local_setsubtensor_of_constants(node):
pass pass
if (replace_x is not None and if (replace_x is not None and
replace_y is not None and replace_y is not None and
replace_x == replace_y): replace_x == replace_y):
return [x] return [x]
else: else:
return False return False
...@@ -2738,7 +2727,7 @@ def local_adv_sub1_adv_inc_sub1(node): ...@@ -2738,7 +2727,7 @@ def local_adv_sub1_adv_inc_sub1(node):
return return
inp = node.inputs[0] inp = node.inputs[0]
if (not inp.owner or if (not inp.owner or
not isinstance(inp.owner.op, AdvancedIncSubtensor1)): not isinstance(inp.owner.op, AdvancedIncSubtensor1)):
return return
idx = node.inputs[1] idx = node.inputs[1]
idx2 = inp.owner.inputs[2] idx2 = inp.owner.inputs[2]
...@@ -2747,13 +2736,13 @@ def local_adv_sub1_adv_inc_sub1(node): ...@@ -2747,13 +2736,13 @@ def local_adv_sub1_adv_inc_sub1(node):
if idx is not idx2: if idx is not idx2:
return return
if (not inp.owner.op.set_instead_of_inc and if (not inp.owner.op.set_instead_of_inc and
T.extract_constant(x) != 0): T.extract_constant(x) != 0):
return return
cond = [T.all(T.and_(T.lt(idx, x.shape[0]), cond = [T.all(T.and_(T.lt(idx, x.shape[0]), T.ge(idx, -x.shape[0])))]
T.ge(idx, -x.shape[0])))]
if not node.fgraph.shape_feature.same_shape(idx, y, 0, 0): if not node.fgraph.shape_feature.same_shape(idx, y, 0, 0):
cond.append(T.eq(idx.shape[0], y.shape[0])) cond.append(T.eq(idx.shape[0], y.shape[0]))
y = Assert("Bad indexing or shapes in a AdvancedIncSubtensor1 that was optimized away")(y, *cond) y = Assert("Bad indexing or shapes in a AdvancedIncSubtensor1 "
"that was optimized away")(y, *cond)
if y.dtype == node.outputs[0].dtype: if y.dtype == node.outputs[0].dtype:
return [y] return [y]
...@@ -2828,33 +2817,34 @@ def local_useless_inc_subtensor_alloc(node): ...@@ -2828,33 +2817,34 @@ def local_useless_inc_subtensor_alloc(node):
# Build `z_broad` explicitly to include extra implicit dimensions. # Build `z_broad` explicitly to include extra implicit dimensions.
z_broad = ((True,) * (xi.ndim - z.ndim) + z.broadcastable) z_broad = ((True,) * (xi.ndim - z.ndim) + z.broadcastable)
cond = [# The shapes of `y` and `xi` must either agree or `y` may cond = [
# also have shape equal to 1 which may be treated as a # The shapes of `y` and `xi` must either agree or `y` may
# broadcastable dimension by the subtensor op. # also have shape equal to 1 which may be treated as a
T.or_(T.eq(y.shape[k], 1), T.eq(y.shape[k], xi.shape[k])) # broadcastable dimension by the subtensor op.
# Loop over all dimensions. T.or_(T.eq(y.shape[k], 1), T.eq(y.shape[k], xi.shape[k]))
for k in xrange(xi.ndim) # Loop over all dimensions.
# We need to check the above shapes, if for k in xrange(xi.ndim)
# * the pre-alloc increment `z` is broadcastable in # We need to check the above shapes, if
# dimension `k` (if it isn't, then the shapes of `z` and # * the pre-alloc increment `z` is broadcastable in
# `y` are the same by the definition of the `Alloc` op in # dimension `k` (if it isn't, then the shapes of `z` and
# this dimension and replacing `y` by `z` will not hide a # `y` are the same by the definition of the `Alloc` op in
# shape error), and # this dimension and replacing `y` by `z` will not hide a
# * `xi` and `y` do not have the same shape in dimension # shape error), and
# `k` or we cannot infer the shape statically (if the # * `xi` and `y` do not have the same shape in dimension
# shapes of `xi` and `y` are not the same, then replacing # `k` or we cannot infer the shape statically (if the
# `y` by `z` will hide the shape error of `y`), and # shapes of `xi` and `y` are not the same, then replacing
# * the shape of `y` is not equal to 1 or we cannot infer # `y` by `z` will hide the shape error of `y`), and
# the shape statically (if the shape of `y` is equal to # * the shape of `y` is not equal to 1 or we cannot infer
# 1, then `y` is broadcasted by the inc_subtensor op # the shape statically (if the shape of `y` is equal to
# internally, so the shapes of `xi` and `y` do not need # 1, then `y` is broadcasted by the inc_subtensor op
# to match in dimension `k`; else we need to check at # internally, so the shapes of `xi` and `y` do not need
# runtime that the shape of `y` is either 1 or the same # to match in dimension `k`; else we need to check at
# as `xi` or otherwise replacing `y` by `z` will hide a # runtime that the shape of `y` is either 1 or the same
# shape error). # as `xi` or otherwise replacing `y` by `z` will hide a
if (z_broad[k] and # shape error).
not same_shape(xi, y, dim_x=k, dim_y=k) and if (z_broad[k] and
shape_of[y][k] != 1)] not same_shape(xi, y, dim_x=k, dim_y=k) and
shape_of[y][k] != 1)]
if len(cond) > 0: if len(cond) > 0:
msg = '`x[i]` and `y` do not have the same shape.' msg = '`x[i]` and `y` do not have the same shape.'
...@@ -2916,7 +2906,7 @@ def local_rebroadcast_lift(node): ...@@ -2916,7 +2906,7 @@ def local_rebroadcast_lift(node):
# compilation phase. # compilation phase.
if hasattr(input, 'clients') and len(input.clients) == 1: if hasattr(input, 'clients') and len(input.clients) == 1:
rval = inode.op.make_node(T.Rebroadcast(*list(op.axis.items()))( rval = inode.op.make_node(T.Rebroadcast(*list(op.axis.items()))(
inode.inputs[0])).outputs inode.inputs[0])).outputs
return rval return rval
if inode and isinstance(inode.op, T.Rebroadcast): if inode and isinstance(inode.op, T.Rebroadcast):
# the "axis" specification in the outer Rebroadcast overrides # the "axis" specification in the outer Rebroadcast overrides
...@@ -3031,11 +3021,11 @@ def local_join_make_vector(node): ...@@ -3031,11 +3021,11 @@ def local_join_make_vector(node):
for idx in xrange(2, len(node.inputs)): for idx in xrange(2, len(node.inputs)):
inp = node.inputs[idx] inp = node.inputs[idx]
if (inp.owner and if (inp.owner and
isinstance(inp.owner.op, MakeVector) and isinstance(inp.owner.op, MakeVector) and
new_inputs[-1].owner and new_inputs[-1].owner and
isinstance(new_inputs[-1].owner.op, MakeVector) and isinstance(new_inputs[-1].owner.op, MakeVector) and
# MakeVector have a dtype parameter # MakeVector have a dtype parameter
inp.owner.op == new_inputs[-1].owner.op): inp.owner.op == new_inputs[-1].owner.op):
inps = new_inputs[-1].owner.inputs + inp.owner.inputs inps = new_inputs[-1].owner.inputs + inp.owner.inputs
new_inputs[-1] = inp.owner.op(*inps) new_inputs[-1] = inp.owner.op(*inps)
else: else:
...@@ -3059,7 +3049,7 @@ def local_remove_switch_const_cond(node): ...@@ -3059,7 +3049,7 @@ def local_remove_switch_const_cond(node):
if cond is constant and cond != 0: left if cond is constant and cond != 0: left
""" """
if (isinstance(node.op, T.Elemwise) and if (isinstance(node.op, T.Elemwise) and
isinstance(node.op.scalar_op, scalar.basic.Switch)): isinstance(node.op.scalar_op, scalar.basic.Switch)):
cond = T.extract_constant(node.inputs[0], elemwise=False) cond = T.extract_constant(node.inputs[0], elemwise=False)
if type(cond) is numpy.ndarray and cond.ndim == 0: if type(cond) is numpy.ndarray and cond.ndim == 0:
if cond == 0: if cond == 0:
...@@ -3241,9 +3231,9 @@ def local_flatten_lift(node): ...@@ -3241,9 +3231,9 @@ def local_flatten_lift(node):
nnet/sigm.py:log1msigm_to_softplus to get applied when there is a flatten. nnet/sigm.py:log1msigm_to_softplus to get applied when there is a flatten.
""" """
if (isinstance(node.op, T.Flatten) and if (isinstance(node.op, T.Flatten) and
node.inputs[0].owner and node.inputs[0].owner and
isinstance(node.inputs[0].owner.op, T.Elemwise) and isinstance(node.inputs[0].owner.op, T.Elemwise) and
len(node.inputs[0].owner.inputs) == 1): len(node.inputs[0].owner.inputs) == 1):
f = node.op(node.inputs[0].owner.inputs[0]) f = node.op(node.inputs[0].owner.inputs[0])
e = node.inputs[0].owner.op(f) e = node.inputs[0].owner.op(f)
return [e] return [e]
...@@ -3290,9 +3280,9 @@ def local_reshape_lift(node): ...@@ -3290,9 +3280,9 @@ def local_reshape_lift(node):
nnet/sigm.py:log1msigm_to_softplus to get applied when there is a reshape. nnet/sigm.py:log1msigm_to_softplus to get applied when there is a reshape.
""" """
if (isinstance(node.op, T.Reshape) and if (isinstance(node.op, T.Reshape) and
node.inputs[0].owner and node.inputs[0].owner and
isinstance(node.inputs[0].owner.op, T.Elemwise) and isinstance(node.inputs[0].owner.op, T.Elemwise) and
len(node.inputs[0].owner.inputs) == 1): len(node.inputs[0].owner.inputs) == 1):
r = node.op(node.inputs[0].owner.inputs[0], node.inputs[1]) r = node.op(node.inputs[0].owner.inputs[0], node.inputs[1])
e = node.inputs[0].owner.op(r) e = node.inputs[0].owner.op(r)
# In rare case the original broadcast was (False, True), but # In rare case the original broadcast was (False, True), but
...@@ -3539,7 +3529,7 @@ class Canonizer(gof.LocalOptimizer): ...@@ -3539,7 +3529,7 @@ class Canonizer(gof.LocalOptimizer):
return [input], [] return [input], []
if input.owner is None or input.owner.op not in [ if input.owner is None or input.owner.op not in [
self.main, self.inverse, self.reciprocal]: self.main, self.inverse, self.reciprocal]:
if input.owner and isinstance(input.owner.op, T.DimShuffle): if input.owner and isinstance(input.owner.op, T.DimShuffle):
# If input is a DimShuffle of some input which does # If input is a DimShuffle of some input which does
# something like this: # something like this:
...@@ -3552,9 +3542,9 @@ class Canonizer(gof.LocalOptimizer): ...@@ -3552,9 +3542,9 @@ class Canonizer(gof.LocalOptimizer):
# the num/denum of its input # the num/denum of its input
dsn = input.owner # dimshuffle node dsn = input.owner # dimshuffle node
dsop = dsn.op # dimshuffle op dsop = dsn.op # dimshuffle op
dsi0 = dsn.inputs[0] # the first input of the
# dimshuffle i.e. the ndarray to # the first input of the dimshuffle i.e. the ndarray to redim
# redim dsi0 = dsn.inputs[0]
# The compatible order is a DimShuffle "new_order" of the form: # The compatible order is a DimShuffle "new_order" of the form:
# ('x', ..., 'x', 0, 1, 2, ..., dimshuffle_input.type.ndim) # ('x', ..., 'x', 0, 1, 2, ..., dimshuffle_input.type.ndim)
...@@ -3566,9 +3556,9 @@ class Canonizer(gof.LocalOptimizer): ...@@ -3566,9 +3556,9 @@ class Canonizer(gof.LocalOptimizer):
# different numbers of dimensions (hence why we can # different numbers of dimensions (hence why we can
# discard its information - we know we can retrieve it # discard its information - we know we can retrieve it
# later on). # later on).
compatible_order = ('x',) * (input.type.ndim compatible_order = (('x',) *
- dsi0.type.ndim) + tuple( (input.type.ndim - dsi0.type.ndim) +
range(dsi0.type.ndim)) tuple(range(dsi0.type.ndim)))
if dsop.new_order == compatible_order: if dsop.new_order == compatible_order:
# If the "new_order" is the one we recognize, # If the "new_order" is the one we recognize,
# we return the num_denum of the dimshuffled input. # we return the num_denum of the dimshuffled input.
...@@ -3815,9 +3805,9 @@ class Canonizer(gof.LocalOptimizer): ...@@ -3815,9 +3805,9 @@ class Canonizer(gof.LocalOptimizer):
new = self.merge_num_denum(num, denum) new = self.merge_num_denum(num, denum)
if new.type.dtype != out.type.dtype: if new.type.dtype != out.type.dtype:
#new = T.fill(out, new) # new = T.fill(out, new)
elem_op = T.Elemwise(scalar.Identity(scalar.specific_out( elem_op = T.Elemwise(scalar.Identity(scalar.specific_out(
getattr(scalar, out.type.dtype)))) getattr(scalar, out.type.dtype))))
new = elem_op(new) new = elem_op(new)
assert (new.type == out.type) == (not (new.type != out.type)) assert (new.type == out.type) == (not (new.type != out.type))
...@@ -3833,12 +3823,12 @@ class Canonizer(gof.LocalOptimizer): ...@@ -3833,12 +3823,12 @@ class Canonizer(gof.LocalOptimizer):
else: else:
_logger.warning(' '.join(('CANONIZE FAILED: new, out = ', _logger.warning(' '.join(('CANONIZE FAILED: new, out = ',
new, ',', out, 'types', new, ',', out, 'types',
new.type, ',', out.type))) new.type, ',', out.type)))
return False return False
def __str__(self): def __str__(self):
return getattr(self, 'name', 'Canonizer(%s, %s, %s)' % ( return getattr(self, 'name', 'Canonizer(%s, %s, %s)' % (
self.main, self.inverse, self.reciprocal)) self.main, self.inverse, self.reciprocal))
def mul_calculate(num, denum, aslist=False, out_type=None): def mul_calculate(num, denum, aslist=False, out_type=None):
...@@ -3872,7 +3862,7 @@ register_canonicalize(local_mul_canonizer, name='local_mul_canonizer') ...@@ -3872,7 +3862,7 @@ register_canonicalize(local_mul_canonizer, name='local_mul_canonizer')
def local_neg_to_mul(node): def local_neg_to_mul(node):
if node.op == T.neg: if node.op == T.neg:
return [T.mul(numpy.array(-1, dtype=node.inputs[0].dtype), return [T.mul(numpy.array(-1, dtype=node.inputs[0].dtype),
node.inputs[0])] node.inputs[0])]
register_canonicalize(local_neg_to_mul) register_canonicalize(local_neg_to_mul)
...@@ -3924,10 +3914,10 @@ def local_elemwise_sub_zeros(node): ...@@ -3924,10 +3914,10 @@ def local_elemwise_sub_zeros(node):
""" """
Elemwise{sub}(X,X) -> zeros_like(X) Elemwise{sub}(X,X) -> zeros_like(X)
""" """
if (isinstance(node.op, T.Elemwise) if (isinstance(node.op, T.Elemwise) and
and node.op.scalar_op.nin == 2 node.op.scalar_op.nin == 2 and
and node.op.scalar_op == scalar.sub node.op.scalar_op == scalar.sub and
and node.inputs[0] == node.inputs[1]): node.inputs[0] == node.inputs[1]):
return [T.zeros_like(node.inputs[0])] return [T.zeros_like(node.inputs[0])]
...@@ -4013,9 +4003,8 @@ def local_sum_div_dimshuffle(node): ...@@ -4013,9 +4003,8 @@ def local_sum_div_dimshuffle(node):
' to False.') ' to False.')
new_denom = T.DimShuffle( new_denom = T.DimShuffle(
thing_dimshuffled.type.broadcastable, thing_dimshuffled.type.broadcastable,
new_new_order new_new_order)(thing_dimshuffled)
)(thing_dimshuffled)
return [T.true_div(node.op(numerator), new_denom)] return [T.true_div(node.op(numerator), new_denom)]
# else: # else:
# print 'incompatible dims:', axis, new_order # print 'incompatible dims:', axis, new_order
...@@ -4052,8 +4041,9 @@ def local_op_of_op(node): ...@@ -4052,8 +4041,9 @@ def local_op_of_op(node):
# We manipulate the graph so this is done to make sure the opt # We manipulate the graph so this is done to make sure the opt
# doesn't affect other computations. # doesn't affect other computations.
if len(node_inps.clients) == 1: if len(node_inps.clients) == 1:
if (node_inps.owner and (isinstance(node_inps.owner.op, T.elemwise.Prod) if (node_inps.owner and
or isinstance(node_inps.owner.op, T.elemwise.Sum))): (isinstance(node_inps.owner.op, T.elemwise.Prod) or
isinstance(node_inps.owner.op, T.elemwise.Sum))):
# check to see either the inner or outer prod is doing a # check to see either the inner or outer prod is doing a
# product over all axis, in which case we can remove it # product over all axis, in which case we can remove it
...@@ -4074,7 +4064,6 @@ def local_op_of_op(node): ...@@ -4074,7 +4064,6 @@ def local_op_of_op(node):
assert len(newaxis) == len(list(node_inps.owner.op.axis) + assert len(newaxis) == len(list(node_inps.owner.op.axis) +
list(node.op.axis)) list(node.op.axis))
# The old bugged logic. We keep it there to generate a warning # The old bugged logic. We keep it there to generate a warning
# when we generated bad code. # when we generated bad code.
alldims = list(range(node_inps.owner.inputs[0].type.ndim)) alldims = list(range(node_inps.owner.inputs[0].type.ndim))
...@@ -4087,20 +4076,20 @@ def local_op_of_op(node): ...@@ -4087,20 +4076,20 @@ def local_op_of_op(node):
if i not in alldims] if i not in alldims]
if (theano.config.warn.sum_sum_bug and if (theano.config.warn.sum_sum_bug and
newaxis != newaxis_old and newaxis != newaxis_old and
len(newaxis) == len(newaxis_old)): len(newaxis) == len(newaxis_old)):
_logger.warn( _logger.warn(
"WARNING (YOUR CURRENT CODE IS FINE): Theano " "WARNING (YOUR CURRENT CODE IS FINE): Theano "
"versions between version 9923a40c7b7a and August " "versions between version 9923a40c7b7a and August "
"2nd, 2010 generated bugged code in this case. " "2nd, 2010 generated bugged code in this case. "
"This happens when there are two consecutive sums " "This happens when there are two consecutive sums "
"in the graph and the intermediate sum is not " "in the graph and the intermediate sum is not "
"used elsewhere in the code. Some safeguard " "used elsewhere in the code. Some safeguard "
"removed some bad code, but not in all cases. You " "removed some bad code, but not in all cases. You "
"are in one such case. To disable this warning " "are in one such case. To disable this warning "
"(that you can safely ignore since this bug has " "(that you can safely ignore since this bug has "
"been fixed) set the theano flag " "been fixed) set the theano flag "
"`warn.sum_sum_bug` to False.") "`warn.sum_sum_bug` to False.")
combined = opt_type(newaxis, dtype=out_dtype) combined = opt_type(newaxis, dtype=out_dtype)
return [combined(node_inps.owner.inputs[0])] return [combined(node_inps.owner.inputs[0])]
...@@ -4126,9 +4115,8 @@ def local_reduce_join(node): ...@@ -4126,9 +4115,8 @@ def local_reduce_join(node):
""" """
if (isinstance(node.op, T.CAReduce) and if (isinstance(node.op, T.CAReduce) and
node.inputs[0].owner and node.inputs[0].owner and
isinstance(node.inputs[0].owner.op, T.Join)): isinstance(node.inputs[0].owner.op, T.Join)):
join = node.inputs[0].owner join = node.inputs[0].owner
if T.extract_constant(join.inputs[0]) != 0: if T.extract_constant(join.inputs[0]) != 0:
return return
...@@ -4149,7 +4137,8 @@ def local_reduce_join(node): ...@@ -4149,7 +4137,8 @@ def local_reduce_join(node):
if not inp: if not inp:
return return
if (not isinstance(inp.op, DimShuffle) or if (not isinstance(inp.op, DimShuffle) or
inp.op.new_order != ('x',) + tuple(range(inp.inputs[0].ndim))): inp.op.new_order != ('x',) +
tuple(range(inp.inputs[0].ndim))):
return return
new_inp.append(inp.inputs[0]) new_inp.append(inp.inputs[0])
ret = Elemwise(node.op.scalar_op)(*new_inp) ret = Elemwise(node.op.scalar_op)(*new_inp)
...@@ -4174,8 +4163,7 @@ def local_reduce_join(node): ...@@ -4174,8 +4163,7 @@ def local_reduce_join(node):
'optimization, that modified the pattern ' 'optimization, that modified the pattern '
'"Reduce{scalar.op}(Join(axis=0, a, b), axis=0)", ' '"Reduce{scalar.op}(Join(axis=0, a, b), axis=0)", '
'did not check the reduction axis. So if the ' 'did not check the reduction axis. So if the '
'reduction axis was not 0, you got a wrong answer.' 'reduction axis was not 0, you got a wrong answer.'))
))
return return
# We add the new check late to don't add extra warning. # We add the new check late to don't add extra warning.
...@@ -4204,7 +4192,7 @@ def local_cut_useless_reduce(node): ...@@ -4204,7 +4192,7 @@ def local_cut_useless_reduce(node):
# theano/tensor/tests/test_opt.py:T_local_reduce.test_local_reduce_broadcast_some_0 # theano/tensor/tests/test_opt.py:T_local_reduce.test_local_reduce_broadcast_some_0
# see gh-790 issue. # see gh-790 issue.
# #
#@register_canonicalize # @register_canonicalize
@register_uncanonicalize @register_uncanonicalize
@register_specialize @register_specialize
@gof.local_optimizer(ALL_REDUCE) @gof.local_optimizer(ALL_REDUCE)
...@@ -4258,7 +4246,7 @@ def local_opt_alloc(node): ...@@ -4258,7 +4246,7 @@ def local_opt_alloc(node):
input = node_inps.owner.inputs[0] input = node_inps.owner.inputs[0]
shapes = node_inps.owner.inputs[1:] shapes = node_inps.owner.inputs[1:]
if (node.op.axis is None or if (node.op.axis is None or
node.op.axis == tuple(range(input.ndim))): node.op.axis == tuple(range(input.ndim))):
try: try:
val = get_scalar_constant_value(input) val = get_scalar_constant_value(input)
assert val.size == 1 assert val.size == 1
...@@ -4346,7 +4334,7 @@ register_canonicalize(local_mul_zero) ...@@ -4346,7 +4334,7 @@ register_canonicalize(local_mul_zero)
@gof.local_optimizer([T.true_div]) @gof.local_optimizer([T.true_div])
def local_div_to_inv(node): def local_div_to_inv(node):
if node.op == T.true_div and N.all( if node.op == T.true_div and N.all(
local_mul_canonizer.get_constant(node.inputs[0]) == 1.0): local_mul_canonizer.get_constant(node.inputs[0]) == 1.0):
out = node.outputs[0] out = node.outputs[0]
new_out = T.inv(local_mul_canonizer.merge_num_denum(node.inputs[1:], new_out = T.inv(local_mul_canonizer.merge_num_denum(node.inputs[1:],
[])) []))
...@@ -4501,7 +4489,8 @@ def local_pow_specialize_device(node): ...@@ -4501,7 +4489,8 @@ def local_pow_specialize_device(node):
if abs(y) > 2: if abs(y) > 2:
# We fuse all the pow together here to make # We fuse all the pow together here to make
# compilation faster # compilation faster
rval1 = Elemwise(theano.scalar.Composite( rval1 = Elemwise(
theano.scalar.Composite(
[pow2_scal[0]], [rval1_scal])).make_node(xsym) [pow2_scal[0]], [rval1_scal])).make_node(xsym)
if y < 0: if y < 0:
rval = [T.inv(rval1)] rval = [T.inv(rval1)]
...@@ -4566,8 +4555,8 @@ def local_mul_specialize(node): ...@@ -4566,8 +4555,8 @@ def local_mul_specialize(node):
else: else:
# The next case would cause a replace by an equivalent case. # The next case would cause a replace by an equivalent case.
if (neg and if (neg and
nb_neg_node == 0 and nb_neg_node == 0 and
nb_cst == 1): nb_cst == 1):
return return
elif neg: elif neg:
# Don't add an extra neg node as we can't # Don't add an extra neg node as we can't
...@@ -4640,8 +4629,8 @@ def check_for_x_over_absX(numerators, denominators): ...@@ -4640,8 +4629,8 @@ def check_for_x_over_absX(numerators, denominators):
# TODO: this function should dig/search through dimshuffles # TODO: this function should dig/search through dimshuffles
# This won't catch a dimshuffled absolute value # This won't catch a dimshuffled absolute value
for den in list(denominators): for den in list(denominators):
if (den.owner and den.owner.op == T.abs_ if (den.owner and den.owner.op == T.abs_ and
and den.owner.inputs[0] in numerators): den.owner.inputs[0] in numerators):
if den.owner.inputs[0].type.dtype.startswith('complex'): if den.owner.inputs[0].type.dtype.startswith('complex'):
# TODO: Make an Op that projects a complex number to # TODO: Make an Op that projects a complex number to
# have unit length but projects 0 to 0. That # have unit length but projects 0 to 0. That
...@@ -4715,8 +4704,8 @@ def local_log1p(node): ...@@ -4715,8 +4704,8 @@ def local_log1p(node):
if node.op == T.log: if node.op == T.log:
log_arg, = node.inputs log_arg, = node.inputs
if log_arg.owner and log_arg.owner.op == T.add: if log_arg.owner and log_arg.owner.op == T.add:
scalars, scalar_inputs, nonconsts = \ scalars, scalar_inputs, nonconsts = scalarconsts_rest(
scalarconsts_rest(log_arg.owner.inputs) log_arg.owner.inputs)
# scalar_inputs are potentially dimshuffled and fill'd scalars # scalar_inputs are potentially dimshuffled and fill'd scalars
if scalars and numpy.allclose(numpy.sum(scalars), 1): if scalars and numpy.allclose(numpy.sum(scalars), 1):
if not nonconsts: if not nonconsts:
...@@ -4748,7 +4737,7 @@ def local_log_add(node): ...@@ -4748,7 +4737,7 @@ def local_log_add(node):
if len(zi) != 2: if len(zi) != 2:
# -- upgrading Maximum to handle multiple inputs wasn't trivial # -- upgrading Maximum to handle multiple inputs wasn't trivial
# TODO # TODO
#raise NotImplementedError() # raise NotImplementedError()
return return
pre_exp = [x.owner.inputs[0] for x in zi pre_exp = [x.owner.inputs[0] for x in zi
if x.owner and x.owner.op == T.exp] if x.owner and x.owner.op == T.exp]
...@@ -4945,8 +4934,7 @@ def constant_folding(node): ...@@ -4945,8 +4934,7 @@ def constant_folding(node):
storage_map[o] = [None] storage_map[o] = [None]
compute_map[o] = [False] compute_map[o] = [False]
if (hasattr(node.op, 'python_constant_folding') and if (hasattr(node.op, 'python_constant_folding') and
node.op.python_constant_folding(node)): node.op.python_constant_folding(node)):
old_value = getattr(node.op, '_op_use_c_code', False) old_value = getattr(node.op, '_op_use_c_code', False)
try: try:
node.op._op_use_c_code = False node.op._op_use_c_code = False
...@@ -5037,9 +5025,9 @@ register_specialize(local_one_minus_erf) ...@@ -5037,9 +5025,9 @@ register_specialize(local_one_minus_erf)
local_one_minus_erf2 = gof.PatternSub((T.add, local_one_minus_erf2 = gof.PatternSub((T.add,
1, 1,
(T.mul, -1, (T.erf, 'x'))), (T.mul, -1, (T.erf, 'x'))),
(T.erfc, 'x'), (T.erfc, 'x'),
allow_multiple_clients=True, allow_multiple_clients=True,
name='local_one_minus_erf2') name='local_one_minus_erf2')
register_canonicalize(local_one_minus_erf2) register_canonicalize(local_one_minus_erf2)
register_stabilize(local_one_minus_erf2) register_stabilize(local_one_minus_erf2)
register_specialize(local_one_minus_erf2) register_specialize(local_one_minus_erf2)
...@@ -5058,7 +5046,7 @@ register_canonicalize(local_one_plus_neg_erf) ...@@ -5058,7 +5046,7 @@ register_canonicalize(local_one_plus_neg_erf)
register_stabilize(local_one_plus_neg_erf) register_stabilize(local_one_plus_neg_erf)
register_specialize(local_one_plus_neg_erf) register_specialize(local_one_plus_neg_erf)
#(-1)+erf(x) => -erfc(x) don't need erf(x)+(-1) as the canonicalize # (-1)+erf(x) => -erfc(x) don't need erf(x)+(-1) as the canonicalize
# will put the -1 as the first argument. # will put the -1 as the first argument.
local_erf_minus_one = gof.PatternSub((T.add, local_erf_minus_one = gof.PatternSub((T.add,
dict(pattern='y', constraint=_is_minus1), dict(pattern='y', constraint=_is_minus1),
...@@ -5124,7 +5112,7 @@ register_canonicalize(local_one_add_neg_erfc) ...@@ -5124,7 +5112,7 @@ register_canonicalize(local_one_add_neg_erfc)
register_stabilize(local_one_add_neg_erfc) register_stabilize(local_one_add_neg_erfc)
register_specialize(local_one_add_neg_erfc) register_specialize(local_one_add_neg_erfc)
#(-1)+erfc(-x)=>erf(x) # (-1)+erfc(-x)=>erf(x)
local_erf_neg_minus_one = gof.PatternSub((T.add, local_erf_neg_minus_one = gof.PatternSub((T.add,
dict(pattern='y', constraint=_is_minus1), dict(pattern='y', constraint=_is_minus1),
(T.erfc, (T.neg, 'x'))), (T.erfc, (T.neg, 'x'))),
...@@ -5137,7 +5125,7 @@ register_canonicalize(local_erf_neg_minus_one) ...@@ -5137,7 +5125,7 @@ register_canonicalize(local_erf_neg_minus_one)
register_stabilize(local_erf_neg_minus_one) register_stabilize(local_erf_neg_minus_one)
register_specialize(local_erf_neg_minus_one) register_specialize(local_erf_neg_minus_one)
#(-1)+erfc(-1*x)=>erf(x) # (-1)+erfc(-1*x)=>erf(x)
local_erf_neg_minus_one2 = gof.PatternSub((T.add, local_erf_neg_minus_one2 = gof.PatternSub((T.add,
dict(pattern='y', constraint=_is_minus1), dict(pattern='y', constraint=_is_minus1),
(T.erfc, (T.mul, -1, 'x'))), (T.erfc, (T.mul, -1, 'x'))),
...@@ -5176,8 +5164,8 @@ def local_log_erfc(node): ...@@ -5176,8 +5164,8 @@ def local_log_erfc(node):
x = node.inputs[0].owner.inputs[0] x = node.inputs[0].owner.inputs[0]
stab_value = (-x ** 2 - T.log(x) - .5 * T.log(numpy.pi) + stab_value = (-x ** 2 - T.log(x) - .5 * T.log(numpy.pi) +
T.log(1 - 1 / (2 * x ** 2) + 3 / (4 * x ** 4) T.log(1 - 1 / (2 * x ** 2) + 3 / (4 * x ** 4) -
- 15 / (8 * x ** 6))) 15 / (8 * x ** 6)))
if (node.outputs[0].dtype == 'float32' or if (node.outputs[0].dtype == 'float32' or
node.outputs[0].dtype == 'float16'): node.outputs[0].dtype == 'float16'):
...@@ -5191,8 +5179,8 @@ def local_log_erfc(node): ...@@ -5191,8 +5179,8 @@ def local_log_erfc(node):
# Stability optimization of the grad of log(erfc(x)) # Stability optimization of the grad of log(erfc(x))
#([y*]exp(-(x**2)))/erfc(x) # The y* is optional # ([y*]exp(-(x**2)))/erfc(x) # The y* is optional
#([y*]exp(x**2))/erfc(-x) => [y*](when x>threashold, # ([y*]exp(x**2))/erfc(-x) => [y*](when x>threashold,
# sqrt(pi)*-x/(1-1/(2*x**2)+3/(4*x**4)-15/(8*x**6))) # sqrt(pi)*-x/(1-1/(2*x**2)+3/(4*x**4)-15/(8*x**6)))
# for float64: threshold=26.63 see at the end of the fct for the explaination # for float64: threshold=26.63 see at the end of the fct for the explaination
# for float32: threshold=9.3 see at the end of the fct for the explaination # for float32: threshold=9.3 see at the end of the fct for the explaination
...@@ -5226,8 +5214,8 @@ def local_grad_log_erfc_neg(node): ...@@ -5226,8 +5214,8 @@ def local_grad_log_erfc_neg(node):
if mul.owner.inputs[0].owner or len(mul.owner.inputs) != 2: if mul.owner.inputs[0].owner or len(mul.owner.inputs) != 2:
return False return False
y = mul.owner.inputs[0] y = mul.owner.inputs[0]
if (not mul.owner.inputs[1].owner if (not mul.owner.inputs[1].owner or
or mul.owner.inputs[1].owner.op != T.exp): mul.owner.inputs[1].owner.op != T.exp):
return False return False
exp = mul.owner.inputs[1] exp = mul.owner.inputs[1]
...@@ -5236,8 +5224,8 @@ def local_grad_log_erfc_neg(node): ...@@ -5236,8 +5224,8 @@ def local_grad_log_erfc_neg(node):
if exp.owner.inputs[0].owner.op == T.neg: if exp.owner.inputs[0].owner.op == T.neg:
neg = exp.owner.inputs[0] neg = exp.owner.inputs[0]
if (not neg.owner.inputs[0].owner if (not neg.owner.inputs[0].owner or
or neg.owner.inputs[0].owner.op != T.sqr): neg.owner.inputs[0].owner.op != T.sqr):
return False return False
sqr = neg.owner.inputs[0] sqr = neg.owner.inputs[0]
x = sqr.owner.inputs[0] x = sqr.owner.inputs[0]
...@@ -5279,8 +5267,8 @@ def local_grad_log_erfc_neg(node): ...@@ -5279,8 +5267,8 @@ def local_grad_log_erfc_neg(node):
return False return False
if len(mul_neg.owner.inputs) == 2: if len(mul_neg.owner.inputs) == 2:
if (not mul_neg.owner.inputs[1].owner if (not mul_neg.owner.inputs[1].owner or
or mul_neg.owner.inputs[1].owner.op != T.sqr): mul_neg.owner.inputs[1].owner.op != T.sqr):
return False return False
sqr = mul_neg.owner.inputs[1] sqr = mul_neg.owner.inputs[1]
x = sqr.owner.inputs[0] x = sqr.owner.inputs[0]
...@@ -5292,8 +5280,8 @@ def local_grad_log_erfc_neg(node): ...@@ -5292,8 +5280,8 @@ def local_grad_log_erfc_neg(node):
return False return False
if cst2 != -1: if cst2 != -1:
if (not erfc_x.owner or erfc_x.owner.op != T.mul if (not erfc_x.owner or erfc_x.owner.op != T.mul or
or len(erfc_x.owner.inputs) != 2): len(erfc_x.owner.inputs) != 2):
# todo implement that case # todo implement that case
return False return False
if erfc_x.owner.inputs[1] is not mul_neg.owner.inputs[1]: if erfc_x.owner.inputs[1] is not mul_neg.owner.inputs[1]:
...@@ -5324,12 +5312,12 @@ def local_grad_log_erfc_neg(node): ...@@ -5324,12 +5312,12 @@ def local_grad_log_erfc_neg(node):
# aaron value # aaron value
stab_value = (x * T.pow(1 - 1 / (2 * (x ** 2)) + stab_value = (x * T.pow(1 - 1 / (2 * (x ** 2)) +
3 / (4 * (x ** 4)) - 15 / (8 * (x ** 6)), -1) 3 / (4 * (x ** 4)) - 15 / (8 * (x ** 6)), -1) *
* T.cast(T.sqrt(numpy.pi), dtype=x.dtype)) T.cast(T.sqrt(numpy.pi), dtype=x.dtype))
if x.dtype == 'float32' or x.dtype == 'float16': if x.dtype == 'float32' or x.dtype == 'float16':
threshold = 9.3 threshold = 9.3
#threshold = 10.1 # threshold = 10.1
elif x.dtype == 'float64': elif x.dtype == 'float64':
threshold = 26.641747557 threshold = 26.641747557
ret = T.switch(x < threshold, true_div_no_mul, stab_value) * y ret = T.switch(x < threshold, true_div_no_mul, stab_value) * y
...@@ -5531,6 +5519,7 @@ def local_elemwise_fusion_op(OP, max_input_fct=lambda node: 32, ...@@ -5531,6 +5519,7 @@ def local_elemwise_fusion_op(OP, max_input_fct=lambda node: 32,
if maker is None: if maker is None:
def maker(node, scalar_op): def maker(node, scalar_op):
return OP(scalar_op) return OP(scalar_op)
def local_fuse(node): def local_fuse(node):
""" """
As part of specialization, we fuse two consecutive elemwise Ops of the As part of specialization, we fuse two consecutive elemwise Ops of the
...@@ -5598,13 +5587,13 @@ def local_elemwise_fusion_op(OP, max_input_fct=lambda node: 32, ...@@ -5598,13 +5587,13 @@ def local_elemwise_fusion_op(OP, max_input_fct=lambda node: 32,
# If a variable is used as multiple into to the same node, # If a variable is used as multiple into to the same node,
# we still want to fusion. So we take the set. # we still want to fusion. So we take the set.
if (i.owner and if (i.owner and
isinstance(i.owner.op, OP) and isinstance(i.owner.op, OP) and
len(set([n for n, idx in i.clients])) == 1 and len(set([n for n, idx in i.clients])) == 1 and
# Do not merge elemwise that don't have the same # Do not merge elemwise that don't have the same
# broadcastable pattern to don't redo duplicate # broadcastable pattern to don't redo duplicate
# computation due to broadcast. # computation due to broadcast.
i.owner.outputs[0].broadcastable == node.outputs[0].broadcastable): i.owner.outputs[0].broadcastable ==
node.outputs[0].broadcastable):
do_fusion = True do_fusion = True
try: try:
tmp_s_input = [] tmp_s_input = []
...@@ -5840,14 +5829,14 @@ def local_add_mul_fusion(node): ...@@ -5840,14 +5829,14 @@ def local_add_mul_fusion(node):
""" """
if (not isinstance(node.op, Elemwise) or if (not isinstance(node.op, Elemwise) or
not isinstance(node.op.scalar_op, (scalar.Add, scalar.Mul))): not isinstance(node.op.scalar_op, (scalar.Add, scalar.Mul))):
return False return False
s_op = node.op.scalar_op.__class__ s_op = node.op.scalar_op.__class__
for inp in node.inputs: for inp in node.inputs:
if (inp.owner and if (inp.owner and
isinstance(inp.owner.op, Elemwise) and isinstance(inp.owner.op, Elemwise) and
isinstance(inp.owner.op.scalar_op, s_op)): isinstance(inp.owner.op.scalar_op, s_op)):
l = list(node.inputs) l = list(node.inputs)
l.remove(inp) l.remove(inp)
return [node.op(*(l + inp.owner.inputs))] return [node.op(*(l + inp.owner.inputs))]
...@@ -5882,13 +5871,15 @@ else: ...@@ -5882,13 +5871,15 @@ else:
# just returns the input, it should be removed from the graph to # just returns the input, it should be removed from the graph to
# make sure all possible optimizations can be applied. # make sure all possible optimizations can be applied.
register_canonicalize(gof.OpRemove(theano.gradient.consider_constant_), register_canonicalize(gof.OpRemove(theano.gradient.consider_constant_),
'fast_compile', 'fast_run', name='remove_consider_constant') 'fast_compile', 'fast_run',
name='remove_consider_constant')
register_canonicalize(gof.OpRemove(theano.gradient.zero_grad_), register_canonicalize(gof.OpRemove(theano.gradient.zero_grad_),
'fast_compile', 'fast_run', name='remove_zero_grad') 'fast_compile', 'fast_run', name='remove_zero_grad')
register_canonicalize(gof.OpRemove(theano.gradient.disconnected_grad_), register_canonicalize(gof.OpRemove(theano.gradient.disconnected_grad_),
'fast_compile', 'fast_run', name='remove_disconnected_grad') 'fast_compile', 'fast_run',
name='remove_disconnected_grad')
@register_canonicalize @register_canonicalize
......
theano/tensor/raw_random.py
浏览文件 @ bd11e130
"""Define random number Type (`RandomStateType`) and Op (`RandomFunction`).""" """Define random number Type (`RandomStateType`) and Op (`RandomFunction`)."""
from __future__ import print_function from __future__ import print_function
__docformat__ = "restructuredtext en"
import sys import sys
from copy import copy from copy import copy
...@@ -15,6 +15,8 @@ from theano import gof ...@@ -15,6 +15,8 @@ from theano import gof
from six import string_types from six import string_types
from theano.compile import optdb from theano.compile import optdb
__docformat__ = "restructuredtext en"
class RandomStateType(gof.Type): class RandomStateType(gof.Type):
"""A Type wrapper for numpy.random.RandomState """A Type wrapper for numpy.random.RandomState
...@@ -85,13 +87,13 @@ class RandomStateType(gof.Type): ...@@ -85,13 +87,13 @@ class RandomStateType(gof.Type):
# Register RandomStateType's C code for ViewOp. # Register RandomStateType's C code for ViewOp.
theano.compile.register_view_op_c_code( theano.compile.register_view_op_c_code(
RandomStateType, RandomStateType,
""" """
Py_XDECREF(%(oname)s); Py_XDECREF(%(oname)s);
%(oname)s = %(iname)s; %(oname)s = %(iname)s;
Py_XINCREF(%(oname)s); Py_XINCREF(%(oname)s);
""", """,
1) 1)
random_state_type = RandomStateType() random_state_type = RandomStateType()
...@@ -135,9 +137,8 @@ class RandomFunction(gof.Op): ...@@ -135,9 +137,8 @@ class RandomFunction(gof.Op):
and self.ndim_added == other.ndim_added and self.ndim_added == other.ndim_added
def __hash__(self): def __hash__(self):
return hash(type(self)) ^ hash(self.fn) \ return (hash(type(self)) ^ hash(self.fn) ^ hash(self.outtype) ^
^ hash(self.outtype) \ hash(self.inplace) ^ hash(self.ndim_added))
^ hash(self.inplace) ^ hash(self.ndim_added)
def __getstate__(self): def __getstate__(self):
return self.state return self.state
...@@ -233,7 +234,6 @@ class RandomFunction(gof.Op): ...@@ -233,7 +234,6 @@ class RandomFunction(gof.Op):
# copy of r if self.inplace is False # copy of r if self.inplace is False
r, shape, args = inputs[0], inputs[1], inputs[2:] r, shape, args = inputs[0], inputs[1], inputs[2:]
assert type(r) == numpy.random.RandomState, (type(r), r) assert type(r) == numpy.random.RandomState, (type(r), r)
r_orig = r
# If shape == [], that means no shape is enforced, and numpy is # If shape == [], that means no shape is enforced, and numpy is
# trusted to draw the appropriate number of samples, numpy uses # trusted to draw the appropriate number of samples, numpy uses
...@@ -245,16 +245,16 @@ class RandomFunction(gof.Op): ...@@ -245,16 +245,16 @@ class RandomFunction(gof.Op):
shape = tuple(shape) shape = tuple(shape)
if (shape is not None and if (shape is not None and
self.outtype.ndim != len(shape) + self.ndim_added): self.outtype.ndim != len(shape) + self.ndim_added):
raise ValueError('Shape mismatch: self.outtype.ndim (%i) !=' raise ValueError('Shape mismatch: self.outtype.ndim (%i) !='
' len(shape) (%i) + self.ndim_added (%i)' ' len(shape) (%i) + self.ndim_added (%i)'
% (self.outtype.ndim, len(shape), self.ndim_added)) % (self.outtype.ndim, len(shape), self.ndim_added))
if not self.inplace: if not self.inplace:
r = copy(r) r = copy(r)
rout[0] = r rout[0] = r
rval = self.fn(r, *(args + [shape])) rval = self.fn(r, *(args + [shape]))
if not isinstance(rval, numpy.ndarray) \ if (not isinstance(rval, numpy.ndarray) or
or str(rval.dtype) != node.outputs[1].type.dtype: str(rval.dtype) != node.outputs[1].type.dtype):
rval = theano._asarray(rval, dtype=node.outputs[1].type.dtype) rval = theano._asarray(rval, dtype=node.outputs[1].type.dtype)
# When shape is None, numpy has a tendency to unexpectedly # When shape is None, numpy has a tendency to unexpectedly
...@@ -288,7 +288,7 @@ class RandomFunction(gof.Op): ...@@ -288,7 +288,7 @@ class RandomFunction(gof.Op):
def grad(self, inputs, outputs): def grad(self, inputs, outputs):
return [theano.gradient.grad_undefined(self, k, inp, return [theano.gradient.grad_undefined(self, k, inp,
'No gradient defined through raw random numbers op') 'No gradient defined through raw random numbers op')
for k, inp in enumerate(inputs)] for k, inp in enumerate(inputs)]
def R_op(self, inputs, eval_points): def R_op(self, inputs, eval_points):
...@@ -325,8 +325,8 @@ def _infer_ndim_bcast(ndim, shape, *args): ...@@ -325,8 +325,8 @@ def _infer_ndim_bcast(ndim, shape, *args):
else: else:
if shape_ndim != ndim: if shape_ndim != ndim:
raise ValueError('ndim should be equal to len(shape), but\n', raise ValueError('ndim should be equal to len(shape), but\n',
'ndim = %s, len(shape) = %s, shape = %s' 'ndim = %s, len(shape) = %s, shape = %s'
% (ndim, shape_ndim, shape)) % (ndim, shape_ndim, shape))
bcast = [] bcast = []
pre_v_shape = [] pre_v_shape = []
...@@ -353,7 +353,8 @@ def _infer_ndim_bcast(ndim, shape, *args): ...@@ -353,7 +353,8 @@ def _infer_ndim_bcast(ndim, shape, *args):
break break
else: else:
if n_a_i == 0: if n_a_i == 0:
raise ValueError(('Auto-shape of -1 must overlap' raise ValueError((
'Auto-shape of -1 must overlap'
'with the shape of one of the broadcastable' 'with the shape of one of the broadcastable'
'inputs')) 'inputs'))
else: else:
...@@ -373,7 +374,7 @@ def _infer_ndim_bcast(ndim, shape, *args): ...@@ -373,7 +374,7 @@ def _infer_ndim_bcast(ndim, shape, *args):
# but we need to know ndim # but we need to know ndim
if not args: if not args:
raise TypeError(('_infer_ndim_bcast cannot infer shape without' raise TypeError(('_infer_ndim_bcast cannot infer shape without'
' either shape or args')) ' either shape or args'))
template = reduce(lambda a, b: a + b, args) template = reduce(lambda a, b: a + b, args)
v_shape = template.shape v_shape = template.shape
bcast = template.broadcastable bcast = template.broadcastable
...@@ -463,7 +464,7 @@ def uniform(random_state, size=None, low=0.0, high=1.0, ndim=None, dtype=None): ...@@ -463,7 +464,7 @@ def uniform(random_state, size=None, low=0.0, high=1.0, ndim=None, dtype=None):
dtype = tensor.scal.upcast(theano.config.floatX, low.dtype, high.dtype) dtype = tensor.scal.upcast(theano.config.floatX, low.dtype, high.dtype)
ndim, size, bcast = _infer_ndim_bcast(ndim, size, low, high) ndim, size, bcast = _infer_ndim_bcast(ndim, size, low, high)
op = RandomFunction('uniform', op = RandomFunction('uniform',
tensor.TensorType(dtype=dtype, broadcastable=bcast)) tensor.TensorType(dtype=dtype, broadcastable=bcast))
return op(random_state, size, low, high) return op(random_state, size, low, high)
...@@ -487,7 +488,7 @@ def normal(random_state, size=None, avg=0.0, std=1.0, ndim=None, dtype=None): ...@@ -487,7 +488,7 @@ def normal(random_state, size=None, avg=0.0, std=1.0, ndim=None, dtype=None):
dtype = tensor.scal.upcast(theano.config.floatX, avg.dtype, std.dtype) dtype = tensor.scal.upcast(theano.config.floatX, avg.dtype, std.dtype)
ndim, size, bcast = _infer_ndim_bcast(ndim, size, avg, std) ndim, size, bcast = _infer_ndim_bcast(ndim, size, avg, std)
op = RandomFunction('normal', op = RandomFunction('normal',
tensor.TensorType(dtype=dtype, broadcastable=bcast)) tensor.TensorType(dtype=dtype, broadcastable=bcast))
return op(random_state, size, avg, std) return op(random_state, size, avg, std)
...@@ -517,7 +518,8 @@ def binomial(random_state, size=None, n=1, p=0.5, ndim=None, ...@@ -517,7 +518,8 @@ def binomial(random_state, size=None, n=1, p=0.5, ndim=None,
# p=numpy.asarray([.1, .2, .3], dtype='float64')) # p=numpy.asarray([.1, .2, .3], dtype='float64'))
n = tensor.cast(n, 'int32') n = tensor.cast(n, 'int32')
op = RandomFunction('binomial', op = RandomFunction('binomial',
tensor.TensorType(dtype=dtype, broadcastable=(False,) * ndim)) tensor.TensorType(dtype=dtype,
broadcastable=(False,) * ndim))
return op(random_state, size, n, p) return op(random_state, size, n, p)
...@@ -583,7 +585,7 @@ def random_integers(random_state, size=None, low=0, high=1, ndim=None, ...@@ -583,7 +585,7 @@ def random_integers(random_state, size=None, low=0, high=1, ndim=None,
high = tensor.as_tensor_variable(high) high = tensor.as_tensor_variable(high)
ndim, size, bcast = _infer_ndim_bcast(ndim, size, low, high) ndim, size, bcast = _infer_ndim_bcast(ndim, size, low, high)
op = RandomFunction(random_integers_helper, op = RandomFunction(random_integers_helper,
tensor.TensorType(dtype=dtype, broadcastable=bcast)) tensor.TensorType(dtype=dtype, broadcastable=bcast))
return op(random_state, size, low, high) return op(random_state, size, low, high)
...@@ -719,8 +721,9 @@ def permutation(random_state, size=None, n=1, ndim=None, dtype='int64'): ...@@ -719,8 +721,9 @@ def permutation(random_state, size=None, n=1, ndim=None, dtype='int64'):
ndim, size, bcast = _infer_ndim_bcast(ndim, size) ndim, size, bcast = _infer_ndim_bcast(ndim, size)
# print "NDIM", ndim, size # print "NDIM", ndim, size
op = RandomFunction(permutation_helper, op = RandomFunction(permutation_helper,
tensor.TensorType(dtype=dtype, broadcastable=bcast + (False,)), tensor.TensorType(dtype=dtype,
ndim_added=1) broadcastable=bcast + (False,)),
ndim_added=1)
return op(random_state, size, n) return op(random_state, size, n)
...@@ -738,14 +741,11 @@ def multinomial_helper(random_state, n, pvals, size): ...@@ -738,14 +741,11 @@ def multinomial_helper(random_state, n, pvals, size):
ndim = len(size) ndim = len(size)
else: else:
ndim = max(n.ndim, pvals.ndim - 1) ndim = max(n.ndim, pvals.ndim - 1)
out_ndim = ndim + 1
# broadcast n to ndim dimensions and pvals to ndim+1 # broadcast n to ndim dimensions and pvals to ndim+1
if n.ndim > ndim: if n.ndim > ndim:
raise ValueError( raise ValueError('n.ndim (%i) should not be larger than len(size) (%i)'
'n.ndim (%i) should not be larger than len(size) (%i)' % (n.ndim, ndim), n, size)
% (n.ndim, ndim),
n, size)
if n.ndim < ndim: if n.ndim < ndim:
n = n.reshape((1,) * (ndim - n.ndim) + n.shape) n = n.reshape((1,) * (ndim - n.ndim) + n.shape)
...@@ -788,7 +788,7 @@ def multinomial_helper(random_state, n, pvals, size): ...@@ -788,7 +788,7 @@ def multinomial_helper(random_state, n, pvals, size):
# because mtrand.pyx has a ValueError that will trigger if # because mtrand.pyx has a ValueError that will trigger if
# sum(pvals[:-1]) > 1.0 # sum(pvals[:-1]) > 1.0
pvi = pvi * (1.0 - 5e-5) pvi = pvi * (1.0 - 5e-5)
#pvi = pvi * .9 # pvi = pvi * .9
pisum = numpy.sum(pvi) pisum = numpy.sum(pvi)
elif pvi[-1] < 5e-5: # will this even work? elif pvi[-1] < 5e-5: # will this even work?
pvi = pvi * (1.0 - 5e-5) pvi = pvi * (1.0 - 5e-5)
...@@ -859,8 +859,9 @@ def multinomial(random_state, size=None, n=1, pvals=[0.5, 0.5], ...@@ -859,8 +859,9 @@ def multinomial(random_state, size=None, n=1, pvals=[0.5, 0.5],
ndim, size, bcast = _infer_ndim_bcast(ndim, size, n, tmp) ndim, size, bcast = _infer_ndim_bcast(ndim, size, n, tmp)
bcast = bcast + (pvals.type.broadcastable[-1],) bcast = bcast + (pvals.type.broadcastable[-1],)
op = RandomFunction(multinomial_helper, op = RandomFunction(multinomial_helper,
tensor.TensorType(dtype=dtype, broadcastable=bcast), tensor.TensorType(dtype=dtype,
ndim_added=1) broadcastable=bcast),
ndim_added=1)
return op(random_state, size, n, pvals) return op(random_state, size, n, pvals)
......
theano/tensor/shared_randomstreams.py
浏览文件 @ bd11e130
"""Define RandomStreams, providing random number variables for Theano """Define RandomStreams, providing random number variables for Theano
graphs. graphs.
""" """
__docformat__ = "restructuredtext en"
import copy import copy
import numpy import numpy
from theano.compile.sharedvalue import (SharedVariable, shared_constructor, from theano.compile.sharedvalue import (SharedVariable, shared_constructor,
shared) shared)
from theano.tensor import raw_random from theano.tensor import raw_random
__docformat__ = "restructuredtext en"
class RandomStateSharedVariable(SharedVariable): class RandomStateSharedVariable(SharedVariable):
pass pass
...@@ -77,7 +79,7 @@ class RandomStreams(raw_random.RandomStreamsBase): ...@@ -77,7 +79,7 @@ class RandomStreams(raw_random.RandomStreamsBase):
for old_r, new_r in self.state_updates: for old_r, new_r in self.state_updates:
old_r_seed = seedgen.randint(2 ** 30) old_r_seed = seedgen.randint(2 ** 30)
old_r.set_value(numpy.random.RandomState(int(old_r_seed)), old_r.set_value(numpy.random.RandomState(int(old_r_seed)),
borrow=True) borrow=True)
def __getitem__(self, item): def __getitem__(self, item):
"""Retrieve the numpy RandomState instance associated with a """Retrieve the numpy RandomState instance associated with a
......
theano/tensor/sharedvar.py
浏览文件 @ bd11e130
...@@ -41,10 +41,10 @@ def tensor_constructor(value, name=None, strict=False, allow_downcast=None, ...@@ -41,10 +41,10 @@ def tensor_constructor(value, name=None, strict=False, allow_downcast=None,
broadcastable = (False,) * len(value.shape) broadcastable = (False,) * len(value.shape)
type = TensorType(value.dtype, broadcastable=broadcastable) type = TensorType(value.dtype, broadcastable=broadcastable)
return TensorSharedVariable(type=type, return TensorSharedVariable(type=type,
value=numpy.array(value, copy=(not borrow)), value=numpy.array(value, copy=(not borrow)),
name=name, name=name,
strict=strict, strict=strict,
allow_downcast=allow_downcast) allow_downcast=allow_downcast)
# TensorSharedVariable brings in the tensor operators, is not ideal, but works # TensorSharedVariable brings in the tensor operators, is not ideal, but works
...@@ -85,8 +85,10 @@ def scalar_constructor(value, name=None, strict=False, allow_downcast=None, ...@@ -85,8 +85,10 @@ def scalar_constructor(value, name=None, strict=False, allow_downcast=None,
# Do not pass the dtype to asarray because we want this to fail if # Do not pass the dtype to asarray because we want this to fail if
# strict is True and the types do not match. # strict is True and the types do not match.
rval = ScalarSharedVariable(type=tensor_type, rval = ScalarSharedVariable(type=tensor_type,
value=numpy.array(value, copy=True), value=numpy.array(value, copy=True),
name=name, strict=strict, allow_downcast=allow_downcast) name=name,
strict=strict,
allow_downcast=allow_downcast)
return rval return rval
except Exception: except Exception:
traceback.print_exc() traceback.print_exc()
......
theano/tensor/slinalg.py
浏览文件 @ bd11e130
import logging import logging
logger = logging.getLogger(__name__)
import numpy
import warnings import warnings
from six.moves import xrange from six.moves import xrange
from theano.gof import Op, Apply import numpy
from theano.tensor import as_tensor_variable, dot, DimShuffle, Dot
from theano.tensor.blas import Dot22
from theano import tensor
import theano.tensor
from theano.tensor.opt import (register_stabilize,
register_specialize, register_canonicalize)
from theano.gof import local_optimizer
from theano.gof.opt import Optimizer
from theano.gradient import DisconnectedType
try: try:
import scipy.linalg import scipy.linalg
...@@ -24,6 +11,13 @@ except ImportError: ...@@ -24,6 +11,13 @@ except ImportError:
# some ops (e.g. Cholesky, Solve, A_Xinv_b) won't work # some ops (e.g. Cholesky, Solve, A_Xinv_b) won't work
imported_scipy = False imported_scipy = False
from theano import tensor
import theano.tensor
from theano.tensor import as_tensor_variable
from theano.gof import Op, Apply
logger = logging.getLogger(__name__)
MATRIX_STRUCTURES = ( MATRIX_STRUCTURES = (
'general', 'general',
'symmetric', 'symmetric',
...@@ -123,7 +117,6 @@ class CholeskyGrad(Op): ...@@ -123,7 +117,6 @@ class CholeskyGrad(Op):
F[k, k] /= (2 * L[k, k]) F[k, k] /= (2 * L[k, k])
else: else:
F = numpy.triu(dz) F = numpy.triu(dz)
M = N - 1
for k in xrange(N - 1, -1, -1): for k in xrange(N - 1, -1, -1):
for j in xrange(k + 1, N): for j in xrange(k + 1, N):
for i in xrange(j, N): for i in xrange(j, N):
...@@ -182,7 +175,7 @@ class Solve(Op): ...@@ -182,7 +175,7 @@ class Solve(Op):
else: else:
rval = scipy.linalg.solve(A, b) rval = scipy.linalg.solve(A, b)
output_storage[0][0] = rval output_storage[0][0] = rval
# computes shape of x where x = inv(A) * b # computes shape of x where x = inv(A) * b
def infer_shape(self, node, shapes): def infer_shape(self, node, shapes):
Ashape, Bshape = shapes Ashape, Bshape = shapes
......
theano/tensor/sort.py
浏览文件 @ bd11e130
...@@ -28,7 +28,7 @@ class SortOp(theano.Op): ...@@ -28,7 +28,7 @@ class SortOp(theano.Op):
def make_node(self, input, axis=-1): def make_node(self, input, axis=-1):
input = theano.tensor.as_tensor_variable(input) input = theano.tensor.as_tensor_variable(input)
if (axis is None or if (axis is None or
(isinstance(axis, theano.Constant) and axis.data is None)): (isinstance(axis, theano.Constant) and axis.data is None)):
axis = theano.Constant(theano.gof.generic, None) axis = theano.Constant(theano.gof.generic, None)
# axis=None flattens the array before sorting # axis=None flattens the array before sorting
out_type = tensor(dtype=input.dtype, broadcastable=[False]) out_type = tensor(dtype=input.dtype, broadcastable=[False])
...@@ -45,7 +45,7 @@ class SortOp(theano.Op): ...@@ -45,7 +45,7 @@ class SortOp(theano.Op):
def infer_shape(self, node, inputs_shapes): def infer_shape(self, node, inputs_shapes):
if (isinstance(node.inputs[1], theano.Constant) and if (isinstance(node.inputs[1], theano.Constant) and
node.inputs[1].data is None): node.inputs[1].data is None):
# That means axis = None, # That means axis = None,
# So the array is flattened before being sorted # So the array is flattened before being sorted
return [(mul(*inputs_shapes[0]),)] return [(mul(*inputs_shapes[0]),)]
...@@ -64,16 +64,17 @@ class SortOp(theano.Op): ...@@ -64,16 +64,17 @@ class SortOp(theano.Op):
" matrix (and axis is None or 0) and tensor3") " matrix (and axis is None or 0) and tensor3")
if a.ndim == 1: if a.ndim == 1:
idx = argsort(*inputs, kind=self.kind, order=self.order) idx = argsort(*inputs, kind=self.kind, order=self.order)
# rev_idx = numpy.where(idx[None, :]==numpy.arange(5)[:,None])[1] # rev_idx = numpy.where(idx[None, :]==numpy.arange(5)[:,None])[1]
rev_idx = theano.tensor.eq(idx[None, :], rev_idx = theano.tensor.eq(idx[None, :],
arange(a.shape[0])[:, None]).nonzero()[1] arange(a.shape[0])[:, None]).nonzero()[1]
inp_grad = output_grads[0][rev_idx] inp_grad = output_grads[0][rev_idx]
elif a.ndim == 2: elif a.ndim == 2:
if (axis is None or if (axis is None or
(isinstance(axis, theano.Constant) and axis.data is None)): (isinstance(axis, theano.Constant) and axis.data is None)):
idx = argsort(*inputs, kind=self.kind, order=self.order) idx = argsort(*inputs, kind=self.kind, order=self.order)
rev_idx = theano.tensor.eq(idx[None, :], rev_idx = theano.tensor.eq(
arange(a.shape[0]*a.shape[1])[:, None]).nonzero()[1] idx[None, :],
arange(a.shape[0] * a.shape[1])[:, None]).nonzero()[1]
inp_grad = output_grads[0][rev_idx].reshape(a.shape) inp_grad = output_grads[0][rev_idx].reshape(a.shape)
elif (axis == 0 or elif (axis == 0 or
(isinstance(axis, theano.Constant) and axis.data == 0)): (isinstance(axis, theano.Constant) and axis.data == 0)):
...@@ -85,7 +86,7 @@ class SortOp(theano.Op): ...@@ -85,7 +86,7 @@ class SortOp(theano.Op):
indices = self.__get_argsort_indices(a, axis) indices = self.__get_argsort_indices(a, axis)
inp_grad = output_grads[0][indices[0], indices[1], indices[2]] inp_grad = output_grads[0][indices[0], indices[1], indices[2]]
elif (axis is None or elif (axis is None or
(isinstance(axis, theano.Constant) and axis.data is None)): (isinstance(axis, theano.Constant) and axis.data is None)):
rev_idx = self.__get_argsort_indices(a, axis) rev_idx = self.__get_argsort_indices(a, axis)
inp_grad = output_grads[0][rev_idx].reshape(a.shape) inp_grad = output_grads[0][rev_idx].reshape(a.shape)
axis_grad = theano.gradient.grad_undefined( axis_grad = theano.gradient.grad_undefined(
...@@ -103,13 +104,13 @@ class SortOp(theano.Op): ...@@ -103,13 +104,13 @@ class SortOp(theano.Op):
list of lenght len(a.shape) otherwise list of lenght len(a.shape) otherwise
""" """
# The goal is to get gradient wrt input from gradient # The goal is to get gradient wrt input from gradient
# wrt sort(input, axis) # wrt sort(input, axis)
idx = argsort(a, axis, kind=self.kind, order=self.order) idx = argsort(a, axis, kind=self.kind, order=self.order)
# rev_idx is the reverse of previous argsort operation # rev_idx is the reverse of previous argsort operation
rev_idx = argsort(idx, axis, kind=self.kind, order=self.order) rev_idx = argsort(idx, axis, kind=self.kind, order=self.order)
if (axis is None or if (axis is None or
(isinstance(axis, theano.Constant) and axis.data is None)): (isinstance(axis, theano.Constant) and axis.data is None)):
return rev_idx return rev_idx
indices = [] indices = []
if axis.data >= 0: if axis.data >= 0:
...@@ -120,7 +121,7 @@ class SortOp(theano.Op): ...@@ -120,7 +121,7 @@ class SortOp(theano.Op):
if i == axis_data: if i == axis_data:
indices.append(rev_idx) indices.append(rev_idx)
else: else:
index_shape = [1] * a.ndim index_shape = [1] * a.ndim
index_shape[i] = a.shape[i] index_shape[i] = a.shape[i]
# it's a way to emulate numpy.ogrid[0: a.shape[0], 0: a.shape[1], 0: a.shape[2]] # it's a way to emulate numpy.ogrid[0: a.shape[0], 0: a.shape[1], 0: a.shape[2]]
indices.append(theano.tensor.arange(a.shape[i]).reshape(index_shape)) indices.append(theano.tensor.arange(a.shape[i]).reshape(index_shape))
...@@ -178,28 +179,27 @@ class ArgSortOp(theano.Op): ...@@ -178,28 +179,27 @@ class ArgSortOp(theano.Op):
return hash(type(self)) ^ hash(self.order) ^ hash(self.kind) return hash(type(self)) ^ hash(self.order) ^ hash(self.kind)
def __str__(self): def __str__(self):
return (self.__class__.__name__ return (self.__class__.__name__ +
+ "{%s, %s}" % (self.kind, str(self.order))) "{%s, %s}" % (self.kind, str(self.order)))
def make_node(self, input, axis=-1): def make_node(self, input, axis=-1):
input = theano.tensor.as_tensor_variable(input) input = theano.tensor.as_tensor_variable(input)
if (axis is None or if (axis is None or
(isinstance(axis, theano.Constant) and axis.data is None)): (isinstance(axis, theano.Constant) and axis.data is None)):
axis = theano.Constant(theano.gof.generic, None) axis = theano.Constant(theano.gof.generic, None)
bcast = [False] bcast = [False]
else: else:
axis = theano.tensor.as_tensor_variable(axis) axis = theano.tensor.as_tensor_variable(axis)
bcast = input.type.broadcastable bcast = input.type.broadcastable
return theano.Apply(self, [input, axis], return theano.Apply(self, [input, axis], [theano.tensor.TensorType(
[theano.tensor.TensorType(dtype="int64", broadcastable=bcast)()]) dtype="int64", broadcastable=bcast)()])
def perform(self, node, inputs, output_storage): def perform(self, node, inputs, output_storage):
a = inputs[0] a = inputs[0]
axis = inputs[1] axis = inputs[1]
z = output_storage[0] z = output_storage[0]
z[0] = theano._asarray( z[0] = theano._asarray(np.argsort(a, axis, self.kind, self.order),
np.argsort(a, axis, self.kind, self.order), dtype=node.outputs[0].dtype)
dtype=node.outputs[0].dtype)
def infer_shape(self, node, inputs_shapes): def infer_shape(self, node, inputs_shapes):
if (isinstance(node.inputs[1], theano.Constant) and if (isinstance(node.inputs[1], theano.Constant) and
......
theano/tensor/subtensor.py
浏览文件 @ bd11e130
from copy import copy from copy import copy
import os
import sys import sys
from textwrap import dedent from textwrap import dedent
import warnings import warnings
import logging import logging
_logger = logging.getLogger("theano.tensor.subtensor")
import numpy import numpy
from six.moves import xrange from six.moves import xrange
...@@ -32,6 +30,7 @@ if config.cxx: ...@@ -32,6 +30,7 @@ if config.cxx:
except ImportError: except ImportError:
pass pass
_logger = logging.getLogger("theano.tensor.subtensor")
# Do a lazy import of the sparse module # Do a lazy import of the sparse module
sparse_module_ref = None sparse_module_ref = None
...@@ -336,9 +335,9 @@ class Subtensor(Op): ...@@ -336,9 +335,9 @@ class Subtensor(Op):
theano.tensor.wscalar, theano.tensor.bscalar] theano.tensor.wscalar, theano.tensor.bscalar]
invalid_tensor_types = [theano.tensor.fscalar, theano.tensor.dscalar, invalid_tensor_types = [theano.tensor.fscalar, theano.tensor.dscalar,
theano.tensor.cscalar, theano.tensor.zscalar] theano.tensor.cscalar, theano.tensor.zscalar]
if (isinstance(entry, gof.Variable) if (isinstance(entry, gof.Variable) and
and (entry.type in invalid_scal_types (entry.type in invalid_scal_types or
or entry.type in invalid_tensor_types)): entry.type in invalid_tensor_types)):
raise TypeError("Expected an integer") raise TypeError("Expected an integer")
if isinstance(entry, gof.Variable) and entry.type in scal_types: if isinstance(entry, gof.Variable) and entry.type in scal_types:
...@@ -346,13 +345,13 @@ class Subtensor(Op): ...@@ -346,13 +345,13 @@ class Subtensor(Op):
elif isinstance(entry, gof.Type) and entry in scal_types: elif isinstance(entry, gof.Type) and entry in scal_types:
return entry return entry
if (isinstance(entry, gof.Variable) if (isinstance(entry, gof.Variable) and
and entry.type in tensor_types entry.type in tensor_types and
and numpy.all(entry.type.broadcastable)): numpy.all(entry.type.broadcastable)):
return scal.get_scalar_type(entry.type.dtype) return scal.get_scalar_type(entry.type.dtype)
elif (isinstance(entry, gof.Type) elif (isinstance(entry, gof.Type) and
and entry in tensor_types entry in tensor_types and
and numpy.all(entry.broadcastable)): numpy.all(entry.broadcastable)):
return scal.get_scalar_type(entry.dtype) return scal.get_scalar_type(entry.dtype)
elif slice_ok and isinstance(entry, slice): elif slice_ok and isinstance(entry, slice):
a = entry.start a = entry.start
...@@ -425,8 +424,9 @@ class Subtensor(Op): ...@@ -425,8 +424,9 @@ class Subtensor(Op):
conv(val.step)) conv(val.step))
else: else:
try: try:
return get_scalar_constant_value(val, return get_scalar_constant_value(
only_process_constants=only_process_constants) val,
only_process_constants=only_process_constants)
except theano.tensor.NotScalarConstantError: except theano.tensor.NotScalarConstantError:
if allow_partial: if allow_partial:
return val return val
...@@ -477,8 +477,8 @@ class Subtensor(Op): ...@@ -477,8 +477,8 @@ class Subtensor(Op):
% (input.type, expected_type)) % (input.type, expected_type))
# infer the broadcasting pattern # infer the broadcasting pattern
padded = (self.get_constant_idx((None,)+inputs, allow_partial=True) padded = (self.get_constant_idx((None,) + inputs, allow_partial=True) +
+ [slice(None, None, None)] * (x.type.ndim - len(idx_list))) [slice(None, None, None)] * (x.type.ndim - len(idx_list)))
broadcastable = [] broadcastable = []
for i, (p, bc) in enumerate(izip(padded, x.type.broadcastable)): for i, (p, bc) in enumerate(izip(padded, x.type.broadcastable)):
if isinstance(p, slice): if isinstance(p, slice):
...@@ -528,9 +528,9 @@ class Subtensor(Op): ...@@ -528,9 +528,9 @@ class Subtensor(Op):
if isinstance(idx, slice): if isinstance(idx, slice):
# If it is the default (None, None, None) slice, or a variant, # If it is the default (None, None, None) slice, or a variant,
# the shape will be xl # the shape will be xl
if ((idx.start in [None, 0]) if ((idx.start in [None, 0]) and
and (idx.stop in [None, sys.maxsize]) (idx.stop in [None, sys.maxsize]) and
and (idx.step is None or idx.step == 1)): (idx.step is None or idx.step == 1)):
outshp.append(xl) outshp.append(xl)
else: else:
cnf = get_canonical_form_slice(idx, xl)[0] cnf = get_canonical_form_slice(idx, xl)[0]
...@@ -556,8 +556,7 @@ class Subtensor(Op): ...@@ -556,8 +556,7 @@ class Subtensor(Op):
first = x.zeros_like().astype(theano.config.floatX) first = x.zeros_like().astype(theano.config.floatX)
else: else:
first = IncSubtensor(self.idx_list)(x.zeros_like(), gz, *rest) first = IncSubtensor(self.idx_list)(x.zeros_like(), gz, *rest)
return ([first] return ([first] + [DisconnectedType()()] * len(rest))
+ [DisconnectedType()()] * len(rest))
def connection_pattern(self, node): def connection_pattern(self, node):
...@@ -1034,8 +1033,7 @@ def inc_subtensor(x, y, inplace=False, set_instead_of_inc=False, ...@@ -1034,8 +1033,7 @@ def inc_subtensor(x, y, inplace=False, set_instead_of_inc=False,
dim_offset = x.ndim - y.ndim dim_offset = x.ndim - y.ndim
for dim in xrange(y.ndim): for dim in xrange(y.ndim):
if (x.broadcastable[dim + dim_offset] if (x.broadcastable[dim + dim_offset] and not y.broadcastable[dim]):
and not y.broadcastable[dim]):
# It is acceptable to try to increment a subtensor with a # It is acceptable to try to increment a subtensor with a
# broadcastable dim with a tensor that is not broadcastable # broadcastable dim with a tensor that is not broadcastable
# on that dimension. However, its length must then be 1. # on that dimension. However, its length must then be 1.
...@@ -2133,9 +2131,9 @@ class AdvancedIncSubtensor(Op): ...@@ -2133,9 +2131,9 @@ class AdvancedIncSubtensor(Op):
return hash((type(self), self.inplace, self.set_instead_of_inc)) return hash((type(self), self.inplace, self.set_instead_of_inc))
def __eq__(self, other): def __eq__(self, other):
return (type(self) == type(other) return (type(self) == type(other) and
and self.inplace == other.inplace self.inplace == other.inplace and
and self.set_instead_of_inc == other.set_instead_of_inc) self.set_instead_of_inc == other.set_instead_of_inc)
def __str__(self): def __str__(self):
return "%s{%s, %s}" % (self.__class__.__name__, return "%s{%s, %s}" % (self.__class__.__name__,
......
theano/tensor/utils.py
浏览文件 @ bd11e130
...@@ -79,11 +79,11 @@ def shape_of_variables(fgraph, input_shapes): ...@@ -79,11 +79,11 @@ def shape_of_variables(fgraph, input_shapes):
if not hasattr(fgraph, 'shape_feature'): if not hasattr(fgraph, 'shape_feature'):
fgraph.attach_feature(theano.tensor.opt.ShapeFeature()) fgraph.attach_feature(theano.tensor.opt.ShapeFeature())
input_dims = [dimension for inp in fgraph.inputs input_dims = [dimension for inp in fgraph.inputs
for dimension in fgraph.shape_feature.shape_of[inp]] for dimension in fgraph.shape_feature.shape_of[inp]]
output_dims = [dimension for shape in fgraph.shape_feature.shape_of.values() output_dims = [dimension for shape in fgraph.shape_feature.shape_of.values()
for dimension in shape] for dimension in shape]
compute_shapes = theano.function(input_dims, output_dims) compute_shapes = theano.function(input_dims, output_dims)
...@@ -93,8 +93,8 @@ def shape_of_variables(fgraph, input_shapes): ...@@ -93,8 +93,8 @@ def shape_of_variables(fgraph, input_shapes):
" interface changed. Now by default, it clones the graph it receives." " interface changed. Now by default, it clones the graph it receives."
" To have the old behavior, give it this new parameter `clone=False`.") " To have the old behavior, give it this new parameter `clone=False`.")
numeric_input_dims = [dim for inp in fgraph.inputs numeric_input_dims = [dim for inp in fgraph.inputs
for dim in input_shapes[inp]] for dim in input_shapes[inp]]
numeric_output_dims = compute_shapes(*numeric_input_dims) numeric_output_dims = compute_shapes(*numeric_input_dims)
sym_to_num_dict = dict(izip(output_dims, numeric_output_dims)) sym_to_num_dict = dict(izip(output_dims, numeric_output_dims))
......
theano/tensor/var.py
浏览文件 @ bd11e130
import copy import copy
import pdb
import sys
import traceback as tb import traceback as tb
import warnings import warnings
...@@ -41,9 +39,9 @@ class _tensor_py_operators: ...@@ -41,9 +39,9 @@ class _tensor_py_operators:
# CASTS # CASTS
# REMOVED THESE BECAUSE PYTHON appears to require __int__ to return # REMOVED THESE BECAUSE PYTHON appears to require __int__ to return
# an int. -JB 20081112 # an int. -JB 20081112
#def __int__(self): return convert_to_int32(self) # def __int__(self): return convert_to_int32(self)
#def __float__(self): return convert_to_float64(self) # def __float__(self): return convert_to_float64(self)
#def __complex__(self): return convert_to_complex128(self) # def __complex__(self): return convert_to_complex128(self)
# COMPARISONS # COMPARISONS
_is_nonzero = True _is_nonzero = True
...@@ -68,7 +66,6 @@ class _tensor_py_operators: ...@@ -68,7 +66,6 @@ class _tensor_py_operators:
rval._is_nonzero = False rval._is_nonzero = False
return rval return rval
def __nonzero__(self): def __nonzero__(self):
# Python 2.x # Python 2.x
return self.__bool__() return self.__bool__()
...@@ -215,7 +212,7 @@ class _tensor_py_operators: ...@@ -215,7 +212,7 @@ class _tensor_py_operators:
# DO NOT USE THESE BECAUSE INPLACE OPS SHOULD BE INSERTED # DO NOT USE THESE BECAUSE INPLACE OPS SHOULD BE INSERTED
# BY OPTIMIZATIONS ONLY # BY OPTIMIZATIONS ONLY
## ARITHMETIC - INPLACE # ARITHMETIC - INPLACE
# def __iadd__(self, other): # def __iadd__(self, other):
# return _add_inplace(self, other) # return _add_inplace(self, other)
# def __isub__(self, other): # def __isub__(self, other):
...@@ -642,7 +639,8 @@ class TensorVariable(_tensor_py_operators, Variable): ...@@ -642,7 +639,8 @@ class TensorVariable(_tensor_py_operators, Variable):
elif config.warn_float64 == "raise": elif config.warn_float64 == "raise":
raise Exception(msg) raise Exception(msg)
elif config.warn_float64 == 'pdb': elif config.warn_float64 == 'pdb':
import pdb; pdb.set_trace() import pdb
pdb.set_trace()
TensorType.Variable = TensorVariable TensorType.Variable = TensorVariable
...@@ -744,8 +742,8 @@ class TensorConstant(_tensor_py_operators, Constant): ...@@ -744,8 +742,8 @@ class TensorConstant(_tensor_py_operators, Constant):
def __init__(self, type, data, name=None): def __init__(self, type, data, name=None):
Constant.__init__(self, type, data, name) Constant.__init__(self, type, data, name)
if (isinstance(data, numpy.ndarray) and if (isinstance(data, numpy.ndarray) and
data.ndim > 0 and data.ndim > 0 and
len(numpy.unique(data)) == 1): len(numpy.unique(data)) == 1):
self.tag.unique_value = numpy.unique(data)[0] self.tag.unique_value = numpy.unique(data)[0]
else: else:
self.tag.unique_value = None self.tag.unique_value = None
......
theano/tensor/xlogx.py
浏览文件 @ bd11e130
...@@ -13,12 +13,15 @@ class XlogX(scalar.UnaryScalarOp): ...@@ -13,12 +13,15 @@ class XlogX(scalar.UnaryScalarOp):
if x == 0.0: if x == 0.0:
return 0.0 return 0.0
return x * numpy.log(x) return x * numpy.log(x)
def impl(self, x): def impl(self, x):
return XlogX.st_impl(x) return XlogX.st_impl(x)
def grad(self, inputs, grads): def grad(self, inputs, grads):
x, = inputs x, = inputs
gz, = grads gz, = grads
return [gz * (1 + scalar.log(x))] return [gz * (1 + scalar.log(x))]
def c_code(self, node, name, inputs, outputs, sub): def c_code(self, node, name, inputs, outputs, sub):
x, = inputs x, = inputs
z, = outputs z, = outputs
...@@ -28,7 +31,8 @@ class XlogX(scalar.UnaryScalarOp): ...@@ -28,7 +31,8 @@ class XlogX(scalar.UnaryScalarOp):
? 0.0 ? 0.0
: %(x)s * log(%(x)s);""" % locals() : %(x)s * log(%(x)s);""" % locals()
raise NotImplementedError('only floatingpoint is implemented') raise NotImplementedError('only floatingpoint is implemented')
scalar_xlogx = XlogX(scalar.upgrade_to_float, name='scalar_xlogx')
scalar_xlogx = XlogX(scalar.upgrade_to_float, name='scalar_xlogx')
xlogx = Elemwise(scalar_xlogx, name='xlogx') xlogx = Elemwise(scalar_xlogx, name='xlogx')
...@@ -41,12 +45,15 @@ class XlogY0(scalar.BinaryScalarOp): ...@@ -41,12 +45,15 @@ class XlogY0(scalar.BinaryScalarOp):
if x == 0.0: if x == 0.0:
return 0.0 return 0.0
return x * numpy.log(y) return x * numpy.log(y)
def impl(self, x, y): def impl(self, x, y):
return XlogY0.st_impl(x, y) return XlogY0.st_impl(x, y)
def grad(self, inputs, grads): def grad(self, inputs, grads):
x, y = inputs x, y = inputs
gz, = grads gz, = grads
return [gz * scalar.log(y), gz * x / y] return [gz * scalar.log(y), gz * x / y]
def c_code(self, node, name, inputs, outputs, sub): def c_code(self, node, name, inputs, outputs, sub):
x, y = inputs x, y = inputs
z, = outputs z, = outputs
...@@ -56,5 +63,6 @@ class XlogY0(scalar.BinaryScalarOp): ...@@ -56,5 +63,6 @@ class XlogY0(scalar.BinaryScalarOp):
? 0.0 ? 0.0
: %(x)s * log(%(y)s);""" % locals() : %(x)s * log(%(y)s);""" % locals()
raise NotImplementedError('only floatingpoint is implemented') raise NotImplementedError('only floatingpoint is implemented')
scalar_xlogy0 = XlogY0(scalar.upgrade_to_float, name='scalar_xlogy0')
scalar_xlogy0 = XlogY0(scalar.upgrade_to_float, name='scalar_xlogy0')
xlogy0 = Elemwise(scalar_xlogy0, name='xlogy0') xlogy0 = Elemwise(scalar_xlogy0, name='xlogy0')
theano/tests/test_flake8.py
浏览文件 @ bd11e130
...@@ -57,30 +57,17 @@ whitelist_flake8 = [ ...@@ -57,30 +57,17 @@ whitelist_flake8 = [
"typed_list/tests/test_type.py", "typed_list/tests/test_type.py",
"typed_list/tests/test_opt.py", "typed_list/tests/test_opt.py",
"typed_list/tests/test_basic.py", "typed_list/tests/test_basic.py",
"tensor/var.py",
"tensor/sharedvar.py",
"tensor/inplace.py",
"tensor/slinalg.py",
"tensor/shared_randomstreams.py",
"tensor/subtensor.py",
"tensor/elemwise.py",
"tensor/xlogx.py",
"tensor/blas_headers.py", "tensor/blas_headers.py",
"tensor/utils.py",
"tensor/type.py", "tensor/type.py",
"tensor/fourier.py", "tensor/fourier.py",
"tensor/sort.py",
"tensor/__init__.py", "tensor/__init__.py",
"tensor/opt_uncanonicalize.py", "tensor/opt_uncanonicalize.py",
"tensor/opt.py",
"tensor/blas.py", "tensor/blas.py",
"tensor/extra_ops.py", "tensor/extra_ops.py",
"tensor/nlinalg.py", "tensor/nlinalg.py",
"tensor/blas_c.py", "tensor/blas_c.py",
"tensor/elemwise_cgen.py", "tensor/elemwise_cgen.py",
"tensor/raw_random.py",
"tensor/blas_scipy.py", "tensor/blas_scipy.py",
"tensor/basic.py",
"tensor/tests/test_subtensor.py", "tensor/tests/test_subtensor.py",
"tensor/tests/test_utils.py", "tensor/tests/test_utils.py",
"tensor/tests/test_nlinalg.py", "tensor/tests/test_nlinalg.py",
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论