Skip to content

P
pytensor
提交 b41e3605 authored 作者: Frederic's avatar Frederic
浏览文件
操作

Move TensorType to its own file as CudaNdarrayType and SparseType.

This advance ticket gh-651
上级 0a1468ed
隐藏空白字符变更
内嵌 并排
正在显示 包含 642 行增加 和 637 行删除
theano/gof/__init__.py
浏览文件 @ b41e3605
......@@ -78,4 +78,4 @@ from theano.gof.type import \
Type, Generic, generic
from theano.gof.utils import \
object2, MethodNotDefined
hashtype, object2, MethodNotDefined
theano/gof/utils.py
浏览文件 @ b41e3605
......@@ -22,6 +22,11 @@ def hashgen():
hashgen.next = 0
def hashtype(self):
t = type(self)
return hash(t.__name__) ^ hash(t.__module__)
class MethodNotDefined(Exception):
"""
To be raised by functions defined as part of an interface.
......
theano/tensor/basic.py
浏览文件 @ b41e3605
......@@ -14,12 +14,13 @@ import theano
from theano.compat import PY3
from theano.configparser import config
from theano import gof
from theano.gof import Apply, Constant, Op, Type, Variable
from theano.gof import Apply, Constant, Op, Variable
from theano.tensor import elemwise
from theano.tensor.type import TensorType
from theano import scalar as scal
from theano.gof.python25 import partial, any, all, maxsize
from theano.gof.utils import MethodNotDefined
from theano.gof.utils import hashtype, MethodNotDefined
from theano import compile, printing
from theano.printing import pprint, min_informative_str
from theano.tensor.utils import hash_from_ndarray
......@@ -94,12 +95,6 @@ def check_equal_numpy(x, y):
compile.register_checker(check_equal_numpy)
def hashtype(self):
t = type(self)
return hash(t.__name__) ^ hash(t.__module__)
elemwise.hashtype = hashtype
__oplist_constructor_list = []
"""List of functions to be listed as op constructors in the oplist
(`gen_oplist`, doc/oplist.txt)."""
......@@ -671,634 +666,6 @@ def get_scalar_constant_value(v):
raise NotScalarConstantError(v)
class TensorType(Type):
"""Symbolic `Type` representing a numpy.ndarray value."""
filter_checks_isfinite = False
"""
When this is True, strict filtering rejects data containing NaN or
Inf entries. (Used in `DebugMode`)
"""
def __init__(self, dtype, broadcastable, name=None, sparse_grad=False):
"""Initialize self.dtype and self.broadcastable.
:Parameters:
- `dtype`: str corresponding to numpy dtype (e.g., 'int64')
The value (ndarray) associated to a `Variable` of this `Type` will
have this dtype.
- `broadcastable`: tuple, list, or array of boolean values
This argument serves two purposes. First, the True elements of this
list indicate the dimensions where the shape of an associated value
must be 1. Secondly, the length of this list is the number of
dimensions that an associated value must have. See
:doc:`broadcasting` for an explanation of how this list is used.
- `name`: str
Optional name for this type.
"""
self.dtype = str(dtype)
if self.dtype == 'floatX':
self.dtype = config.floatX
### broadcastable is immutable, and all elements are either
### True or False
self.broadcastable = tuple(bool(b) for b in broadcastable)
self.dtype_specs() # error checking is done there
self.name = name
self.numpy_dtype = numpy.dtype(self.dtype)
self.sparse_grad = sparse_grad
if sparse_grad:
warnings.warn(
"DEPRECATION WARNING: You use an old interface to"
" AdvancedSubtensor1 sparse_grad. Now use"
" theano.sparse_grad(a_tensor[an_int_vector]).")
def filter(self, data, strict=False, allow_downcast=None):
"""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
`Linker` instances to use when running a compiled graph.
"""
# Explicit error message when one accidentally uses a Variable as
# input (typical mistake, especially with shared variables).
if isinstance(data, Variable):
raise TypeError(
'Expected an array-like object, but found a Variable: '
'maybe you are trying to call a function on a (possibly '
'shared) variable instead of a numeric array?')
if ((type(data) is numpy.ndarray)
and (data.dtype == self.numpy_dtype)):
if data.dtype.num != self.numpy_dtype.num:
data = theano._asarray(data, dtype=self.dtype)
# -- now fall through to ndim check
elif((type(data) is numpy.memmap)
and (data.dtype == self.numpy_dtype)):
# numpy.memmap is a "safe" subclass of ndarray,
# so we can use it whereever we expect a base ndarray.
# however, casting it would defeat the purpose of not
# loading the whole data into memory
pass
elif strict:
# If any of the two conditions above was not met,
# we raise a meaningful TypeError.
if not (type(data) is numpy.ndarray):
raise TypeError("%s expected a ndarray object." % self,
data, type(data))
if data.dtype != self.numpy_dtype:
raise TypeError(("%s expected a ndarray object with "
"dtype = %s (got %s).") % (
self, self.numpy_dtype, data.dtype))
assert False, "This point should never be reached."
else:
if allow_downcast:
# Convert to self.dtype, regardless of the type of data
data = theano._asarray(data, dtype=self.dtype)
# TODO: consider to pad shape with ones to make it consistent
# with self.broadcastable... like vector->row type thing
else:
if isinstance(data, numpy.ndarray):
# Check if self.dtype can accurately represent data
# (do not try to convert the data)
up_dtype = scal.upcast(self.dtype, data.dtype)
if up_dtype == self.dtype:
# Bug in the following line when data is a
# scalar array, see
# http://projects.scipy.org/numpy/ticket/1611
# data = data.astype(self.dtype)
data = theano._asarray(data, dtype=self.dtype)
if up_dtype != self.dtype:
err_msg = (
'%s cannot store a value of dtype %s without '
'risking loss of precision. If you do not mind '
'this loss, you can: '
'1) explicitly cast your data to %s, or '
'2) set "allow_input_downcast=True" when calling '
'"function".'
% (self, data.dtype, self.dtype))
raise TypeError(err_msg, data)
elif (allow_downcast is None and
type(data) is float and
self.dtype == theano.config.floatX):
# Special case where we allow downcasting of Python float
# literals to floatX, even when floatX=='float32'
data = theano._asarray(data, self.dtype)
else:
# data has to be converted.
# Check that this conversion is lossless
converted_data = theano._asarray(data, self.dtype)
# We use the `values_eq` static function from TensorType
# to handle NaN values.
if TensorType.values_eq(numpy.asarray(data),
converted_data,
force_same_dtype=False):
data = converted_data
else:
# Do not print a too long description of data
# (ndarray truncates it, but it's not sure for data)
str_data = str(data)
if len(str_data) > 80:
str_data = str_data[:75] + '(...)'
err_msg = (
'%s cannot store accurately value %s, '
'it would be represented as %s. '
'If you do not mind this precision loss, you can: '
'1) explicitly convert your data to a numpy array '
'of dtype %s, or '
'2) set "allow_input_downcast=True" when calling '
'"function".'
% (self, data, converted_data, self.dtype))
raise TypeError(err_msg, data)
if self.ndim != data.ndim:
raise TypeError("Wrong number of dimensions: expected %s,"
" got %s with shape %s." % (self.ndim, data.ndim,
data.shape))
if not data.flags.aligned:
try:
msg = "object buffer" + str(data.data)
except AttributeError:
msg = ""
raise TypeError("The numpy.ndarray object is not aligned."
" Theano C code does not support that.",
msg,
"object shape", data.shape,
"object strides", data.strides)
i = 0
for b in self.broadcastable:
if b and data.shape[i] != 1:
raise TypeError("Non-unit value on shape on a broadcastable"
" dimension.", data.shape, self.broadcastable)
i += 1
if (self.filter_checks_isfinite and
not numpy.all(numpy.isfinite(data))):
raise ValueError("non-finite elements not allowed")
return data
def filter_variable(self, other):
"""Convert a symbolic Variable into a TensorType, if compatible.
For the moment, only a TensorType or CudaNdarrayType will be
converted, provided they have the same number of dimensions,
broadcastable pattern, and dtype.
"""
if hasattr(other, '_as_TensorVariable'):
other = other._as_TensorVariable()
if not isinstance(other, Variable):
# The value is not a Variable: we cast it into
# a Constant of the appropriate Type.
other = self.Constant(type=self, data=other)
if other.type == self:
return other
raise TypeError(
'Cannot convert Type %(othertype)s '
'(of Variable %(other)s) into Type %(self)s. '
'You can try to manually convert %(other)s into a %(self)s.'
% dict(
othertype=other.type,
other=other,
self=self)
)
def value_validity_msg(self, a):
try:
self.filter(a, strict=True)
except Exception, e:
return str(e)
return "value is valid"
def dtype_specs(self):
"""Return a tuple (python type, c type, numpy typenum) that corresponds
to self.dtype.
This function is used internally as part of C code generation.
"""
# TODO: add more type correspondances for e.g. int32, int64, float32,
# complex64, etc.
try:
return {
'float32': (float, 'npy_float32', 'NPY_FLOAT32'),
'float64': (float, 'npy_float64', 'NPY_FLOAT64'),
'uint8': (int, 'npy_uint8', 'NPY_UINT8'),
'int8': (int, 'npy_int8', 'NPY_INT8'),
'uint16': (int, 'npy_uint16', 'NPY_UINT16'),
'int16': (int, 'npy_int16', 'NPY_INT16'),
'uint32': (int, 'npy_uint32', 'NPY_UINT32'),
'int32': (int, 'npy_int32', 'NPY_INT32'),
'uint64': (int, 'npy_uint64', 'NPY_UINT64'),
'int64': (int, 'npy_int64', 'NPY_INT64'),
'complex128': (complex, 'theano_complex128', 'NPY_COMPLEX128'),
'complex64': (complex, 'theano_complex64', 'NPY_COMPLEX64')
}[self.dtype]
except KeyError:
raise TypeError("Unsupported dtype for %s: %s"
% (self.__class__.__name__, self.dtype))
def to_scalar_type(self):
return scal.Scalar(dtype=self.dtype)
def __eq__(self, other):
"""Compare True iff other is the same kind of TensorType"""
return type(self) == type(other) and other.dtype == self.dtype \
and other.broadcastable == self.broadcastable
@staticmethod
def may_share_memory(a, b):
# This is a method of TensorType, so both a and b should be ndarrays
if isinstance(a, numpy.ndarray) and isinstance(b, numpy.ndarray):
return numpy.may_share_memory(a, b)
else:
return False
@staticmethod
def values_eq(a, b, force_same_dtype=True):
# TODO: check to see if the shapes must match
# for now, we err on safe side...
if a.shape != b.shape:
return False
if force_same_dtype and a.dtype != b.dtype:
return False
a_eq_b = (a == b)
r = numpy.all(a_eq_b)
if r:
return True
# maybe the trouble is that there are NaNs
a_missing = numpy.isnan(a)
if a_missing.any():
b_missing = numpy.isnan(b)
return numpy.all(a_eq_b + (a_missing == b_missing))
else:
return False
@staticmethod
def values_eq_approx(a, b, allow_remove_inf=False, allow_remove_nan=False,
rtol=None, atol=None):
"""
:param allow_remove_inf: If True, when there is an inf in a,
we allow any value in b in that position.
Event -inf
:param allow_remove_nan: If True, when there is a nan in a,
we allow any value in b in that position.
Event +-inf
:param rtol: relative tolerance, passed to _allclose
:param atol: absolute tolerance, passed to _allclose
"""
if isinstance(a, numpy.ndarray) and isinstance(b, numpy.ndarray):
if a.shape != b.shape:
return False
if a.dtype != b.dtype:
return False
if 'int' in str(a.dtype):
return numpy.all(a == b)
else:
# work around a numpy.allclose bug:
# http://projects.scipy.org/numpy/ticket/1672
if a.ndim == 0 and numpy.isinf(a):
a = a.reshape(1)
b = b.reshape(1)
cmp = _allclose(a, b, rtol=rtol, atol=atol)
if cmp:
# Numpy claims they are close, this is good enough for us.
return True
# Numpy is unhappy, but it does not necessarily mean that a and
# b are different. Indeed, Numpy does not like missing values
# and will return False whenever some are found in a or b.
# The proper way would be to use the MaskArray stuff available
# in Numpy. However, it looks like it has been added to Numpy's
# core recently, so it may not be available to everyone. Thus,
# for now we use a home-made recipe, that should probably be
# revisited in the future.
a_missing = numpy.isnan(a)
a_inf = numpy.isinf(a)
if not (a_missing.any() or (allow_remove_inf and a_inf.any())):
# There are no missing values in a, thus this is not the
# reason why numpy.allclose(a, b) returned False.
_logger.info(
'numpy allclose failed for abs_err %f and rel_err %f',
numpy.max(abs(a - b)),
numpy.max(abs(a - b) / (abs(a) + abs(b))))
return False
# The following line is what numpy.allclose bases its decision
# upon, according to its documentation.
rtol = 1.0000000000000001e-05
atol = 1e-8
cmp_elemwise = (numpy.absolute(a - b) <=
(atol + rtol * numpy.absolute(b)))
# Find places where both a and b have missing values.
both_missing = a_missing * numpy.isnan(b)
# Find places where both a and b have inf of the same sign.
both_inf = a_inf * numpy.isinf(b)
# cmp_elemwise is weird when we have inf and -inf.
# set it to False
cmp_elemwise = numpy.where(
both_inf & cmp_elemwise,
a == b,
cmp_elemwise)
# check the sign of the inf
both_inf = numpy.where(both_inf, (a == b), both_inf)
if allow_remove_inf:
both_inf += a_inf
if allow_remove_nan:
both_missing += a_missing
# Combine all information.
return (cmp_elemwise + both_missing + both_inf).all()
return False
@staticmethod
def values_eq_approx_remove_inf(a, b):
return TensorType.values_eq_approx(a, b, True)
@staticmethod
def values_eq_approx_remove_nan(a, b):
return TensorType.values_eq_approx(a, b, False, True)
@staticmethod
def values_eq_approx_remove_inf_nan(a, b):
return TensorType.values_eq_approx(a, b, True, True)
def __hash__(self):
"""Hash equal for same kinds of TensorType"""
return hashtype(self) ^ hash(self.dtype) ^ hash(self.broadcastable)
ndim = property(lambda self: len(self.broadcastable),
doc="number of dimensions")
"""Number of dimensions
This read-only property is the preferred way to get the number of
dimensions of a `TensorType`.
"""
def make_variable(self, name=None):
"""Return a `TensorVariable` of this type
:Parameters:
- `name`: str
A pretty name to identify this `Variable` when printing and
debugging
"""
return TensorVariable(self, name=name)
def __str__(self):
if self.name:
return self.name
else:
b = self.broadcastable
named_broadcastable = {(): 'scalar',
(False,): 'vector',
(False, True): 'col',
(True, False): 'row',
(False, False): 'matrix'}
if b in named_broadcastable:
bcast = named_broadcastable[b]
else:
if python_any(b):
bcast = str(b)
else:
bcast = '%iD' % len(b)
return "TensorType(%s, %s)" % (str(self.dtype), bcast)
def __repr__(self):
return str(self)
#"TensorType{%s, %s}" % (str(self.dtype), str(self.broadcastable))
def c_declare(self, name, sub):
"""Override `CLinkerOp.c_declare` """
return """
PyArrayObject* %(name)s;
int type_num_%(name)s;
typedef %(dtype)s dtype_%(name)s;
""" % dict(sub, name=name, dtype=self.dtype_specs()[1])
def c_init(self, name, sub):
"""Override `CLinkerOp.c_init` """
return """
%(name)s = NULL;
type_num_%(name)s = %(type_num)s;
""" % dict(sub, name=name, type_num=self.dtype_specs()[2])
def c_extract(self, name, sub):
"""Override `CLinkerOp.c_extract` """
return """
%(name)s = NULL;
if (py_%(name)s == Py_None) {
// We can either fail here or set %(name)s to NULL and rely on Ops
// using tensors to handle the NULL case, but if they fail to do so
// they'll end up with nasty segfaults, so this is public service.
PyErr_SetString(PyExc_ValueError, "expected an ndarray, not None");
%(fail)s
}
if (!PyArray_Check(py_%(name)s)) {
PyErr_SetString(PyExc_ValueError, "expected an ndarray");
%(fail)s
}
// We expect %(type_num)s
type_num_%(name)s = ((PyArrayObject*)py_%(name)s)->descr->type_num;
if (!PyArray_ISALIGNED(py_%(name)s)) {
PyErr_Format(PyExc_NotImplementedError,
"expected an aligned array of type %%ld "
"(%(type_num)s), got non-aligned array of type %%ld"
" with %%ld dimensions, with 3 last dims "
"%%ld, %%ld, %%ld"
" and 3 last strides %%ld %%ld, %%ld.",
(long int) %(type_num)s,
(long int) type_num_%(name)s,
(long int) PyArray_NDIM(py_%(name)s),
(long int) PyArray_NDIM(py_%(name)s) >= 3 ?
PyArray_DIMS(py_%(name)s)[PyArray_NDIM(py_%(name)s)-3] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 2 ?
PyArray_DIMS(py_%(name)s)[PyArray_NDIM(py_%(name)s)-2] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 1 ?
PyArray_DIMS(py_%(name)s)[PyArray_NDIM(py_%(name)s)-1] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 3 ?
PyArray_STRIDES(py_%(name)s)[PyArray_NDIM(py_%(name)s)-3] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 2 ?
PyArray_STRIDES(py_%(name)s)[PyArray_NDIM(py_%(name)s)-2] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 1 ?
PyArray_STRIDES(py_%(name)s)[PyArray_NDIM(py_%(name)s)-1] : -1
);
%(fail)s
}
// This is a TypeError to be consistent with DEBUG_MODE
// Note: DEBUG_MODE also tells the name of the container
if (type_num_%(name)s != %(type_num)s) {
PyErr_Format(PyExc_TypeError,
"expected type_num %%d (%(type_num)s) got %%d",
%(type_num)s, type_num_%(name)s);
%(fail)s
}
%(name)s = (PyArrayObject*)(py_%(name)s);
Py_XINCREF(%(name)s);
""" % dict(sub, name=name, type_num=self.dtype_specs()[2])
def c_cleanup(self, name, sub):
"""Override `CLinkerOp.c_cleanup` """
return """
if (%(name)s) {
Py_XDECREF(%(name)s);
}
""" % locals()
def c_sync(self, name, sub):
"""Override `CLinkerOp.c_sync` """
fail = sub['fail']
type_num = self.dtype_specs()[2]
return """
{Py_XDECREF(py_%(name)s);}
if (!%(name)s) {
Py_INCREF(Py_None);
py_%(name)s = Py_None;
}
else if ((void*)py_%(name)s != (void*)%(name)s) {
py_%(name)s = (PyObject*)%(name)s;
}
{Py_XINCREF(py_%(name)s);}
if (!PyArray_ISALIGNED(py_%(name)s)) {
PyErr_Format(PyExc_NotImplementedError,
"c_sync: expected an aligned array of type %%ld "
"(%(type_num)s), got non-aligned array of type %%ld"
" with %%ld dimensions, with 3 last dims "
"%%ld, %%ld, %%ld"
" and 3 last strides %%ld %%ld, %%ld.",
(long int) %(type_num)s,
(long int) type_num_%(name)s,
(long int) PyArray_NDIM(py_%(name)s),
(long int) PyArray_NDIM(py_%(name)s) >= 3 ?
PyArray_DIMS(py_%(name)s)[PyArray_NDIM(py_%(name)s)-3] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 2 ?
PyArray_DIMS(py_%(name)s)[PyArray_NDIM(py_%(name)s)-2] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 1 ?
PyArray_DIMS(py_%(name)s)[PyArray_NDIM(py_%(name)s)-1] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 3 ?
PyArray_STRIDES(py_%(name)s)[PyArray_NDIM(py_%(name)s)-3] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 2 ?
PyArray_STRIDES(py_%(name)s)[PyArray_NDIM(py_%(name)s)-2] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 1 ?
PyArray_STRIDES(py_%(name)s)[PyArray_NDIM(py_%(name)s)-1] : -1
);
%(fail)s
}
""" % locals()
def c_headers(self):
"""Override `CLinkerOp.c_headers` """
return scal.Scalar(self.dtype).c_headers()
def c_libraries(self):
return scal.Scalar(self.dtype).c_libraries()
def c_compile_args(self):
return scal.Scalar(self.dtype).c_compile_args()
def c_support_code(self):
"""Override `CLinkerOp.c_support_code` """
return scal.Scalar(self.dtype).c_support_code()
def c_code_cache_version(self):
scalar_version = scal.Scalar(self.dtype).c_code_cache_version()
if scalar_version:
return (9,) + scalar_version
else:
return ()
def value_zeros(self, shape):
"""
Create an numpy ndarray full of 0 values.
"""
return numpy.zeros(shape, dtype=self.dtype)
def get_shape_info(self, obj):
"""
Return the information needed to compute the memory size of ``obj``.
The memory size is only the data, so this excludes the container.
For an ndarray, this is the data, but not the ndarray object and
other data structures such as shape and strides.
``get_shape_info()`` and ``get_size()`` work in tandem for the memory
profiler.
``get_shape_info()`` is called during the execution of the function.
So it is better that it is not too slow.
``get_size()`` will be called on the output of this function
when printing the memory profile.
:param obj: The object that this Type represents during execution
:return: Python object that ``self.get_size()`` understands
"""
return obj.shape
def get_size(self, shape_info):
""" Number of bytes taken by the object represented by shape_info.
:param shape_info: the output of the call to get_shape_info()
:return: the number of bytes taken by the object described by
``shape_info``.
"""
if shape_info:
return numpy.prod(shape_info) * numpy.dtype(self.dtype).itemsize
else: # a scalar
return numpy.dtype(self.dtype).itemsize
theano.compile.ops.expandable_types += (TensorType,)
# Register TensorType C code for ViewOp.
theano.compile.register_view_op_c_code(
TensorType,
"""
Py_XDECREF(%(oname)s);
%(oname)s = %(iname)s;
Py_XINCREF(%(oname)s);
""",
version=1)
# Register TensorType C code for DeepCopyOp
theano.compile.register_deep_copy_op_c_code(
TensorType,
"""
int alloc = %(oname)s == NULL;
for(int i=0; !alloc && i<PyArray_NDIM(%(oname)s); i++) {
if(PyArray_DIMS(%(iname)s)[i] != PyArray_DIMS(%(oname)s)[i]) {
alloc = true;
break;
}
}
if(alloc) {
Py_XDECREF(%(oname)s);
%(oname)s = (PyArrayObject*)PyArray_NewCopy(%(iname)s,
NPY_ANYORDER);
if (!%(oname)s)
{
PyErr_SetString(PyExc_ValueError,
"DeepCopyOp: the copy failed!");
%(fail)s;
}
} else {
if(PyArray_CopyInto(%(oname)s, %(iname)s)){
PyErr_SetString(PyExc_ValueError,
"DeepCopyOp: the copy failed into already allocated space!");
%(fail)s;
}
}
""",
version=2)
# Easy constructors
def tensor(*args, **kwargs):
......
theano/tensor/type.py 0 → 100644
浏览文件 @ b41e3605
import numpy
import theano
from theano.gof import hashtype, Type, Variable
from theano import scalar as scal
class TensorType(Type):
"""Symbolic `Type` representing a numpy.ndarray value."""
filter_checks_isfinite = False
"""
When this is True, strict filtering rejects data containing NaN or
Inf entries. (Used in `DebugMode`)
"""
def __init__(self, dtype, broadcastable, name=None, sparse_grad=False):
"""Initialize self.dtype and self.broadcastable.
:Parameters:
- `dtype`: str corresponding to numpy dtype (e.g., 'int64')
The value (ndarray) associated to a `Variable` of this `Type` will
have this dtype.
- `broadcastable`: tuple, list, or array of boolean values
This argument serves two purposes. First, the True elements of this
list indicate the dimensions where the shape of an associated value
must be 1. Secondly, the length of this list is the number of
dimensions that an associated value must have. See
:doc:`broadcasting` for an explanation of how this list is used.
- `name`: str
Optional name for this type.
"""
self.dtype = str(dtype)
if self.dtype == 'floatX':
self.dtype = config.floatX
### broadcastable is immutable, and all elements are either
### True or False
self.broadcastable = tuple(bool(b) for b in broadcastable)
self.dtype_specs() # error checking is done there
self.name = name
self.numpy_dtype = numpy.dtype(self.dtype)
self.sparse_grad = sparse_grad
if sparse_grad:
warnings.warn(
"DEPRECATION WARNING: You use an old interface to"
" AdvancedSubtensor1 sparse_grad. Now use"
" theano.sparse_grad(a_tensor[an_int_vector]).")
def filter(self, data, strict=False, allow_downcast=None):
"""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
`Linker` instances to use when running a compiled graph.
"""
# Explicit error message when one accidentally uses a Variable as
# input (typical mistake, especially with shared variables).
if isinstance(data, Variable):
raise TypeError(
'Expected an array-like object, but found a Variable: '
'maybe you are trying to call a function on a (possibly '
'shared) variable instead of a numeric array?')
if ((type(data) is numpy.ndarray)
and (data.dtype == self.numpy_dtype)):
if data.dtype.num != self.numpy_dtype.num:
data = theano._asarray(data, dtype=self.dtype)
# -- now fall through to ndim check
elif((type(data) is numpy.memmap)
and (data.dtype == self.numpy_dtype)):
# numpy.memmap is a "safe" subclass of ndarray,
# so we can use it whereever we expect a base ndarray.
# however, casting it would defeat the purpose of not
# loading the whole data into memory
pass
elif strict:
# If any of the two conditions above was not met,
# we raise a meaningful TypeError.
if not (type(data) is numpy.ndarray):
raise TypeError("%s expected a ndarray object." % self,
data, type(data))
if data.dtype != self.numpy_dtype:
raise TypeError(("%s expected a ndarray object with "
"dtype = %s (got %s).") % (
self, self.numpy_dtype, data.dtype))
assert False, "This point should never be reached."
else:
if allow_downcast:
# Convert to self.dtype, regardless of the type of data
data = theano._asarray(data, dtype=self.dtype)
# TODO: consider to pad shape with ones to make it consistent
# with self.broadcastable... like vector->row type thing
else:
if isinstance(data, numpy.ndarray):
# Check if self.dtype can accurately represent data
# (do not try to convert the data)
up_dtype = scal.upcast(self.dtype, data.dtype)
if up_dtype == self.dtype:
# Bug in the following line when data is a
# scalar array, see
# http://projects.scipy.org/numpy/ticket/1611
# data = data.astype(self.dtype)
data = theano._asarray(data, dtype=self.dtype)
if up_dtype != self.dtype:
err_msg = (
'%s cannot store a value of dtype %s without '
'risking loss of precision. If you do not mind '
'this loss, you can: '
'1) explicitly cast your data to %s, or '
'2) set "allow_input_downcast=True" when calling '
'"function".'
% (self, data.dtype, self.dtype))
raise TypeError(err_msg, data)
elif (allow_downcast is None and
type(data) is float and
self.dtype == theano.config.floatX):
# Special case where we allow downcasting of Python float
# literals to floatX, even when floatX=='float32'
data = theano._asarray(data, self.dtype)
else:
# data has to be converted.
# Check that this conversion is lossless
converted_data = theano._asarray(data, self.dtype)
# We use the `values_eq` static function from TensorType
# to handle NaN values.
if TensorType.values_eq(numpy.asarray(data),
converted_data,
force_same_dtype=False):
data = converted_data
else:
# Do not print a too long description of data
# (ndarray truncates it, but it's not sure for data)
str_data = str(data)
if len(str_data) > 80:
str_data = str_data[:75] + '(...)'
err_msg = (
'%s cannot store accurately value %s, '
'it would be represented as %s. '
'If you do not mind this precision loss, you can: '
'1) explicitly convert your data to a numpy array '
'of dtype %s, or '
'2) set "allow_input_downcast=True" when calling '
'"function".'
% (self, data, converted_data, self.dtype))
raise TypeError(err_msg, data)
if self.ndim != data.ndim:
raise TypeError("Wrong number of dimensions: expected %s,"
" got %s with shape %s." % (self.ndim, data.ndim,
data.shape))
if not data.flags.aligned:
try:
msg = "object buffer" + str(data.data)
except AttributeError:
msg = ""
raise TypeError("The numpy.ndarray object is not aligned."
" Theano C code does not support that.",
msg,
"object shape", data.shape,
"object strides", data.strides)
i = 0
for b in self.broadcastable:
if b and data.shape[i] != 1:
raise TypeError("Non-unit value on shape on a broadcastable"
" dimension.", data.shape, self.broadcastable)
i += 1
if (self.filter_checks_isfinite and
not numpy.all(numpy.isfinite(data))):
raise ValueError("non-finite elements not allowed")
return data
def filter_variable(self, other):
"""Convert a symbolic Variable into a TensorType, if compatible.
For the moment, only a TensorType or CudaNdarrayType will be
converted, provided they have the same number of dimensions,
broadcastable pattern, and dtype.
"""
if hasattr(other, '_as_TensorVariable'):
other = other._as_TensorVariable()
if not isinstance(other, Variable):
# The value is not a Variable: we cast it into
# a Constant of the appropriate Type.
other = self.Constant(type=self, data=other)
if other.type == self:
return other
raise TypeError(
'Cannot convert Type %(othertype)s '
'(of Variable %(other)s) into Type %(self)s. '
'You can try to manually convert %(other)s into a %(self)s.'
% dict(
othertype=other.type,
other=other,
self=self)
)
def value_validity_msg(self, a):
try:
self.filter(a, strict=True)
except Exception, e:
return str(e)
return "value is valid"
def dtype_specs(self):
"""Return a tuple (python type, c type, numpy typenum) that corresponds
to self.dtype.
This function is used internally as part of C code generation.
"""
# TODO: add more type correspondances for e.g. int32, int64, float32,
# complex64, etc.
try:
return {
'float32': (float, 'npy_float32', 'NPY_FLOAT32'),
'float64': (float, 'npy_float64', 'NPY_FLOAT64'),
'uint8': (int, 'npy_uint8', 'NPY_UINT8'),
'int8': (int, 'npy_int8', 'NPY_INT8'),
'uint16': (int, 'npy_uint16', 'NPY_UINT16'),
'int16': (int, 'npy_int16', 'NPY_INT16'),
'uint32': (int, 'npy_uint32', 'NPY_UINT32'),
'int32': (int, 'npy_int32', 'NPY_INT32'),
'uint64': (int, 'npy_uint64', 'NPY_UINT64'),
'int64': (int, 'npy_int64', 'NPY_INT64'),
'complex128': (complex, 'theano_complex128', 'NPY_COMPLEX128'),
'complex64': (complex, 'theano_complex64', 'NPY_COMPLEX64')
}[self.dtype]
except KeyError:
raise TypeError("Unsupported dtype for %s: %s"
% (self.__class__.__name__, self.dtype))
def to_scalar_type(self):
return scal.Scalar(dtype=self.dtype)
def __eq__(self, other):
"""Compare True iff other is the same kind of TensorType"""
return type(self) == type(other) and other.dtype == self.dtype \
and other.broadcastable == self.broadcastable
@staticmethod
def may_share_memory(a, b):
# This is a method of TensorType, so both a and b should be ndarrays
if isinstance(a, numpy.ndarray) and isinstance(b, numpy.ndarray):
return numpy.may_share_memory(a, b)
else:
return False
@staticmethod
def values_eq(a, b, force_same_dtype=True):
# TODO: check to see if the shapes must match
# for now, we err on safe side...
if a.shape != b.shape:
return False
if force_same_dtype and a.dtype != b.dtype:
return False
a_eq_b = (a == b)
r = numpy.all(a_eq_b)
if r:
return True
# maybe the trouble is that there are NaNs
a_missing = numpy.isnan(a)
if a_missing.any():
b_missing = numpy.isnan(b)
return numpy.all(a_eq_b + (a_missing == b_missing))
else:
return False
@staticmethod
def values_eq_approx(a, b, allow_remove_inf=False, allow_remove_nan=False,
rtol=None, atol=None):
"""
:param allow_remove_inf: If True, when there is an inf in a,
we allow any value in b in that position.
Event -inf
:param allow_remove_nan: If True, when there is a nan in a,
we allow any value in b in that position.
Event +-inf
:param rtol: relative tolerance, passed to _allclose
:param atol: absolute tolerance, passed to _allclose
"""
if isinstance(a, numpy.ndarray) and isinstance(b, numpy.ndarray):
if a.shape != b.shape:
return False
if a.dtype != b.dtype:
return False
if 'int' in str(a.dtype):
return numpy.all(a == b)
else:
# work around a numpy.allclose bug:
# http://projects.scipy.org/numpy/ticket/1672
if a.ndim == 0 and numpy.isinf(a):
a = a.reshape(1)
b = b.reshape(1)
cmp = _allclose(a, b, rtol=rtol, atol=atol)
if cmp:
# Numpy claims they are close, this is good enough for us.
return True
# Numpy is unhappy, but it does not necessarily mean that a and
# b are different. Indeed, Numpy does not like missing values
# and will return False whenever some are found in a or b.
# The proper way would be to use the MaskArray stuff available
# in Numpy. However, it looks like it has been added to Numpy's
# core recently, so it may not be available to everyone. Thus,
# for now we use a home-made recipe, that should probably be
# revisited in the future.
a_missing = numpy.isnan(a)
a_inf = numpy.isinf(a)
if not (a_missing.any() or (allow_remove_inf and a_inf.any())):
# There are no missing values in a, thus this is not the
# reason why numpy.allclose(a, b) returned False.
_logger.info(
'numpy allclose failed for abs_err %f and rel_err %f',
numpy.max(abs(a - b)),
numpy.max(abs(a - b) / (abs(a) + abs(b))))
return False
# The following line is what numpy.allclose bases its decision
# upon, according to its documentation.
rtol = 1.0000000000000001e-05
atol = 1e-8
cmp_elemwise = (numpy.absolute(a - b) <=
(atol + rtol * numpy.absolute(b)))
# Find places where both a and b have missing values.
both_missing = a_missing * numpy.isnan(b)
# Find places where both a and b have inf of the same sign.
both_inf = a_inf * numpy.isinf(b)
# cmp_elemwise is weird when we have inf and -inf.
# set it to False
cmp_elemwise = numpy.where(
both_inf & cmp_elemwise,
a == b,
cmp_elemwise)
# check the sign of the inf
both_inf = numpy.where(both_inf, (a == b), both_inf)
if allow_remove_inf:
both_inf += a_inf
if allow_remove_nan:
both_missing += a_missing
# Combine all information.
return (cmp_elemwise + both_missing + both_inf).all()
return False
@staticmethod
def values_eq_approx_remove_inf(a, b):
return TensorType.values_eq_approx(a, b, True)
@staticmethod
def values_eq_approx_remove_nan(a, b):
return TensorType.values_eq_approx(a, b, False, True)
@staticmethod
def values_eq_approx_remove_inf_nan(a, b):
return TensorType.values_eq_approx(a, b, True, True)
def __hash__(self):
"""Hash equal for same kinds of TensorType"""
return hashtype(self) ^ hash(self.dtype) ^ hash(self.broadcastable)
ndim = property(lambda self: len(self.broadcastable),
doc="number of dimensions")
"""Number of dimensions
This read-only property is the preferred way to get the number of
dimensions of a `TensorType`.
"""
def make_variable(self, name=None):
"""Return a `TensorVariable` of this type
:Parameters:
- `name`: str
A pretty name to identify this `Variable` when printing and
debugging
"""
return self.Variable(self, name=name)
def __str__(self):
if self.name:
return self.name
else:
b = self.broadcastable
named_broadcastable = {(): 'scalar',
(False,): 'vector',
(False, True): 'col',
(True, False): 'row',
(False, False): 'matrix'}
if b in named_broadcastable:
bcast = named_broadcastable[b]
else:
if python_any(b):
bcast = str(b)
else:
bcast = '%iD' % len(b)
return "TensorType(%s, %s)" % (str(self.dtype), bcast)
def __repr__(self):
return str(self)
#"TensorType{%s, %s}" % (str(self.dtype), str(self.broadcastable))
def c_declare(self, name, sub):
"""Override `CLinkerOp.c_declare` """
return """
PyArrayObject* %(name)s;
int type_num_%(name)s;
typedef %(dtype)s dtype_%(name)s;
""" % dict(sub, name=name, dtype=self.dtype_specs()[1])
def c_init(self, name, sub):
"""Override `CLinkerOp.c_init` """
return """
%(name)s = NULL;
type_num_%(name)s = %(type_num)s;
""" % dict(sub, name=name, type_num=self.dtype_specs()[2])
def c_extract(self, name, sub):
"""Override `CLinkerOp.c_extract` """
return """
%(name)s = NULL;
if (py_%(name)s == Py_None) {
// We can either fail here or set %(name)s to NULL and rely on Ops
// using tensors to handle the NULL case, but if they fail to do so
// they'll end up with nasty segfaults, so this is public service.
PyErr_SetString(PyExc_ValueError, "expected an ndarray, not None");
%(fail)s
}
if (!PyArray_Check(py_%(name)s)) {
PyErr_SetString(PyExc_ValueError, "expected an ndarray");
%(fail)s
}
// We expect %(type_num)s
type_num_%(name)s = ((PyArrayObject*)py_%(name)s)->descr->type_num;
if (!PyArray_ISALIGNED(py_%(name)s)) {
PyErr_Format(PyExc_NotImplementedError,
"expected an aligned array of type %%ld "
"(%(type_num)s), got non-aligned array of type %%ld"
" with %%ld dimensions, with 3 last dims "
"%%ld, %%ld, %%ld"
" and 3 last strides %%ld %%ld, %%ld.",
(long int) %(type_num)s,
(long int) type_num_%(name)s,
(long int) PyArray_NDIM(py_%(name)s),
(long int) PyArray_NDIM(py_%(name)s) >= 3 ?
PyArray_DIMS(py_%(name)s)[PyArray_NDIM(py_%(name)s)-3] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 2 ?
PyArray_DIMS(py_%(name)s)[PyArray_NDIM(py_%(name)s)-2] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 1 ?
PyArray_DIMS(py_%(name)s)[PyArray_NDIM(py_%(name)s)-1] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 3 ?
PyArray_STRIDES(py_%(name)s)[PyArray_NDIM(py_%(name)s)-3] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 2 ?
PyArray_STRIDES(py_%(name)s)[PyArray_NDIM(py_%(name)s)-2] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 1 ?
PyArray_STRIDES(py_%(name)s)[PyArray_NDIM(py_%(name)s)-1] : -1
);
%(fail)s
}
// This is a TypeError to be consistent with DEBUG_MODE
// Note: DEBUG_MODE also tells the name of the container
if (type_num_%(name)s != %(type_num)s) {
PyErr_Format(PyExc_TypeError,
"expected type_num %%d (%(type_num)s) got %%d",
%(type_num)s, type_num_%(name)s);
%(fail)s
}
%(name)s = (PyArrayObject*)(py_%(name)s);
Py_XINCREF(%(name)s);
""" % dict(sub, name=name, type_num=self.dtype_specs()[2])
def c_cleanup(self, name, sub):
"""Override `CLinkerOp.c_cleanup` """
return """
if (%(name)s) {
Py_XDECREF(%(name)s);
}
""" % locals()
def c_sync(self, name, sub):
"""Override `CLinkerOp.c_sync` """
fail = sub['fail']
type_num = self.dtype_specs()[2]
return """
{Py_XDECREF(py_%(name)s);}
if (!%(name)s) {
Py_INCREF(Py_None);
py_%(name)s = Py_None;
}
else if ((void*)py_%(name)s != (void*)%(name)s) {
py_%(name)s = (PyObject*)%(name)s;
}
{Py_XINCREF(py_%(name)s);}
if (!PyArray_ISALIGNED(py_%(name)s)) {
PyErr_Format(PyExc_NotImplementedError,
"c_sync: expected an aligned array of type %%ld "
"(%(type_num)s), got non-aligned array of type %%ld"
" with %%ld dimensions, with 3 last dims "
"%%ld, %%ld, %%ld"
" and 3 last strides %%ld %%ld, %%ld.",
(long int) %(type_num)s,
(long int) type_num_%(name)s,
(long int) PyArray_NDIM(py_%(name)s),
(long int) PyArray_NDIM(py_%(name)s) >= 3 ?
PyArray_DIMS(py_%(name)s)[PyArray_NDIM(py_%(name)s)-3] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 2 ?
PyArray_DIMS(py_%(name)s)[PyArray_NDIM(py_%(name)s)-2] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 1 ?
PyArray_DIMS(py_%(name)s)[PyArray_NDIM(py_%(name)s)-1] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 3 ?
PyArray_STRIDES(py_%(name)s)[PyArray_NDIM(py_%(name)s)-3] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 2 ?
PyArray_STRIDES(py_%(name)s)[PyArray_NDIM(py_%(name)s)-2] : -1,
(long int) PyArray_NDIM(py_%(name)s) >= 1 ?
PyArray_STRIDES(py_%(name)s)[PyArray_NDIM(py_%(name)s)-1] : -1
);
%(fail)s
}
""" % locals()
def c_headers(self):
"""Override `CLinkerOp.c_headers` """
return scal.Scalar(self.dtype).c_headers()
def c_libraries(self):
return scal.Scalar(self.dtype).c_libraries()
def c_compile_args(self):
return scal.Scalar(self.dtype).c_compile_args()
def c_support_code(self):
"""Override `CLinkerOp.c_support_code` """
return scal.Scalar(self.dtype).c_support_code()
def c_code_cache_version(self):
scalar_version = scal.Scalar(self.dtype).c_code_cache_version()
if scalar_version:
return (9,) + scalar_version
else:
return ()
def value_zeros(self, shape):
"""
Create an numpy ndarray full of 0 values.
"""
return numpy.zeros(shape, dtype=self.dtype)
def get_shape_info(self, obj):
"""
Return the information needed to compute the memory size of ``obj``.
The memory size is only the data, so this excludes the container.
For an ndarray, this is the data, but not the ndarray object and
other data structures such as shape and strides.
``get_shape_info()`` and ``get_size()`` work in tandem for the memory
profiler.
``get_shape_info()`` is called during the execution of the function.
So it is better that it is not too slow.
``get_size()`` will be called on the output of this function
when printing the memory profile.
:param obj: The object that this Type represents during execution
:return: Python object that ``self.get_size()`` understands
"""
return obj.shape
def get_size(self, shape_info):
""" Number of bytes taken by the object represented by shape_info.
:param shape_info: the output of the call to get_shape_info()
:return: the number of bytes taken by the object described by
``shape_info``.
"""
if shape_info:
return numpy.prod(shape_info) * numpy.dtype(self.dtype).itemsize
else: # a scalar
return numpy.dtype(self.dtype).itemsize
theano.compile.ops.expandable_types += (TensorType,)
# Register TensorType C code for ViewOp.
theano.compile.register_view_op_c_code(
TensorType,
"""
Py_XDECREF(%(oname)s);
%(oname)s = %(iname)s;
Py_XINCREF(%(oname)s);
""",
version=1)
# Register TensorType C code for DeepCopyOp
theano.compile.register_deep_copy_op_c_code(
TensorType,
"""
int alloc = %(oname)s == NULL;
for(int i=0; !alloc && i<PyArray_NDIM(%(oname)s); i++) {
if(PyArray_DIMS(%(iname)s)[i] != PyArray_DIMS(%(oname)s)[i]) {
alloc = true;
break;
}
}
if(alloc) {
Py_XDECREF(%(oname)s);
%(oname)s = (PyArrayObject*)PyArray_NewCopy(%(iname)s,
NPY_ANYORDER);
if (!%(oname)s)
{
PyErr_SetString(PyExc_ValueError,
"DeepCopyOp: the copy failed!");
%(fail)s;
}
} else {
if(PyArray_CopyInto(%(oname)s, %(iname)s)){
PyErr_SetString(PyExc_ValueError,
"DeepCopyOp: the copy failed into already allocated space!");
%(fail)s;
}
}
""",
version=2)
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