Skip to content

P
pytensor
提交 d7e4f9d2 authored 作者: Frederic's avatar Frederic
浏览文件
操作

pep8

上级 4d5e8de0
显示空白字符变更
内嵌 并排
正在显示 包含 412 行增加 和 229 行删除
theano/tensor/basic.py
浏览文件 @ d7e4f9d2
...@@ -5,13 +5,13 @@ __docformat__ = "restructuredtext en" ...@@ -5,13 +5,13 @@ __docformat__ = "restructuredtext en"
import __builtin__ import __builtin__
import sys # for sys.maxint import sys # for sys.maxint
from theano.configparser import config from theano.configparser import config
import traceback #for overriding Op.__call__
import warnings import warnings
from itertools import izip from itertools import izip
import numpy, theano import numpy
#from copy import copy as python_copy #from copy import copy as python_copy
import theano
from theano import gof from theano import gof
from theano.gof import Apply, Constant, Op, Type, Value, Variable from theano.gof import Apply, Constant, Op, Type, Value, Variable
...@@ -30,7 +30,7 @@ import theano.scalar.sharedvar ...@@ -30,7 +30,7 @@ import theano.scalar.sharedvar
from elemwise import Elemwise, DimShuffle, CAReduce, Sum from elemwise import Elemwise, DimShuffle, CAReduce, Sum
import logging import logging
_logger=logging.getLogger("theano.tensor.basic") _logger = logging.getLogger("theano.tensor.basic")
#This is needed as we will hide it later #This is needed as we will hide it later
python_complex = complex python_complex = complex
...@@ -55,14 +55,18 @@ def check_equal_numpy(x, y): ...@@ -55,14 +55,18 @@ def check_equal_numpy(x, y):
shape if x and y are numpy.ndarray instances). shape if x and y are numpy.ndarray instances).
""" """
if isinstance(x, numpy.ndarray) and isinstance(y, numpy.ndarray): if isinstance(x, numpy.ndarray) and isinstance(y, numpy.ndarray):
return x.dtype == y.dtype and x.shape == y.shape and numpy.any(abs(x - y) < 1e-10) return (x.dtype == y.dtype and x.shape == y.shape and
elif isinstance(x, numpy.random.RandomState) and isinstance(y, numpy.random.RandomState): numpy.any(abs(x - y) < 1e-10))
return python_all(numpy.all(a==b) for a, b in zip(x.__getstate__(), y.__getstate__())) elif (isinstance(x, numpy.random.RandomState) and
isinstance(y, numpy.random.RandomState)):
return python_all(numpy.all(a == b) for a, b in
zip(x.__getstate__(), y.__getstate__()))
else: else:
return x == y return x == y
compile.register_checker(check_equal_numpy) compile.register_checker(check_equal_numpy)
def hashtype(self): def hashtype(self):
t = type(self) t = type(self)
return hash(t.__name__) ^ hash(t.__module__) return hash(t.__name__) ^ hash(t.__module__)
...@@ -70,14 +74,20 @@ elemwise.hashtype = hashtype ...@@ -70,14 +74,20 @@ elemwise.hashtype = hashtype
__oplist_constructor_list = [] __oplist_constructor_list = []
"""List of functions to be listed as op constructors in the oplist (`gen_oplist`, doc/oplist.txt).""" """List of functions to be listed as op constructors in the oplist
(`gen_oplist`, doc/oplist.txt)."""
def constructor(f): def constructor(f):
"""Add `f` to :doc:`oplist`. """Add `f` to :doc:`oplist`.
Make `f` appear as a constructor in the oplist (`gen_oplist`, doc/oplist.txt). Make `f` appear as a constructor in the oplist (`gen_oplist`,
doc/oplist.txt).
""" """
__oplist_constructor_list.append(f) __oplist_constructor_list.append(f)
return f return f
def __oplist_tag(thing, tag): def __oplist_tag(thing, tag):
tags = getattr(thing, '__oplist_tags', []) tags = getattr(thing, '__oplist_tags', [])
tags.append(tag) tags.append(tag)
...@@ -89,30 +99,36 @@ if 0: ...@@ -89,30 +99,36 @@ if 0:
# the one place where this is used we should also allow for sparse # the one place where this is used we should also allow for sparse
# variables # variables
# - JB 20100226 # - JB 20100226
def as_cuda_or_tensor_variable(x, name = None, ndim=None): def as_cuda_or_tensor_variable(x, name=None, ndim=None):
""" """
This function do the same as_tensor_variable, but don't transfert the value on the gpu This function do the same as_tensor_variable, but don't
transfert the value on the gpu
""" """
if hasattr(x, '_as_CudaNdarrayVariable'): if hasattr(x, '_as_CudaNdarrayVariable'):
return x._as_CudaNdarrayVariable() #TODO: pass name and ndim arguments #TODO: pass name and ndim arguments
return x._as_CudaNdarrayVariable()
return as_tensor_variable(x, name, ndim) return as_tensor_variable(x, name, ndim)
def as_tensor_variable(x, name=None, ndim=None): def as_tensor_variable(x, name=None, ndim=None):
"""Return `x`, transformed into a `TensorType` """Return `x`, transformed into a `TensorType`
This function is often used by `make_node` methods of `Op` subclasses to This function is often used by `make_node` methods of `Op`
turn ndarrays, numbers, `Scalar` instances, `Apply` instances and `TensorType` subclasses to turn ndarrays, numbers, `Scalar` instances, `Apply`
instances into valid input list elemnts. instances and `TensorType` instances into valid input list
elemnts.
:Parameters: :Parameters:
- `x`: Apply instance, Variable instance, numpy.ndarray, or number - `x`: Apply instance, Variable instance, numpy.ndarray, or number
This thing will be transformed into a `Variable` in a sensible way. An This thing will be transformed into a `Variable` in a sensible way. An
ndarray argument will not be copied, but a list of numbers will be copied ndarray argument will not be copied, but a list of numbers will be
to make an ndarray. copied to make an ndarray.
- `name`: str or None - `name`: str or None
If a new `Variable` instance is created, it will be named with this string. If a new `Variable` instance is created, it will be named with this
string.
- `ndim`: None or integer - `ndim`: None or integer
Return a Variable with this many dimensions. Raise TypeError if it's not possible. Return a Variable with this many dimensions. Raise TypeError if it's
not possible.
:Exceptions: :Exceptions:
- `ValueError`: raised if an `Apply` with no default output is fetched - `ValueError`: raised if an `Apply` with no default output is fetched
...@@ -120,12 +136,14 @@ def as_tensor_variable(x, name=None, ndim=None): ...@@ -120,12 +136,14 @@ def as_tensor_variable(x, name=None, ndim=None):
""" """
if hasattr(x, '_as_TensorVariable'): if hasattr(x, '_as_TensorVariable'):
return x._as_TensorVariable() #TODO: pass name and ndim arguments return x._as_TensorVariable() # TODO: pass name and ndim arguments
if isinstance(x, gof.Apply): if isinstance(x, gof.Apply):
#TODO: use Apply's default output mechanism #TODO: use Apply's default output mechanism
if len(x.outputs) != 1: if len(x.outputs) != 1:
raise ValueError("It is ambiguous which output of a multi-output Op has to be fetched.", x) raise ValueError(
"It is ambiguous which output of a multi-output Op has"
" to be fetched.", x)
else: else:
x = x.outputs[0] x = x.outputs[0]
if isinstance(x, Variable): if isinstance(x, Variable):
...@@ -133,28 +151,34 @@ def as_tensor_variable(x, name=None, ndim=None): ...@@ -133,28 +151,34 @@ def as_tensor_variable(x, name=None, ndim=None):
x = tensor_from_scalar(x) x = tensor_from_scalar(x)
if not isinstance(x.type, TensorType): if not isinstance(x.type, TensorType):
raise TypeError("Variable type field must be a TensorType.", x, x.type) raise TypeError(
"Variable type field must be a TensorType.", x, x.type)
if ndim is None: if ndim is None:
return x return x
else: else:
if (x.type.ndim > ndim): if (x.type.ndim > ndim):
#TODO: strip off leading broadcastable dimensions #TODO: strip off leading broadcastable dimensions
raise ValueError('TensorType could not be cast to have %i dimensions' % ndim, x.type) raise ValueError(
'TensorType could not be cast to have %i dimensions' %
ndim, x.type)
elif (x.type.ndim < ndim): elif (x.type.ndim < ndim):
return shape_padleft(x, n_ones=(ndim - x.type.ndim)) return shape_padleft(x, n_ones=(ndim - x.type.ndim))
else: else:
return x return x
if isinstance(x, (tuple, list)) and python_any(isinstance(xi, Variable) for xi in x): if isinstance(x, (tuple, list)) and python_any(isinstance(xi, Variable)
for xi in x):
try: try:
return stack(*x) return stack(*x)
except (TypeError, ValueError): except (TypeError, ValueError):
pass pass
if isinstance(x, bool): if isinstance(x, bool):
raise TypeError("Cannot cast True or False as a tensor variable. Please use 1 or 0. " raise TypeError(
"This error might be caused by using the == operator on Variables. " "Cannot cast True or False as a tensor variable. Please use 1 or "
"v == w does not do what you think it does, use theano.tensor.eq(v, w) instead.") "0. This error might be caused by using the == operator on "
"Variables. v == w does not do what you think it does, "
"use theano.tensor.eq(v, w) instead.")
try: try:
return constant(x, name=name, ndim=ndim) return constant(x, name=name, ndim=ndim)
...@@ -165,12 +189,15 @@ def as_tensor_variable(x, name=None, ndim=None): ...@@ -165,12 +189,15 @@ def as_tensor_variable(x, name=None, ndim=None):
str_x = repr(x) str_x = repr(x)
raise TypeError("Cannot convert %s to TensorType" % str_x, type(x)) raise TypeError("Cannot convert %s to TensorType" % str_x, type(x))
# this has a different name, because _as_tensor_variable is the function which ops use # this has a different name, because _as_tensor_variable is the
# to upcast their arguments... this internal-use function is a good place to put debugging stuff, better than the global astensor. # function which ops use to upcast their arguments... this
# internal-use function is a good place to put debugging stuff, better
# than the global astensor.
_as_tensor_variable = as_tensor_variable _as_tensor_variable = as_tensor_variable
as_tensor = as_tensor_variable as_tensor = as_tensor_variable
class NumpyAutocaster(object): class NumpyAutocaster(object):
""" """
This class is used to cast python ints and floats to numpy arrays. This class is used to cast python ints and floats to numpy arrays.
...@@ -253,15 +280,19 @@ class NumpyAutocaster(object): ...@@ -253,15 +280,19 @@ class NumpyAutocaster(object):
autocast_int = NumpyAutocaster(('int8', 'int16', 'int32', 'int64')) autocast_int = NumpyAutocaster(('int8', 'int16', 'int32', 'int64'))
autocast_float = NumpyAutocaster(('float32', 'float64')) autocast_float = NumpyAutocaster(('float32', 'float64'))
# autocast_float dtypes might be manipulated in tensor.__init__ # autocast_float dtypes might be manipulated in tensor.__init__
# #
# Note: it's a bit weird for a compiler to automatically downcast literals like this, and it might # Note: it's a bit weird for a compiler to automatically downcast
# have implications for efficiency when mixing types. For example when you add 1.0 + # literals like this, and it might have implications for efficiency
# dmatrix(), the 1.0 could be converted to float32, and require upcasting for the + operation # when mixing types. For example when you add 1.0 + dmatrix(), the
# at every position in the dmatrix. using theano._asarray(1.0, dtype='float64') will circumvent # 1.0 could be converted to float32, and require upcasting for the +
# this autocasting, and in future, our ops might be smarter about factoring out upcasts. The # operation at every position in the dmatrix. using
# advantage of this mechanism is to combine it with floatX so that 1.0 + xmatrix() will always # theano._asarray(1.0, dtype='float64') will circumvent this
# have the same type as the xmatrix(). # autocasting, and in future, our ops might be smarter about factoring
# out upcasts. The advantage of this mechanism is to combine it with
# floatX so that 1.0 + xmatrix() will always have the same type as the
# xmatrix().
# #
class autocast_float_as(object): class autocast_float_as(object):
""" """
...@@ -274,20 +305,23 @@ class autocast_float_as(object): ...@@ -274,20 +305,23 @@ class autocast_float_as(object):
>>> assert (fvector() + 1.1).dtype == 'float32' # temporary downcasting >>> assert (fvector() + 1.1).dtype == 'float32' # temporary downcasting
>>> assert (fvector() + 1.1).dtype == 'float64' # back to default behaviour >>> assert (fvector() + 1.1).dtype == 'float64' # back to default behaviour
This class might be convenient in some code, but it definitely helps to test the This class might be convenient in some code, but it definitely
autocasting mechanism. helps to test the autocasting mechanism.
""" """
def __init__(self, *dtypes): def __init__(self, *dtypes):
self.dtypes = dtypes self.dtypes = dtypes
assert config.cast_policy == 'custom' assert config.cast_policy == 'custom'
def __enter__(self): def __enter__(self):
assert config.cast_policy == 'custom' assert config.cast_policy == 'custom'
self.old_dtypes = autocast_float.dtypes self.old_dtypes = autocast_float.dtypes
autocast_float.dtypes = self.dtypes autocast_float.dtypes = self.dtypes
def __exit__(self, *args): def __exit__(self, *args):
assert config.cast_policy == 'custom' assert config.cast_policy == 'custom'
autocast_float.dtypes = self.old_dtypes autocast_float.dtypes = self.old_dtypes
def constant_or_value(x, rtype, name=None, ndim=None, dtype=None): def constant_or_value(x, rtype, name=None, ndim=None, dtype=None):
"""Return a symbolic `Constant` with value `x` """Return a symbolic `Constant` with value `x`
...@@ -330,28 +364,35 @@ def constant_or_value(x, rtype, name=None, ndim=None, dtype=None): ...@@ -330,28 +364,35 @@ def constant_or_value(x, rtype, name=None, ndim=None, dtype=None):
bcastable = [True] * (ndim - len(bcastable)) + bcastable bcastable = [True] * (ndim - len(bcastable)) + bcastable
elif len(bcastable) > ndim: elif len(bcastable) > ndim:
#TODO: strip off dimensions of size 1 #TODO: strip off dimensions of size 1
raise ValueError('ndarray could not be cast to constant with %i dimensions' % ndim) raise ValueError(
'ndarray could not be cast to constant with %i dimensions' %
ndim)
assert len(bcastable) == ndim assert len(bcastable) == ndim
try: try:
if rtype is TensorConstant: if rtype is TensorConstant:
rval = rtype( rval = rtype(
TensorType(dtype = x_.dtype, broadcastable = bcastable), TensorType(dtype=x_.dtype, broadcastable=bcastable),
x_.copy(), x_.copy(),
name=name) name=name)
return rval return rval
else: else:
# leave the shape out of the type # leave the shape out of the type
return rtype(TensorType(dtype = x_.dtype, broadcastable = bcastable), x_, name=name) return rtype(TensorType(dtype=x_.dtype, broadcastable=bcastable),
x_, name=name)
except Exception: except Exception:
raise TypeError("Could not convert %s to TensorType" % x, type(x)) raise TypeError("Could not convert %s to TensorType" % x, type(x))
def constant(x, name=None, ndim=None, dtype=None): def constant(x, name=None, ndim=None, dtype=None):
return constant_or_value(x, rtype=TensorConstant, name=name, ndim=ndim, return constant_or_value(x, rtype=TensorConstant, name=name, ndim=ndim,
dtype=dtype) dtype=dtype)
def value(x, name=None, ndim=None, dtype=None): def value(x, name=None, ndim=None, dtype=None):
return constant_or_value(x, rtype=TensorValue, name=name, ndim=ndim, dtype=dtype) return constant_or_value(x, rtype=TensorValue, name=name,
ndim=ndim, dtype=dtype)
def _obj_is_wrappable_as_tensor(x): def _obj_is_wrappable_as_tensor(x):
try: try:
...@@ -359,11 +400,15 @@ def _obj_is_wrappable_as_tensor(x): ...@@ -359,11 +400,15 @@ def _obj_is_wrappable_as_tensor(x):
return True return True
except TypeError: except TypeError:
return False return False
def _wrap_tensor_into_member(x): def _wrap_tensor_into_member(x):
return compile.module.Member(constant(x)) return compile.module.Member(constant(x))
compile.module.register_wrapper(_obj_is_wrappable_as_tensor, _wrap_tensor_into_member) compile.module.register_wrapper(_obj_is_wrappable_as_tensor,
_wrap_tensor_into_member)
if int(config.tensor.cmp_sloppy)>1: if int(config.tensor.cmp_sloppy) > 1:
# This config variable is a quick-and-dirty way to get low-precision # This config variable is a quick-and-dirty way to get low-precision
# comparisons. For a more precise setting of these tolerances set # comparisons. For a more precise setting of these tolerances set
# them explicitly in your user code by assigning, for example, # them explicitly in your user code by assigning, for example,
...@@ -382,8 +427,8 @@ elif int(config.tensor.cmp_sloppy): ...@@ -382,8 +427,8 @@ elif int(config.tensor.cmp_sloppy):
float64_rtol = 1e-4 float64_rtol = 1e-4
float64_atol = 1e-3 float64_atol = 1e-3
else: else:
#If you change those value in test don't forget to put them back when the test end. #If you change those value in test don't forget to put them back
#Don't forget the case when the test fail. #when the test end. Don't forget the case when the test fail.
float32_atol = 1e-5 float32_atol = 1e-5
float32_rtol = 1e-5 float32_rtol = 1e-5
...@@ -393,6 +438,7 @@ else: ...@@ -393,6 +438,7 @@ else:
#more strict. Atleast float32 precision. #more strict. Atleast float32 precision.
float64_rtol = 1.0000000000000001e-06 float64_rtol = 1.0000000000000001e-06
def _allclose(a, b, rtol=None, atol=None): def _allclose(a, b, rtol=None, atol=None):
narrow = 'float32', 'complex64' narrow = 'float32', 'complex64'
if (str(a.dtype) in narrow) or (str(b.dtype) in narrow): if (str(a.dtype) in narrow) or (str(b.dtype) in narrow):
...@@ -412,6 +458,7 @@ def _allclose(a, b, rtol=None, atol=None): ...@@ -412,6 +458,7 @@ def _allclose(a, b, rtol=None, atol=None):
return numpy.allclose(a, b, atol=atol_, rtol=rtol_) return numpy.allclose(a, b, atol=atol_, rtol=rtol_)
def get_constant_value(v): def get_constant_value(v):
"""return the constant scalar(0-D) value underlying variable `v` """return the constant scalar(0-D) value underlying variable `v`
...@@ -420,8 +467,8 @@ def get_constant_value(v): ...@@ -420,8 +467,8 @@ def get_constant_value(v):
If `v` is not some view of constant data, then raise a TypeError. If `v` is not some view of constant data, then raise a TypeError.
:note: There may be another function similar to this one in the code, but I'm not sure where it :note: There may be another function similar to this one in the
is. code, but I'm not sure where it is.
""" """
if isinstance(v, Constant): if isinstance(v, Constant):
...@@ -430,10 +477,12 @@ def get_constant_value(v): ...@@ -430,10 +477,12 @@ def get_constant_value(v):
else: else:
data = v.data data = v.data
try: try:
numpy.complex(data) #works for all numeric scalars numpy.complex(data) # works for all numeric scalars
return data return data
except Exception: except Exception:
raise TypeError('v.data is non-numeric, non-scalar, or has more than one unique value', v) raise TypeError(
'v.data is non-numeric, non-scalar, or has more than one'
' unique value', v)
if v.owner: if v.owner:
if isinstance(v.owner.op, Alloc): if isinstance(v.owner.op, Alloc):
return get_constant_value(v.owner.inputs[0]) return get_constant_value(v.owner.inputs[0])
...@@ -445,19 +494,22 @@ def get_constant_value(v): ...@@ -445,19 +494,22 @@ def get_constant_value(v):
shape, val = v.owner.inputs shape, val = v.owner.inputs
# fill(a,b) fills the shape of 'a' filled with 'b' # fill(a,b) fills the shape of 'a' filled with 'b'
return get_constant_value(val) return get_constant_value(val)
#Don't act as the constant_folding optimization here as this fct is used too early in the optimization phase. #Don't act as the constant_folding optimization here as this
#This would mess with the stabilization optimization. #fct is used too early in the optimization phase. This would
if isinstance(v.owner.op, Elemwise) and isinstance(v.owner.op.scalar_op, scal.Cast): #mess with the stabilization optimization.
if isinstance(v.owner.op, Elemwise) and isinstance(
v.owner.op.scalar_op, scal.Cast):
const = get_constant_value(v.owner.inputs[0]) const = get_constant_value(v.owner.inputs[0])
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]
if isinstance(v.owner.op, Subtensor) and v.ndim==0: if isinstance(v.owner.op, Subtensor) and v.ndim == 0:
if isinstance(v.owner.inputs[0], TensorConstant): if isinstance(v.owner.inputs[0], TensorConstant):
return v.owner.inputs[0].data.__getitem__(tuple(v.owner.op.idx_list)) return v.owner.inputs[0].data.__getitem__(
tuple(v.owner.op.idx_list))
# The index list 'idx_list' should have length the same shape as the # The index list 'idx_list' should have length the same
# input. # shape as the input.
# TODO: implement the case where we take a scalar in a matrix # TODO: implement the case where we take a scalar in a matrix
assert len(v.owner.op.idx_list) == v.owner.inputs[0].ndim assert len(v.owner.op.idx_list) == v.owner.inputs[0].ndim
...@@ -468,12 +520,14 @@ def get_constant_value(v): ...@@ -468,12 +520,14 @@ def get_constant_value(v):
# Ensure the Join is joining only scalar variables (so that # Ensure the Join is joining only scalar variables (so that
# the constant value can be found at the same index as the one # the constant value can be found at the same index as the one
# used in the sub-tensor). # used in the sub-tensor).
python_all(var.ndim==0 for var in v.owner.inputs[0].owner.inputs) and python_all(var.ndim == 0 for var in
v.owner.inputs[0].owner.inputs) and
len(v.owner.op.idx_list) == 1): len(v.owner.op.idx_list) == 1):
# Note the '+ 1' is because the first argument to Join is the # Note the '+ 1' is because the first argument to Join is the
# axis. # axis.
ret = v.owner.inputs[0].owner.inputs[v.owner.op.idx_list[0]+1] ret = v.owner.inputs[0].owner.inputs[
v.owner.op.idx_list[0] + 1]
ret = get_constant_value(ret) ret = get_constant_value(ret)
#join can cast implicitly its input in some case. #join can cast implicitly its input in some case.
return theano._asarray(ret, dtype=v.type.dtype) return theano._asarray(ret, dtype=v.type.dtype)
...@@ -482,7 +536,8 @@ def get_constant_value(v): ...@@ -482,7 +536,8 @@ def get_constant_value(v):
theano.tensor.opt.MakeVector) and theano.tensor.opt.MakeVector) and
# MakeVector normally accept only scalar as input. # MakeVector normally accept only scalar as input.
# We put this check in case there is change in the future # We put this check in case there is change in the future
python_all(var.ndim==0 for var in v.owner.inputs[0].owner.inputs) and python_all(var.ndim == 0 for var in
v.owner.inputs[0].owner.inputs) and
len(v.owner.op.idx_list) == 1): len(v.owner.op.idx_list) == 1):
ret = v.owner.inputs[0].owner.inputs[v.owner.op.idx_list[0]] ret = v.owner.inputs[0].owner.inputs[v.owner.op.idx_list[0]]
...@@ -495,7 +550,8 @@ def get_constant_value(v): ...@@ -495,7 +550,8 @@ def get_constant_value(v):
if (v.owner.inputs[0].owner and if (v.owner.inputs[0].owner and
isinstance(v.owner.inputs[0].owner.op, isinstance(v.owner.inputs[0].owner.op,
theano.tensor.Shape)): theano.tensor.Shape)):
if v.owner.inputs[0].owner.inputs[0].type.broadcastable[v.owner.op.idx_list[0]]: if v.owner.inputs[0].owner.inputs[0].type.broadcastable[
v.owner.op.idx_list[0]]:
return numpy.asarray(1) return numpy.asarray(1)
raise TypeError(v) raise TypeError(v)
...@@ -506,16 +562,17 @@ class TensorType(Type): ...@@ -506,16 +562,17 @@ class TensorType(Type):
filter_checks_isfinite = False filter_checks_isfinite = False
""" """
When this is True, strict filtering rejects data containing NaN or Inf entries. (Used in `DebugMode`) When this is True, strict filtering rejects data containing NaN or
Inf entries. (Used in `DebugMode`)
""" """
def __init__(self, dtype, broadcastable, name = None): def __init__(self, dtype, broadcastable, name=None):
"""Initialize self.dtype and self.broadcastable. """Initialize self.dtype and self.broadcastable.
:Parameters: :Parameters:
- `dtype`: str corresponding to numpy dtype (e.g., 'int64') - `dtype`: str corresponding to numpy dtype (e.g., 'int64')
The value (ndarray) associated to a `Variable` of this `Type` will have The value (ndarray) associated to a `Variable` of this `Type` will
this dtype. have this dtype.
- `broadcastable`: tuple, list, or array of boolean values - `broadcastable`: tuple, list, or array of boolean values
This argument serves two purposes. First, the True elements of this This argument serves two purposes. First, the True elements of this
list indicate the dimensions where the shape of an associated value list indicate the dimensions where the shape of an associated value
...@@ -526,16 +583,18 @@ class TensorType(Type): ...@@ -526,16 +583,18 @@ class TensorType(Type):
Optional name for this type. Optional name for this type.
""" """
self.dtype = str(dtype) self.dtype = str(dtype)
if self.dtype=='floatX': if self.dtype == 'floatX':
self.dtype=config.floatX self.dtype = config.floatX
### broadcastable is immutable, and all elements are either True or False ### broadcastable is immutable, and all elements are either
### True or False
self.broadcastable = tuple(bool(b) for b in broadcastable) self.broadcastable = tuple(bool(b) for b in broadcastable)
self.dtype_specs() # error checking is done there self.dtype_specs() # error checking is done there
self.name = name self.name = name
self.numpy_dtype = numpy.dtype(self.dtype) self.numpy_dtype = numpy.dtype(self.dtype)
def filter(self, data, strict=False, allow_downcast=None): def filter(self, data, strict=False, allow_downcast=None):
"""Convert `data` to something which can be associated to a `TensorVariable`. """Convert `data` to something which can be associated to a
`TensorVariable`.
This function is not meant to be called in user code. It is for This function is not meant to be called in user code. It is for
`Linker` instances to use when running a compiled graph. `Linker` instances to use when running a compiled graph.
...@@ -576,9 +635,10 @@ class TensorType(Type): ...@@ -576,9 +635,10 @@ class TensorType(Type):
# (do not try to convert the data) # (do not try to convert the data)
up_dtype = scal.upcast(self.dtype, data.dtype) up_dtype = scal.upcast(self.dtype, data.dtype)
if up_dtype == self.dtype: if up_dtype == self.dtype:
# Bug in the following line when data is a scalar array, # Bug in the following line when data is a
# see http://projects.scipy.org/numpy/ticket/1611 # scalar array, see
#data = data.astype(self.dtype) # http://projects.scipy.org/numpy/ticket/1611
# data = data.astype(self.dtype)
data = theano._asarray(data, dtype=self.dtype) data = theano._asarray(data, dtype=self.dtype)
if up_dtype != self.dtype: if up_dtype != self.dtype:
err_msg = ( err_msg = (
...@@ -625,13 +685,17 @@ class TensorType(Type): ...@@ -625,13 +685,17 @@ class TensorType(Type):
raise TypeError(err_msg, data) raise TypeError(err_msg, data)
if self.ndim != data.ndim: if self.ndim != data.ndim:
raise TypeError("Wrong number of dimensions: expected %s, got %s with shape %s." % (self.ndim, data.ndim, data.shape), data) raise TypeError("Wrong number of dimensions: expected %s,"
" got %s with shape %s." % (self.ndim, data.ndim,
data.shape), data)
i = 0 i = 0
for b in self.broadcastable: for b in self.broadcastable:
if b and data.shape[i] != 1: if b and data.shape[i] != 1:
raise TypeError("Non-unit value on shape on a broadcastable dimension.", data.shape, self.broadcastable) raise TypeError("Non-unit value on shape on a broadcastable"
i+=1 " dimension.", data.shape, self.broadcastable)
if self.filter_checks_isfinite and (not numpy.all(numpy.isfinite(data))): i += 1
if (self.filter_checks_isfinite and
not numpy.all(numpy.isfinite(data))):
raise ValueError("non-finite elements not allowed") raise ValueError("non-finite elements not allowed")
return data return data
...@@ -3948,10 +4012,12 @@ class IncSubtensor(Op): ...@@ -3948,10 +4012,12 @@ class IncSubtensor(Op):
return [gx, gy] + [None]*len(idx_list) return [gx, gy] + [None]*len(idx_list)
def split(x, splits_size, n_splits, axis=0): def split(x, splits_size, n_splits, axis=0):
the_split = Split(n_splits) the_split = Split(n_splits)
return the_split(x, axis, splits_size) return the_split(x, axis, splits_size)
class Split(Op): class Split(Op):
"""Partition a `TensorVariable` along some axis. """Partition a `TensorVariable` along some axis.
...@@ -3973,8 +4039,8 @@ class Split(Op): ...@@ -3973,8 +4039,8 @@ class Split(Op):
""" """
len_splits = None len_splits = None
"""A Split instance will have this many outputs, and require that the splits argument to """A Split instance will have this many outputs, and require that
`perform` have exactly this many elements. the splits argument to `perform` have exactly this many elements.
""" """
def __init__(self, len_splits): def __init__(self, len_splits):
...@@ -3997,7 +4063,8 @@ class Split(Op): ...@@ -3997,7 +4063,8 @@ class Split(Op):
splits = as_tensor_variable(splits) splits = as_tensor_variable(splits)
if splits.type not in int_vector_types: if splits.type not in int_vector_types:
raise TypeError('splits must have type tensor.lvector', splits.type) raise TypeError('splits must have type tensor.lvector',
splits.type)
if axis.type not in int_types: if axis.type not in int_types:
raise TypeError('axis must have type lscalar', axis.type) raise TypeError('axis must have type lscalar', axis.type)
...@@ -4012,31 +4079,32 @@ class Split(Op): ...@@ -4012,31 +4079,32 @@ class Split(Op):
return Apply(self, inputs, outputs) return Apply(self, inputs, outputs)
def perform(self, node, inputs, outputs): def perform(self, node, inputs, outputs):
"""WRITEME""" """WRITEME"""
x, axis, splits = inputs x, axis, splits = inputs
#in python 2.4, x.shape[numpy.asarray(1)] don't work. #in python 2.4, x.shape[numpy.asarray(1)] don't work.
if sys.version_info[0:2]==(2, 4) and axis.size==1: if sys.version_info[0:2] == (2, 4) and axis.size == 1:
axis=int(axis) axis = int(axis)
try: try:
len_along_axis = x.shape[axis] len_along_axis = x.shape[axis]
except : except:
raise ValueError('Split.perform() with axis=(%s) is invalid for x.shape==(%s)' raise ValueError('Split.perform() with axis=(%s) is invalid'
%(axis, x.shape)) ' for x.shape==(%s)'
% (axis, x.shape))
if len(splits) != self.len_splits: if len(splits) != self.len_splits:
raise ValueError('In Split.perform(), len(splits) != len_splits.', raise ValueError('In Split.perform(), len(splits) != len_splits.',
(len(splits), self.len_splits)) (len(splits), self.len_splits))
if numpy.sum(splits) != len_along_axis: if numpy.sum(splits) != len_along_axis:
raise ValueError('The splits sum to %s, expected %s' % (numpy.sum(splits), len_along_axis)) raise ValueError('The splits sum to %s, expected %s' %
(numpy.sum(splits), len_along_axis))
if not python_all(splits): if not python_all(splits):
raise ValueError('Cannot have a split of zero.') raise ValueError('Cannot have a split of zero.')
# Checking is done, let's roll the splitting algorithm! # Checking is done, let's roll the splitting algorithm!
# Basically we step along the given axis of x, extracting subtensors of size splits[i] # Basically we step along the given axis of x, extracting
# as we go along. # subtensors of size splits[i] as we go along.
general_key = [slice(None, None, None) for s in x.shape] general_key = [slice(None, None, None) for s in x.shape]
lower_idx = 0 lower_idx = 0
...@@ -4056,6 +4124,7 @@ class Split(Op): ...@@ -4056,6 +4124,7 @@ class Split(Op):
return [None for i in self.len_splits] return [None for i in self.len_splits]
return self.make_node(eval_points[0], *inputs[1:]).outputs return self.make_node(eval_points[0], *inputs[1:]).outputs
class Rebroadcast(Op): class Rebroadcast(Op):
""" """
Change the input's broadcastable fields in Change the input's broadcastable fields in
...@@ -4068,43 +4137,58 @@ class Rebroadcast(Op): ...@@ -4068,43 +4137,58 @@ class Rebroadcast(Op):
..note: work inplace and work for CudaNdarrayType ..note: work inplace and work for CudaNdarrayType
""" """
view_map = {0: [0]} view_map = {0: [0]}
def __init__(self, *axis): def __init__(self, *axis):
self.axis = dict(axis) self.axis = dict(axis)
def __eq__(self, other): def __eq__(self, other):
return type(self) == type(other) and self.axis == other.axis return type(self) == type(other) and self.axis == other.axis
def __hash__(self): def __hash__(self):
items = self.axis.items() items = self.axis.items()
items.sort() #no ambiguity because each item key is unique items.sort() # no ambiguity because each item key is unique
return hash(type(self)) ^ hash(tuple(items)) return hash(type(self)) ^ hash(tuple(items))
def __str__(self): def __str__(self):
if len(self.axis) == 0: if len(self.axis) == 0:
broadcast_pattern = [] broadcast_pattern = []
else: else:
broadcast_pattern = ['?' for i in xrange(1+numpy.max(self.axis.keys()))] broadcast_pattern = ['?' for i
for k,v in self.axis.iteritems(): in xrange(1 + numpy.max(self.axis.keys()))]
for k, v in self.axis.iteritems():
broadcast_pattern[k] = str(int(v)) broadcast_pattern[k] = str(int(v))
return '%s{%s}' % (self.__class__.__name__, ','.join(broadcast_pattern)) return '%s{%s}' % (self.__class__.__name__,
','.join(broadcast_pattern))
def make_node(self, x): def make_node(self, x):
if self.axis.keys() and (x.ndim <= numpy.max(self.axis.keys())): if self.axis.keys() and (x.ndim <= numpy.max(self.axis.keys())):
raise ValueError('Trying to rebroadcast nonexistant dimension') raise ValueError('Trying to rebroadcast nonexistant dimension')
t = x.type.__class__(dtype = x.type.dtype, t = x.type.__class__(dtype=x.type.dtype,
broadcastable = [self.axis.get(i, b) broadcastable=[self.axis.get(i, b)
for i, b in enumerate(x.type.broadcastable)]) for i, b in enumerate(
x.type.broadcastable)])
return Apply(self, [x], [t()]) return Apply(self, [x], [t()])
def perform(self, node, inp, out_): def perform(self, node, inp, out_):
x, = inp x, = inp
out, = out_ out, = out_
for axis, value in self.axis.iteritems(): for axis, value in self.axis.iteritems():
if value and x.shape[axis] != 1: if value and x.shape[axis] != 1:
raise ValueError('Dimension %s in Rebroadcast\'s input was supposed to be 1 (got %s instead)' % (axis, x.shape[axis])) raise ValueError('Dimension %s in Rebroadcast\'s input was'
' supposed to be 1 (got %s instead)' %
(axis, x.shape[axis]))
out[0] = x out[0] = x
def grad(self, inp, grads): def grad(self, inp, grads):
x, = inp x, = inp
gz, = grads gz, = grads
# restore the broadcasting pattern of the input # restore the broadcasting pattern of the input
return Rebroadcast(*[(axis, x.type.broadcastable[axis]) for axis, value in self.axis.iteritems()])(gz), return Rebroadcast(*[(axis, x.type.broadcastable[axis])
for axis, value in self.axis.iteritems()])(gz),
def infer_shape(self, node, ishapes): def infer_shape(self, node, ishapes):
assert len(ishapes)==1 assert len(ishapes) == 1
l = [] l = []
one = constant(1) one = constant(1)
for ax in xrange(len(ishapes[0])): for ax in xrange(len(ishapes[0])):
...@@ -4115,7 +4199,6 @@ class Rebroadcast(Op): ...@@ -4115,7 +4199,6 @@ class Rebroadcast(Op):
return [tuple(l)] return [tuple(l)]
def R_op(self, inputs, eval_points): def R_op(self, inputs, eval_points):
if eval_points[0] is None: if eval_points[0] is None:
return [None] return [None]
...@@ -4126,7 +4209,8 @@ def addbroadcast(x, *axes): ...@@ -4126,7 +4209,8 @@ def addbroadcast(x, *axes):
""" """
Make the input broadcastable in the specified axes. Make the input broadcastable in the specified axes.
We apply the opt here not to pollute the graph especially during the gpu optimization We apply the opt here not to pollute the graph especially during
the gpu optimization
""" """
rval = Rebroadcast(*[(axis, True) for axis in axes])(x) rval = Rebroadcast(*[(axis, True) for axis in axes])(x)
return theano.tensor.opt.apply_rebroadcast_opt(rval) return theano.tensor.opt.apply_rebroadcast_opt(rval)
...@@ -4135,11 +4219,13 @@ def unbroadcast(x, *axes): ...@@ -4135,11 +4219,13 @@ def unbroadcast(x, *axes):
""" """
Make the input impossible to broadcast in the specified axes. Make the input impossible to broadcast in the specified axes.
We apply the opt here not to pollute the graph especially during the gpu optimization We apply the opt here not to pollute the graph especially during
the gpu optimization
""" """
rval = Rebroadcast(*[(axis, False) for axis in axes])(x) rval = Rebroadcast(*[(axis, False) for axis in axes])(x)
return theano.tensor.opt.apply_rebroadcast_opt(rval) return theano.tensor.opt.apply_rebroadcast_opt(rval)
def patternbroadcast(x, broadcastable): def patternbroadcast(x, broadcastable):
""" """
Make the input adopt a specific broadcasting pattern. Make the input adopt a specific broadcasting pattern.
...@@ -4149,6 +4235,7 @@ def patternbroadcast(x, broadcastable): ...@@ -4149,6 +4235,7 @@ def patternbroadcast(x, broadcastable):
rval = Rebroadcast(*[(i,broadcastable[i]) for i in xrange(len(broadcastable))])(x) rval = Rebroadcast(*[(i,broadcastable[i]) for i in xrange(len(broadcastable))])(x)
return theano.tensor.opt.apply_rebroadcast_opt(rval) return theano.tensor.opt.apply_rebroadcast_opt(rval)
class Join(Op): class Join(Op):
""" """
Concatenate multiple `TensorVariable`s along some axis. Concatenate multiple `TensorVariable`s along some axis.
...@@ -4172,31 +4259,35 @@ class Join(Op): ...@@ -4172,31 +4259,35 @@ class Join(Op):
""" """
def __eq__(self, other): def __eq__(self, other):
return type(self) == type(other) return type(self) == type(other)
def __hash__(self): def __hash__(self):
return hash(type(self)) return hash(type(self))
def __str__(self): def __str__(self):
return '%s' %(self.__class__.__name__) return '%s' % (self.__class__.__name__)
def make_node(self, *axis_and_tensors): def make_node(self, *axis_and_tensors):
""" """
:param axis: an Int or integer-valued Variable :param axis: an Int or integer-valued Variable
:param tensors: a variable number (but not zero) of tensors to concatenate along the :param tensors: a variable number (but not zero) of tensors to
specified axis. These tensors must have the same shape along all dimensions other than this axis. concatenate along the specified axis. These tensors must have
the same shape along all dimensions other than this axis.
:returns: a symbolic Variable. It has the same ndim as the input tensors, and the most :returns: a symbolic Variable. It has the same ndim as the
inclusive dtype. input tensors, and the most inclusive dtype.
""" """
axis, tensors = axis_and_tensors[0], axis_and_tensors[1:] axis, tensors = axis_and_tensors[0], axis_and_tensors[1:]
if not tensors: if not tensors:
raise ValueError('Cannot join an empty list of tensors') raise ValueError('Cannot join an empty list of tensors')
as_tensor_variable_args= [as_tensor_variable(x) for x in tensors] as_tensor_variable_args = [as_tensor_variable(x) for x in tensors]
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, broadcastable=bcastable) output_maker = lambda bcastable: tensor(dtype=out_dtype,
broadcastable=bcastable)
return self._make_node_internal(axis, tensors, return self._make_node_internal(axis, tensors,
as_tensor_variable_args, output_maker) as_tensor_variable_args, output_maker)
...@@ -4204,8 +4295,10 @@ class Join(Op): ...@@ -4204,8 +4295,10 @@ class Join(Op):
def _make_node_internal(self, axis, tensors, def _make_node_internal(self, axis, tensors,
as_tensor_variable_args, output_maker): as_tensor_variable_args, output_maker):
orig = as_tensor_variable_args orig = as_tensor_variable_args
if not python_all(targs.type.ndim for targs in as_tensor_variable_args): if not python_all(targs.type.ndim for targs
raise TypeError('Join cannot handle arguments of dimension 0. For joining scalar values, see @stack'); in as_tensor_variable_args):
raise TypeError('Join cannot handle arguments of dimension 0.'
' For joining scalar values, see @stack')
# Handle single-tensor joins immediately. # Handle single-tensor joins immediately.
if len(as_tensor_variable_args) == 1: if len(as_tensor_variable_args) == 1:
bcastable = list(as_tensor_variable_args[0].type.broadcastable) bcastable = list(as_tensor_variable_args[0].type.broadcastable)
...@@ -4216,7 +4309,8 @@ class Join(Op): ...@@ -4216,7 +4309,8 @@ class Join(Op):
# except for the axis dimension. # except for the axis dimension.
# Initialize bcastable all false, and then fill in some trues with # Initialize bcastable all false, and then fill in some trues with
# the loops. # the loops.
bcastable = [False] * len(as_tensor_variable_args[0].type.broadcastable) bcastable = [False] * len(
as_tensor_variable_args[0].type.broadcastable)
ndim = len(bcastable) ndim = len(bcastable)
# Axis can also be a constant # Axis can also be a constant
if not isinstance(axis, int): if not isinstance(axis, int):
...@@ -4228,12 +4322,13 @@ class Join(Op): ...@@ -4228,12 +4322,13 @@ class Join(Op):
except TypeError: except TypeError:
pass pass
if isinstance(axis, int): if isinstance(axis, int):
# Basically, broadcastable -> length 1, but the converse does not # Basically, broadcastable -> length 1, but the
# hold. So we permit e.g. T/F/T joins, and if they fail at runtime # converse does not hold. So we permit e.g. T/F/T
# they fail, but if they don't then it means that the argument # joins, and if they fail at runtime they fail, but if
# where that broadcastable flag was False had length 1 along this # they don't then it means that the argument where
# dimension, and therefore this dimension should be broadcastable # that broadcastable flag was False had length 1 along
# for the output. # this dimension, and therefore this dimension should
# be broadcastable for the output.
for x in as_tensor_variable_args: for x in as_tensor_variable_args:
for current_axis, bflag in enumerate(x.type.broadcastable): for current_axis, bflag in enumerate(x.type.broadcastable):
# Not sure if this Op supports/supported/will support # Not sure if this Op supports/supported/will support
...@@ -4245,19 +4340,24 @@ class Join(Op): ...@@ -4245,19 +4340,24 @@ class Join(Op):
try: try:
bcastable[axis] = False bcastable[axis] = False
except IndexError, e: except IndexError, e:
raise ValueError('Join argument "axis" is out of range (given input dimensions)') raise ValueError('Join argument "axis" is out of range'
as_tensor_variable_args = [unbroadcast(x, axis) for x in as_tensor_variable_args] ' (given input dimensions)')
as_tensor_variable_args = [unbroadcast(x, axis)
for x in as_tensor_variable_args]
else: else:
# These unbroadcasts are for the gradient... not sure exactly # These unbroadcasts are for the gradient... not sure exactly
# why... # why...
as_tensor_variable_args = [unbroadcast(x, *range(x.type.ndim)) for x in as_tensor_variable_args] as_tensor_variable_args = [unbroadcast(x, *range(x.type.ndim))
for x in as_tensor_variable_args]
# When the axis may vary, no dimension can be guaranteed to be # When the axis may vary, no dimension can be guaranteed to be
# broadcastable. # broadcastable.
bcastable = [False] * len(as_tensor_variable_args[0].type.broadcastable) bcastable = [False] * len(
as_tensor_variable_args[0].type.broadcastable)
inputs = [as_tensor_variable(axis)] + list(as_tensor_variable_args) inputs = [as_tensor_variable(axis)] + list(as_tensor_variable_args)
if inputs[0].type not in int_types: if inputs[0].type not in int_types:
raise TypeError('Axis could not be cast to an integer type', axis, inputs[0].type, int_types) raise TypeError('Axis could not be cast to an integer type',
axis, inputs[0].type, int_types)
outputs = [output_maker(bcastable)] outputs = [output_maker(bcastable)]
...@@ -4267,7 +4367,7 @@ class Join(Op): ...@@ -4267,7 +4367,7 @@ class Join(Op):
def perform(self, node, axis_and_tensors, out_): def perform(self, node, axis_and_tensors, out_):
out, = out_ out, = out_
axis, tensors = axis_and_tensors[0], axis_and_tensors[1:] axis, tensors = axis_and_tensors[0], axis_and_tensors[1:]
out[0] = theano._asarray(numpy.concatenate(tensors, axis = axis), out[0] = theano._asarray(numpy.concatenate(tensors, axis=axis),
dtype=node.outputs[0].type.dtype) dtype=node.outputs[0].type.dtype)
def R_op(self, inputs, eval_points): def R_op(self, inputs, eval_points):
...@@ -4276,15 +4376,16 @@ class Join(Op): ...@@ -4276,15 +4376,16 @@ class Join(Op):
return self.make_node(inputs[0], *eval_points[1:]).outputs return self.make_node(inputs[0], *eval_points[1:]).outputs
def grad(self, axis_and_tensors, grads): def grad(self, axis_and_tensors, grads):
""" The gradient wrt a join op is a `Split`, used to partition the gradient along the """ The gradient wrt a join op is a `Split`, used to partition
`axis` which was used for joining. the gradient along the `axis` which was used for joining.
""" """
gz, = grads gz, = grads
axis, tensors = axis_and_tensors[0], axis_and_tensors[1:] axis, tensors = axis_and_tensors[0], axis_and_tensors[1:]
if 'float' in tensors[0].dtype or 'complex' in tensors[0].dtype: if 'float' in tensors[0].dtype or 'complex' in tensors[0].dtype:
# assume that this is differentiable # assume that this is differentiable
split = Split(len(tensors)) split = Split(len(tensors))
split_gz = split(gz, axis, stack(*[shape(x)[axis] for x in tensors])) split_gz = split(gz, axis, stack(*[shape(x)[axis]
for x in tensors]))
# If there is only one split, it might not be in a list. # If there is only one split, it might not be in a list.
if not isinstance(split_gz, list): if not isinstance(split_gz, list):
split_gz = [split_gz] split_gz = [split_gz]
...@@ -4334,11 +4435,12 @@ class Join(Op): ...@@ -4334,11 +4435,12 @@ class Join(Op):
for shape in ishapes[2:]: for shape in ishapes[2:]:
t_side = t_side + shape[dim] t_side = t_side + shape[dim]
# return the dimensions found # return the dimensions found
out_shapes.append( switch(eq(dim, node.inputs[0]), out_shapes.append(switch(eq(dim, node.inputs[0]),
t_side, f_side)) t_side, f_side))
return [tuple(out_shapes)] return [tuple(out_shapes)]
@_redefine_asRoutine(Join()) @_redefine_asRoutine(Join())
def join(axis, *tensors): def join(axis, *tensors):
""" """
...@@ -4348,14 +4450,20 @@ def join(axis, *tensors): ...@@ -4348,14 +4450,20 @@ def join(axis, *tensors):
- `tensors` : list of tensors (or list-like) - `tensors` : list of tensors (or list-like)
A list of tensors to be concatenated along the given axis. A list of tensors to be concatenated along the given axis.
- `axis` : int (symbolic or literal) - `axis` : int (symbolic or literal)
On which dimension should the tensors be joined? The `axis` must be a valid index into
the shape of the tensors to be concatenated.
The `axis` parameter may either be an integer or an object that can be converted to a
scalar using `as_scalar`(`axis`). In the former case, the axis is fixed at construction,
while in the latter it may vary over time depending on the value of the `axis` variable.
The shapes of the tensors to be concatenated must be all identical, except in the dimension On which dimension should the tensors be joined? The `axis`
(`axis`) on which they are to be joined. must be a valid index into the shape of the tensors to be
concatenated.
The `axis` parameter may either be an integer or an object that
can be converted to a scalar using `as_scalar`(`axis`). In the
former case, the axis is fixed at construction, while in the
latter it may vary over time depending on the value of the
`axis` variable.
The shapes of the tensors to be concatenated must be all
identical, except in the dimension (`axis`) on which they are to
be joined.
""" """
...@@ -4414,9 +4522,10 @@ def shape_padleft(t, n_ones=1): ...@@ -4414,9 +4522,10 @@ def shape_padleft(t, n_ones=1):
""" """
_t = as_tensor_variable(t) _t = as_tensor_variable(t)
pattern = ['x']*n_ones + [i for i in xrange(_t.type.ndim)] pattern = ['x'] * n_ones + [i for i in xrange(_t.type.ndim)]
return DimShuffle(_t.broadcastable, pattern)(_t) return DimShuffle(_t.broadcastable, pattern)(_t)
@constructor @constructor
def shape_padright(t, n_ones=1): def shape_padright(t, n_ones=1):
"""Reshape `t` by right-padding the shape with `n_ones` 1s """Reshape `t` by right-padding the shape with `n_ones` 1s
...@@ -4425,17 +4534,20 @@ def shape_padright(t, n_ones=1): ...@@ -4425,17 +4534,20 @@ def shape_padright(t, n_ones=1):
""" """
_t = as_tensor_variable(t) _t = as_tensor_variable(t)
pattern = [i for i in xrange(_t.type.ndim)] + ['x']*n_ones pattern = [i for i in xrange(_t.type.ndim)] + ['x'] * n_ones
return DimShuffle(_t.broadcastable, pattern)(_t) return DimShuffle(_t.broadcastable, pattern)(_t)
@constructor @constructor
def stack(*tensors): def stack(*tensors):
"""Insert the arguments as slices into a tensor of 1 rank greater. """Insert the arguments as slices into a tensor of 1 rank greater.
The size in dimension 0 of the result will be equal to the number of tensors passed. The size in dimension 0 of the result will be equal to the number
of tensors passed.
""" """
if len(tensors)==0: if len(tensors) == 0:
raise Exception('theano.tensor.stack(*tensors) must have at least one parameter') raise Exception('theano.tensor.stack(*tensors) must have at least'
' one parameter')
# If all tensors are scalars of the same type, call make_vector. # If all tensors are scalars of the same type, call make_vector.
# It makes the graph simpler, by not adding DimShuffles and Rebroadcasts # It makes the graph simpler, by not adding DimShuffles and Rebroadcasts
...@@ -4451,16 +4563,19 @@ def stack(*tensors): ...@@ -4451,16 +4563,19 @@ def stack(*tensors):
isinstance(t.type, TensorType) and isinstance(t.type, TensorType) and
t.ndim==0) t.ndim==0)
for t in tensors]): for t in tensors]):
tensors = map(as_tensor_variable,tensors)#in case their is direct int #in case their is direct int
tensors = map(as_tensor_variable, tensors)
dtype = scal.upcast(*[i.dtype for i in tensors]) dtype = scal.upcast(*[i.dtype for i in tensors])
return theano.tensor.opt.MakeVector(dtype)(*tensors) return theano.tensor.opt.MakeVector(dtype)(*tensors)
return join(0, *[shape_padleft(t, 1) for t in tensors]) return join(0, *[shape_padleft(t, 1) for t in tensors])
@constructor @constructor
def concatenate(tensor_list, axis=0): def concatenate(tensor_list, axis=0):
"""Alias for `join`(axis, *tensor_list). """Alias for `join`(axis, *tensor_list).
This function is similar to `join`, but uses the signature of numpy's concatenate function. This function is similar to `join`, but uses the signature of
numpy's concatenate function.
This function This function
:Exceptions: :Exceptions:
...@@ -4477,6 +4592,7 @@ def concatenate(tensor_list, axis=0): ...@@ -4477,6 +4592,7 @@ def concatenate(tensor_list, axis=0):
"arguments of concatenate.", tensor_list) "arguments of concatenate.", tensor_list)
return join(axis, *tensor_list) return join(axis, *tensor_list)
def get_vector_length(v): def get_vector_length(v):
"""Return the run-time length of a symbolic vector. """Return the run-time length of a symbolic vector.
...@@ -4487,9 +4603,9 @@ def get_vector_length(v): ...@@ -4487,9 +4603,9 @@ def get_vector_length(v):
- `TypeError` : `v` hasn't the proper type. - `TypeError` : `v` hasn't the proper type.
- `ValueError` : No special case applies, the length is not known. - `ValueError` : No special case applies, the length is not known.
In general this is not possible, but for a number of special cases the length can be In general this is not possible, but for a number of special cases
determined at compile / graph-construction time. This function implements these special the length can be determined at compile / graph-construction time.
cases. This function implements these special cases.
""" """
v = as_tensor_variable(v) v = as_tensor_variable(v)
...@@ -4505,6 +4621,7 @@ def get_vector_length(v): ...@@ -4505,6 +4621,7 @@ def get_vector_length(v):
return v.owner.inputs[0].type.ndim return v.owner.inputs[0].type.ndim
raise ValueError("length not known") raise ValueError("length not known")
@constructor @constructor
def horizontal_stack(*args): def horizontal_stack(*args):
""" """
...@@ -4520,15 +4637,19 @@ def horizontal_stack(*args): ...@@ -4520,15 +4637,19 @@ def horizontal_stack(*args):
# trying to get closer to Numpy's way of doing things. In the meantime, # trying to get closer to Numpy's way of doing things. In the meantime,
# better keep different names to emphasize the implementation divergences. # better keep different names to emphasize the implementation divergences.
assert len(args) >= 2 assert len(args) >= 2
for arg in args: assert arg.type.ndim == 2 for arg in args:
assert arg.type.ndim == 2
return concatenate(args, axis=1) return concatenate(args, axis=1)
@constructor @constructor
def vertical_stack(*args): def vertical_stack(*args):
assert len(args) >= 2 assert len(args) >= 2
for arg in args: assert arg.type.ndim == 2 for arg in args:
assert arg.type.ndim == 2
return concatenate(args, axis=0) return concatenate(args, axis=0)
if 0: #vertical and horizontal stacking are deprecated. Better to use stack() and join(). if 0: #vertical and horizontal stacking are deprecated. Better to use stack() and join().
class VerticalStack(Op): class VerticalStack(Op):
""" """
...@@ -4548,23 +4669,27 @@ if 0: #vertical and horizontal stacking are deprecated. Better to use stack() a ...@@ -4548,23 +4669,27 @@ if 0: #vertical and horizontal stacking are deprecated. Better to use stack() a
raise NotImplementedError raise NotImplementedError
inputs = [x, y] inputs = [x, y]
bcastable = (False, ) + x.type.broadcastable[1:] bcastable = (False, ) + x.type.broadcastable[1:]
outputs = [tensor(dtype = x.type.dtype, outputs = [tensor(dtype=x.type.dtype,
broadcastable = bcastable)] broadcastable=bcastable)]
return Apply(self, inputs, outputs) return Apply(self, inputs, outputs)
def perform(self, node, inp, out_): def perform(self, node, inp, out_):
x, y = inp x, y = inp
out, = out_ out, = out_
assert x.ndim == y.ndim assert x.ndim == y.ndim
# Make sure every dimension (save the first) is the same # Make sure every dimension (save the first) is the same
for i in xrange(x.ndim): assert i == 0 or x.shape[i] == y.shape[i] for i in xrange(x.ndim):
assert i == 0 or x.shape[i] == y.shape[i]
out[0] = numpy.vstack([x, y]) out[0] = numpy.vstack([x, y])
def grad(self, inp, grads): def grad(self, inp, grads):
""" """
@todo: Make VSplit (or this grad implementation) its own L{Op}, @todo: Make VSplit (or this grad implementation) its own L{Op},
that way we can do more sanity-checking:: that way we can do more sanity-checking::
assert x.ndim == y.ndim assert x.ndim == y.ndim
# Make sure every dimension (save the first) is the same # Make sure every dimension (save the first) is the same
for i in xrange(x.data.ndim): assert i == 0 or x.data.shape[i] == y.shape[i] for i in xrange(x.data.ndim):
assert i == 0 or x.data.shape[i] == y.shape[i]
etc... etc...
""" """
x, y = inp x, y = inp
...@@ -4580,21 +4705,26 @@ else: ...@@ -4580,21 +4705,26 @@ else:
class Reshape(Op): class Reshape(Op):
"""Perform a reshape operation of the input x to the new shape shp. """Perform a reshape operation of the input x to the new shape shp.
The number of dimensions to which to reshape to (ndim) must be known at graph
build time.""" The number of dimensions to which to reshape to (ndim) must be
view_map = {0: [0]} #output 0 is potentially aliased to inputs [0] known at graph build time."""
def __init__(self, ndim, name = None): view_map = {0: [0]} # output 0 is potentially aliased to inputs [0]
def __init__(self, ndim, name=None):
self.ndim = ndim self.ndim = ndim
self.name = name self.name = name
def __eq__(self, other): def __eq__(self, other):
# .name does not participate because it doesn't affect computations # .name does not participate because it doesn't affect computations
return (type(other) is type(self)) and (other.ndim == self.ndim) return (type(other) is type(self)) and (other.ndim == self.ndim)
def __hash__(self): def __hash__(self):
# .name does not participate because it doesn't affect computations # .name does not participate because it doesn't affect computations
return hash(type(self)) ^ hash(self.ndim) return hash(type(self)) ^ hash(self.ndim)
def __str__(self): def __str__(self):
return '%s{%s}' %(self.__class__.__name__, self.ndim) return '%s{%s}' % (self.__class__.__name__, self.ndim)
def make_node(self, x, shp): def make_node(self, x, shp):
x = as_tensor_variable(x) x = as_tensor_variable(x)
shp_orig = shp shp_orig = shp
...@@ -4603,12 +4733,12 @@ class Reshape(Op): ...@@ -4603,12 +4733,12 @@ class Reshape(Op):
raise TypeError("Shape must be integers", shp, shp.dtype) raise TypeError("Shape must be integers", shp, shp.dtype)
assert shp.ndim == 1 assert shp.ndim == 1
if isinstance(shp, TensorConstant): if isinstance(shp, TensorConstant):
bcast = [s==1 for s in shp.data] bcast = [s == 1 for s in shp.data]
return gof.Apply(self, [x, shp], [tensor(x.type.dtype, bcast)]) return gof.Apply(self, [x, shp], [tensor(x.type.dtype, bcast)])
else: else:
bcasts = [False] * self.ndim bcasts = [False] * self.ndim
shp_list = shp_orig shp_list = shp_orig
if hasattr(shp_orig,"ndim") and shp_orig.ndim==0: if hasattr(shp_orig, "ndim") and shp_orig.ndim == 0:
shp_list = [shp_orig] shp_list = [shp_orig]
for index in xrange(self.ndim): for index in xrange(self.ndim):
y = shp_list[index] y = shp_list[index]
...@@ -4616,20 +4746,25 @@ class Reshape(Op): ...@@ -4616,20 +4746,25 @@ class Reshape(Op):
# Try to see if we can infer that y has a constant value of 1. # Try to see if we can infer that y has a constant value of 1.
# If so, that dimension should be broadcastable. # If so, that dimension should be broadcastable.
try: try:
bcasts[index] = (hasattr(y, 'get_constant_value') and y.get_constant_value() == 1) bcasts[index] = (hasattr(y, 'get_constant_value') and
y.get_constant_value() == 1)
except TypeError: except TypeError:
pass pass
return gof.Apply(self, [x, shp], [tensor(x.type.dtype, bcasts)]) return gof.Apply(self, [x, shp], [tensor(x.type.dtype, bcasts)])
def perform(self, node, inp, out_): def perform(self, node, inp, out_):
x, shp = inp x, shp = inp
out, = out_ out, = out_
if (len(shp) != self.ndim): if (len(shp) != self.ndim):
raise ValueError('shape argument to Reshape.perform has incorrect length %i' raise ValueError('shape argument to Reshape.perform has incorrect'
' length %i'
', should be %i' % (len(shp), self.ndim), shp) ', should be %i' % (len(shp), self.ndim), shp)
try: try:
out[0] = numpy.reshape(x, shp) out[0] = numpy.reshape(x, shp)
except Exception, e: except Exception, e:
raise ValueError('Cannot reshape input of shape %s to shape %s' % (x.shape,shp)) raise ValueError('Cannot reshape input of shape %s to shape %s' %
(x.shape, shp))
def grad(self, inp, grads): def grad(self, inp, grads):
x, shp = inp x, shp = inp
g_out, = grads g_out, = grads
...@@ -4640,7 +4775,6 @@ class Reshape(Op): ...@@ -4640,7 +4775,6 @@ class Reshape(Op):
return [None] return [None]
return self.make_node(eval_points[0], *inputs[1:]).outputs return self.make_node(eval_points[0], *inputs[1:]).outputs
def infer_shape(self, node, ishapes): def infer_shape(self, node, ishapes):
# inputs[1] can contain at most one value of '-1', meaning the actual # inputs[1] can contain at most one value of '-1', meaning the actual
# shape of the output will be automatically computed by reshape, so # shape of the output will be automatically computed by reshape, so
...@@ -4862,6 +4996,7 @@ class ARange(Op): ...@@ -4862,6 +4996,7 @@ class ARange(Op):
def infer_shape(self, node, i_shapes): def infer_shape(self, node, i_shapes):
start, stop, step = node.inputs start, stop, step = node.inputs
def is_constant_value(var, value): def is_constant_value(var, value):
try: try:
v = get_constant_value(var) v = get_constant_value(var)
...@@ -4874,10 +5009,10 @@ class ARange(Op): ...@@ -4874,10 +5009,10 @@ class ARange(Op):
if is_constant_value(start, 0): if is_constant_value(start, 0):
return [(cast(stop, 'int64'),)] return [(cast(stop, 'int64'),)]
else: else:
return [(maximum(cast(stop-start, 'int64'),0),)] return [(maximum(cast(stop - start, 'int64'), 0),)]
else: else:
return [(maximum(cast(ceil(cast((stop-start),'float64') return [(maximum(cast(ceil(cast((stop - start), 'float64')
/step),'int64'),0),)] / step), 'int64'), 0),)]
def perform(self, node, inp, out_): def perform(self, node, inp, out_):
start, stop, step = inp start, stop, step = inp
...@@ -4894,6 +5029,8 @@ class ARange(Op): ...@@ -4894,6 +5029,8 @@ class ARange(Op):
def R_op(self, inputs, eval_points): def R_op(self, inputs, eval_points):
return [None] return [None]
_arange = {} _arange = {}
def arange(start, stop=None, step=1, dtype=None): def arange(start, stop=None, step=1, dtype=None):
# If only one argument is provided, it is in fact the "stop" argument, # If only one argument is provided, it is in fact the "stop" argument,
# and start is 0. # and start is 0.
...@@ -4969,7 +5106,8 @@ class PermuteRowElements(Op): ...@@ -4969,7 +5106,8 @@ class PermuteRowElements(Op):
inverse = as_tensor_variable(0) inverse = as_tensor_variable(0)
# y should contain integers # y should contain integers
assert y.type.dtype.startswith('int') or y.type.dtype.startswith('uint') assert (y.type.dtype.startswith('int') or
y.type.dtype.startswith('uint'))
# Inverse should be an integer scalar # Inverse should be an integer scalar
assert inverse.type.ndim == 0 and\ assert inverse.type.ndim == 0 and\
(inverse.type.dtype.startswith('int') or\ (inverse.type.dtype.startswith('int') or\
...@@ -4985,15 +5123,17 @@ class PermuteRowElements(Op): ...@@ -4985,15 +5123,17 @@ class PermuteRowElements(Op):
x = shape_padleft(x, n_ones=(y_dim - x_dim)) x = shape_padleft(x, n_ones=(y_dim - x_dim))
# Compute the broadcastable pattern of the output # Compute the broadcastable pattern of the output
out_broadcastable = [xb and yb for xb, yb in zip(x.type.broadcastable, y.type.broadcastable)] out_broadcastable = [xb and yb for xb, yb in
out_type = tensor(dtype = x.type.dtype, broadcastable = out_broadcastable) zip(x.type.broadcastable, y.type.broadcastable)]
out_type = tensor(dtype=x.type.dtype, broadcastable=out_broadcastable)
inputlist = [x, y, inverse] inputlist = [x, y, inverse]
outputlist = [out_type] outputlist = [out_type]
return Apply(self, inputlist, outputlist) return Apply(self, inputlist, outputlist)
def _rec_perform(self, node, x, y, inverse, out, curdim): def _rec_perform(self, node, x, y, inverse, out, curdim):
"""Perform the permutation by doing a recursion over the input dimensions. """Perform the permutation by doing a recursion over the input
dimensions.
For every dimension, starting with the leftmost, the right set of For every dimension, starting with the leftmost, the right set of
indices is determined (depending if broadcasting or not), then indices is determined (depending if broadcasting or not), then
...@@ -5027,15 +5167,18 @@ class PermuteRowElements(Op): ...@@ -5027,15 +5167,18 @@ class PermuteRowElements(Op):
ys0 = y.shape[0] ys0 = y.shape[0]
if xs0 == ys0: if xs0 == ys0:
for i in xrange(xs0): for i in xrange(xs0):
self._rec_perform(node, x[i], y[i], inverse, out[i], curdim+1) self._rec_perform(node, x[i], y[i], inverse, out[i],
curdim+1)
elif ys0 == 1 and node.inputs[1].type.broadcastable[curdim]: elif ys0 == 1 and node.inputs[1].type.broadcastable[curdim]:
# Broadcast y # Broadcast y
for i in xrange(xs0): for i in xrange(xs0):
self._rec_perform(node, x[i], y[0], inverse, out[i], curdim+1) self._rec_perform(node, x[i], y[0], inverse, out[i],
curdim+1)
elif xs0 == 1 and node.inputs[0].type.broadcastable[curdim]: elif xs0 == 1 and node.inputs[0].type.broadcastable[curdim]:
# Broadcast x # Broadcast x
for i in xrange(ys0): for i in xrange(ys0):
self._rec_perform(node, x[0], y[i], inverse, out[i], curdim+1) self._rec_perform(node, x[0], y[i], inverse, out[i],
curdim+1)
else: else:
raise ValueError('Dimension mismatch: %s, %s' % (xs0, ys0)) raise ValueError('Dimension mismatch: %s, %s' % (xs0, ys0))
...@@ -5075,9 +5218,10 @@ class PermuteRowElements(Op): ...@@ -5075,9 +5218,10 @@ class PermuteRowElements(Op):
# If x has been broadcasted along some axes, we need to sum # If x has been broadcasted along some axes, we need to sum
# 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] and not gz.type.broadcastable[dim]] if x.type.broadcastable[dim]
gx = Sum(axis = broadcasted_dims)(gx) and not gz.type.broadcastable[dim]]
gx = Sum(axis=broadcasted_dims)(gx)
# Sum(...) removed the dimensions in broadcasted_dims, # Sum(...) removed the dimensions in broadcasted_dims,
# so we need to put them back. # so we need to put them back.
...@@ -5093,11 +5237,13 @@ class PermuteRowElements(Op): ...@@ -5093,11 +5237,13 @@ class PermuteRowElements(Op):
gx = DimShuffle(gx.type.broadcastable, newdims)(gx) gx = DimShuffle(gx.type.broadcastable, newdims)(gx)
assert gx.type.broadcastable == x.type.broadcastable assert gx.type.broadcastable == x.type.broadcastable
return [gx, None, None] return [gx, None, None]
_permute_row_elements = PermuteRowElements() _permute_row_elements = PermuteRowElements()
def permute_row_elements(x, y, inverse=0): def permute_row_elements(x, y, inverse=0):
return _permute_row_elements(x, y, inverse) return _permute_row_elements(x, y, inverse)
def inverse_permutation(perm): def inverse_permutation(perm):
"""Computes the inverse of permutations. """Computes the inverse of permutations.
Each row of input should contain a permutation of the first integers. Each row of input should contain a permutation of the first integers.
...@@ -5114,6 +5260,7 @@ def inverse_permutation(perm): ...@@ -5114,6 +5260,7 @@ def inverse_permutation(perm):
# Should reproduce numpy's behaviour: # Should reproduce numpy's behaviour:
# http://docs.scipy.org/doc/numpy/reference/arrays.indexing.html#advanced-indexing # http://docs.scipy.org/doc/numpy/reference/arrays.indexing.html#advanced-indexing
class AdvancedSubtensor1(Op): class AdvancedSubtensor1(Op):
"""Implement x[ilist] where ilist is a vector of integers.""" """Implement x[ilist] where ilist is a vector of integers."""
...@@ -5290,19 +5437,28 @@ class AdvancedSubtensor(Op): ...@@ -5290,19 +5437,28 @@ class AdvancedSubtensor(Op):
if x.ndim == 2 and len(inputs) == 2: if x.ndim == 2 and len(inputs) == 2:
ind1 = as_tensor_variable(inputs[0]) ind1 = as_tensor_variable(inputs[0])
ind2 = as_tensor_variable(inputs[1]) ind2 = as_tensor_variable(inputs[1])
if not (ind1.type.dtype.startswith('int') or ind1.type.dtype.startswith('uint')): if (not (ind1.type.dtype.startswith('int') or
raise TypeError('the indices into a matrix must be int or uint. It is ',ind1.type.dtype) ind1.type.dtype.startswith('uint'))):
if not (ind2.type.dtype.startswith('int') or ind2.type.dtype.startswith('uint')): raise TypeError(
raise TypeError('the indices into a matrix must be int or uint. It is ',ind2.type.dtype) 'the indices into a matrix must be int or uint. It is ',
ind1.type.dtype)
if (not (ind2.type.dtype.startswith('int') or
ind2.type.dtype.startswith('uint'))):
raise TypeError(
'the indices into a matrix must be int or uint. It is ',
ind2.type.dtype)
if ind1.ndim == 1 and ind2.ndim == 1: if ind1.ndim == 1 and ind2.ndim == 1:
return gof.Apply(self, return gof.Apply(self,
(x,) + inputs, (x,) + inputs,
[tensor(dtype = x.type.dtype, [tensor(dtype=x.type.dtype,
broadcastable = [False])]) broadcastable=[False])])
raise NotImplementedError('Advanced indexing of x (of dimension %i) with these argument dimensions (%s) not supported yet'\ raise NotImplementedError(
'Advanced indexing of x (of dimension %i) with these argument'
' dimensions (%s) not supported yet'
% (x.ndim, ','.join(str(input.ndim) for input in inputs))) % (x.ndim, ','.join(str(input.ndim) for input in inputs)))
raise NotImplementedError('Advanced indexing of x with arguments (%s) not supported yet'\ raise NotImplementedError(
'Advanced indexing of x with arguments (%s) not supported yet'
% ','.join(str(input) for input in inputs)) % ','.join(str(input) for input in inputs))
def R_op(self, inputs, eval_points): def R_op(self, inputs, eval_points):
...@@ -5327,8 +5483,8 @@ class AdvancedSubtensor(Op): ...@@ -5327,8 +5483,8 @@ class AdvancedSubtensor(Op):
def perform(self, node, inputs, out_): def perform(self, node, inputs, out_):
out, = out_ out, = out_
# TODO: in general, we need to re-pack the inputs into a valid index, just like # TODO: in general, we need to re-pack the inputs into a valid
# subtensor # index, just like subtensor
out[0] = inputs[0].__getitem__(inputs[1:]) out[0] = inputs[0].__getitem__(inputs[1:])
if (numpy.__version__ <= '1.6.1' and if (numpy.__version__ <= '1.6.1' and
out[0].size != numpy.uint32(out[0].size)): out[0].size != numpy.uint32(out[0].size)):
...@@ -5345,7 +5501,8 @@ class AdvancedSubtensor(Op): ...@@ -5345,7 +5501,8 @@ class AdvancedSubtensor(Op):
gz, = grads gz, = grads
x = inputs[0] x = inputs[0]
rest = inputs[1:] rest = inputs[1:]
return [AdvancedIncSubtensor()(zeros_like(x), gz, *rest)] + [None]*len(rest) return [AdvancedIncSubtensor()(zeros_like(x), gz,
*rest)] + [None] * len(rest)
class AdvancedIncSubtensor(Op): class AdvancedIncSubtensor(Op):
...@@ -5371,11 +5528,17 @@ class AdvancedIncSubtensor(Op): ...@@ -5371,11 +5528,17 @@ class AdvancedIncSubtensor(Op):
if ind1.ndim == 1 and ind2.ndim == 1: if ind1.ndim == 1 and ind2.ndim == 1:
return gof.Apply(self, return gof.Apply(self,
(x, y) + inputs, (x, y) + inputs,
[tensor(dtype = x.type.dtype, [tensor(dtype=x.type.dtype,
broadcastable = x.type.broadcastable)]) broadcastable=x.type.broadcastable)])
raise NotImplementedError('Advanced indexing increment of x (of dimension %i) by y (of dimension %i) with these argument dimensions (%s) not supported yet'\ raise NotImplementedError(
% (x.ndim, y.ndim, ','.join(str(input.ndim) for input in inputs))) 'Advanced indexing increment of x (of dimension %i) by y'
raise NotImplementedError('Advanced indexing increment of x (of dim %i) by y (of dim %i) with arguments (%s) not supported yet'\ ' (of dimension %i) with these argument dimensions (%s) not'
' supported yet'
% (x.ndim, y.ndim,
','.join(str(input.ndim) for input in inputs)))
raise NotImplementedError(
'Advanced indexing increment of x (of dim %i) by y (of dim %i)'
' with arguments (%s) not supported yet'
% (x.ndim, y.ndim, ','.join(str(input) for input in inputs))) % (x.ndim, y.ndim, ','.join(str(input) for input in inputs)))
def perform(self, node, inputs, out_): def perform(self, node, inputs, out_):
...@@ -5403,11 +5566,8 @@ class AdvancedIncSubtensor(Op): ...@@ -5403,11 +5566,8 @@ class AdvancedIncSubtensor(Op):
def R_op(self, inputs, eval_points): def R_op(self, inputs, eval_points):
if None in eval_points[:2]: if None in eval_points[:2]:
return [None] return [None]
return self.make_node(eval_points[0], eval_points[1], *inputs[2:]).outputs return self.make_node(eval_points[0], eval_points[1],
*inputs[2:]).outputs
######################### #########################
...@@ -5421,18 +5581,23 @@ class AdvancedIncSubtensor(Op): ...@@ -5421,18 +5581,23 @@ class AdvancedIncSubtensor(Op):
class Dot(Op): class Dot(Op):
"""Compute matrix-matrix, matrix-vector products and vector inner-products. """Compute matrix-matrix, matrix-vector products and vector inner-products.
:note: matrix-matrix products are sometimes optimized to Dot22 ops (see tensor.blas) :note: matrix-matrix products are sometimes optimized to Dot22 ops
(see tensor.blas)
:note: non matrix-matrix products (including matrix-vector products) are handled by numpy. Ensure that you have linked numpy with a fast BLAS. :note: non matrix-matrix products (including matrix-vector
products) are handled by numpy. Ensure that you have linked numpy
with a fast BLAS.
""" """
def __eq__(self, other): def __eq__(self, other):
return type(self) == type(other) return type(self) == type(other)
def __hash__(self): def __hash__(self):
return hash(type(self)) return hash(type(self))
# the rationale for Dot22 is related to getting GEMM Ops into the graph. See Dot22 in tensor.blas for details. # the rationale for Dot22 is related to getting GEMM Ops into the
# graph. See Dot22 in tensor.blas for details.
def make_node(self, *inputs): def make_node(self, *inputs):
inputs = map(as_tensor_variable, inputs) inputs = map(as_tensor_variable, inputs)
...@@ -5441,32 +5606,38 @@ class Dot(Op): ...@@ -5441,32 +5606,38 @@ class Dot(Op):
if numpy_semantics: if numpy_semantics:
#numpy defines dot for tensor pairs with any rank #numpy defines dot for tensor pairs with any rank
if len(inputs) != 2: if len(inputs) != 2:
raise TypeError("Wrong number of inputs for %s (got %i, expected 2)" % self) raise TypeError(
"Wrong number of inputs for %s (got %i, expected 2)" %
self)
i_broadcastables = [input.type.broadcastable for input in inputs] i_broadcastables = [input.type.broadcastable for input in inputs]
bx, by = i_broadcastables bx, by = i_broadcastables
if len(bx) == 0: # x is a scalar if len(bx) == 0: # x is a scalar
bz = by bz = by
else: else:
if len(by) >= 2: #y is a matrix or tensor if len(by) >= 2: # y is a matrix or tensor
bz = bx[:-1] + by[:-2] + by[-1:] bz = bx[:-1] + by[:-2] + by[-1:]
elif len(by)==1: #y is vector elif len(by) == 1: # y is vector
bz = bx[:-1] bz = bx[:-1]
else: #y is a scalar else: # y is a scalar
bz = bx bz = bx
else: else:
if len(inputs) != 2: if len(inputs) != 2:
raise TypeError('theanor.tensor.Dot: 2 arguments required, %d given ' % len(inputs)) raise TypeError(
'theanor.tensor.Dot: 2 arguments required, %d given ' %
len(inputs))
x, y = inputs x, y = inputs
nx = x.type.ndim nx = x.type.ndim
ny = y.type.ndim ny = y.type.ndim
if nx not in (1,2): if nx not in (1, 2):
raise TypeError(('dot supports matrix and vector args: email theano-dev about' raise TypeError(
' enabling numpy dot semantics if you want them'), x) ('dot supports matrix and vector args: email theano-dev '
if ny not in (1,2): 'about enabling numpy dot semantics if you want them'), x)
raise TypeError(('dot supports matrix and vector args: email theano-dev about' if ny not in (1, 2):
' enabling numpy dot semantics if you want them'), y) raise TypeError(
('dot supports matrix and vector args: email theano-dev '
'about enabling numpy dot semantics if you want them'), y)
if nx == 2 and ny == 2: if nx == 2 and ny == 2:
bz = [x.type.broadcastable[0], y.type.broadcastable[1]] bz = [x.type.broadcastable[0], y.type.broadcastable[1]]
...@@ -5485,8 +5656,9 @@ class Dot(Op): ...@@ -5485,8 +5656,9 @@ class Dot(Op):
x, y = inp x, y = inp
z, = out z, = out
try: try:
# the asarray is here because dot between two vectors gives a numpy float object # the asarray is here because dot between two vectors
# but we need to return a 0d ndarray # gives a numpy float object but we need to return a 0d
# ndarray
z[0] = numpy.asarray(numpy.dot(x, y)) z[0] = numpy.asarray(numpy.dot(x, y))
except ValueError, e: except ValueError, e:
# The error raised by numpy has no shape information, we mean to # The error raised by numpy has no shape information, we mean to
...@@ -5530,42 +5702,48 @@ class Dot(Op): ...@@ -5530,42 +5702,48 @@ class Dot(Op):
try: try:
iv0 = gof.op.get_test_value(inputs[0]) iv0 = gof.op.get_test_value(inputs[0])
except AttributeError: except AttributeError:
gof.op.missing_test_message('first input passed to Dot.R_op has no test value') gof.op.missing_test_message(
'first input passed to Dot.R_op has no test value')
debugger_available = False debugger_available = False
try: try:
iv1 = gof.op.get_test_value(inputs[1]) iv1 = gof.op.get_test_value(inputs[1])
except AttributeError: except AttributeError:
gof.op.missing_test_message('second input passed to Dot.R_op has no test value') gof.op.missing_test_message(
'second input passed to Dot.R_op has no test value')
debugger_available = False debugger_available = False
try: try:
ev0 = gof.op.get_test_value(eval_points[0]) ev0 = gof.op.get_test_value(eval_points[0])
except AttributeError: except AttributeError:
gof.op.missing_test_message('first eval point passed to Dot.R_op has no test value') gof.op.missing_test_message(
'first eval point passed to Dot.R_op has no test value')
debugger_available = False debugger_available = False
try: try:
ev1 = gof.op.get_test_value(eval_points[1]) ev1 = gof.op.get_test_value(eval_points[1])
except AttributeError: except AttributeError:
gof.op.missing_test_message('second eval point passed to Dot.R_op has no test value') gof.op.missing_test_message(
'second eval point passed to Dot.R_op has no test value')
debugger_available = False debugger_available = False
if debugger_available: if debugger_available:
input_values = [ iv0, iv1] input_values = [iv0, iv1]
eval_point_values = [ ev0, ev1 ] eval_point_values = [ev0, ev1]
for i in xrange(2): for i in xrange(2):
if input_values[i].shape != eval_point_values[i].shape: if input_values[i].shape != eval_point_values[i].shape:
raise ValueError('input '+str(i)+' and eval_point '+str(i)+' to Dot.R_op ' raise ValueError('input ' + str(i) + ' and eval_point ' +
str(i) + ' to Dot.R_op '
'should have the ' 'should have the '
'same shape, but their shapes are %s and %s, respectively' % ( \ 'same shape, but their shapes are'
str(input_values[i].shape), str(eval_point_values[i].shape) ) ) ' %s and %s, respectively' % (
str(input_values[i].shape),
str(eval_point_values[i].shape)))
t1 = self(eval_points[0], inputs[1]) t1 = self(eval_points[0], inputs[1])
t2 = self(inputs[0], eval_points[1]) t2 = self(inputs[0], eval_points[1])
return [t1+t2] return [t1 + t2]
def infer_shape(self, node, shapes): def infer_shape(self, node, shapes):
xshp, yshp = shapes xshp, yshp = shapes
...@@ -5583,7 +5761,11 @@ class Dot(Op): ...@@ -5583,7 +5761,11 @@ class Dot(Op):
def __str__(self): def __str__(self):
return "dot" return "dot"
dot = Dot() dot = Dot()
pprint.assign(dot, printing.OperatorPrinter(printing.special['middle_dot'], -1, 'left'))
pprint.assign(dot, printing.OperatorPrinter(printing.special['middle_dot'],
-1, 'left'))
######################### #########################
# Linalg : TensorDot # Linalg : TensorDot
...@@ -5603,14 +5785,14 @@ class TensorDotGrad(Op): ...@@ -5603,14 +5785,14 @@ class TensorDotGrad(Op):
assert isinstance(y, Variable) assert isinstance(y, Variable)
assert isinstance(gz, Variable) assert isinstance(gz, Variable)
gx = tensor(dtype=scal.upcast(gz.dtype, y.dtype), gx = tensor(dtype=scal.upcast(gz.dtype, y.dtype),
broadcastable = x.broadcastable) broadcastable=x.broadcastable)
gy = tensor(dtype=scal.upcast(x.dtype, gz.dtype), gy = tensor(dtype=scal.upcast(x.dtype, gz.dtype),
broadcastable = y.broadcastable) broadcastable=y.broadcastable)
op = self op = self
if isinstance(self.axes,int): if isinstance(self.axes, int):
axes = [range(x.ndim-self.axes,x.ndim),range(self.axes)] axes = [range(x.ndim - self.axes, x.ndim), range(self.axes)]
op = TensorDotGrad(axes) op = TensorDotGrad(axes)
return Apply(op, [x,y,gz], [gx, gy]) return Apply(op, [x, y, gz], [gx, gy])
def perform(self, node, inp, out): def perform(self, node, inp, out):
x, y, gz = inp x, y, gz = inp
...@@ -5844,7 +6026,8 @@ def sort(a, axis=-1, kind='quicksort', order=None): ...@@ -5844,7 +6026,8 @@ def sort(a, axis=-1, kind='quicksort', order=None):
Tensor to be sorted Tensor to be sorted
axis : Tensor axis : Tensor
Axis along which to sort. If None, the array is flattened before sorting. Axis along which to sort. If None, the array is
flattened before sorting.
kind : {'quicksort', 'mergesort', 'heapsort'}, optional kind : {'quicksort', 'mergesort', 'heapsort'}, optional
......
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