Skip to content

P
pytensor
提交 bf9413c8 authored 作者: Pascal Lamblin's avatar Pascal Lamblin 提交者： GitHub
浏览文件
操作

Merge pull request #5091 from abergeron/bool

Add bool dtype in scalar and tensor.
上级 58164835 46a30357
隐藏空白字符变更
内嵌 并排
正在显示 包含 368 行增加 和 568 行删除
doc/tutorial/gradients.txt
浏览文件 @ bf9413c8
......@@ -124,7 +124,7 @@ do is to loop over the entries in *y* and compute the gradient of
>>> import theano.tensor as T
>>> x = T.dvector('x')
>>> y = x ** 2
>>> J, updates = theano.scan(lambda i, y,x : T.grad(y[i], x), sequences=T.arange(y.shape[0]), non_sequences=[y,x])
>>> J, updates = theano.scan(lambda i, y, x : T.grad(y[i], x), sequences=T.arange(y.shape[0]), non_sequences=[y, x])
>>> f = theano.function([x], J, updates=updates)
>>> f([4, 4])
array([[ 8., 0.],
......
setup.py
浏览文件 @ bf9413c8
......@@ -53,7 +53,7 @@ PLATFORMS = ["Windows", "Linux", "Solaris", "Mac OS-X", "Unix"]
MAJOR = 0
MINOR = 9
MICRO = 0
SUFFIX = "dev3" # Should be blank except for rc's, betas, etc.
SUFFIX = "dev4" # Should be blank except for rc's, betas, etc.
ISRELEASED = False
VERSION = '%d.%d.%d%s' % (MAJOR, MINOR, MICRO, SUFFIX)
......
theano/gof/tests/test_fg.py
浏览文件 @ bf9413c8
from __future__ import absolute_import, print_function, division
import os
import pickle
import sys
import unittest
from nose.plugins.skip import SkipTest
import theano
from theano.compat import PY3
from theano.gof import CachedConstantError, FunctionGraph
......@@ -32,14 +29,11 @@ class TFunctionGraph(unittest.TestCase):
pickle.loads(s)
def test_node_outputs_not_used(self):
"""In the past, we where removing some not used variable from
fgraph.variables event if the apply had other output used in
the graph. This caused a crash.
This test run the pickle that reproduce this case.
"""
if sys.version_info[:2] < (2, 7):
raise SkipTest("This test need python 2.7 or more recent.")
# In the past, we where removing some not used variable from
# fgraph.variables event if the apply had other output used in
# the graph. This caused a crash.
# This test run the pickle that reproduce this case.
with open(os.path.join(os.path.dirname(__file__),
'test_fg_old_crash.pkl'),
'rb') as f:
......
theano/gpuarray/elemwise.py
浏览文件 @ bf9413c8
......@@ -129,7 +129,8 @@ class GpuElemwise(HideC, Elemwise):
support_code += fake_node.op.c_support_code()
except MethodNotDefined:
pass
for npy, ga in [("npy_uint8", "ga_ubyte"),
for npy, ga in [("npy_bool", "ga_bool"),
("npy_uint8", "ga_ubyte"),
("npy_uint16", "ga_ushort"),
("npy_uint32", "ga_uint"),
("npy_uint64", "ga_ulong"),
......
theano/gpuarray/type.py
浏览文件 @ bf9413c8
......@@ -408,6 +408,7 @@ class GpuArrayType(Type):
'float16': (float, 'npy_float16', 'NPY_FLOAT16'),
'float32': (float, 'npy_float32', 'NPY_FLOAT32'),
'float64': (float, 'npy_float64', 'NPY_FLOAT64'),
'bool': (int, 'npy_bool', 'NPY_BOOL'),
'uint8': (int, 'npy_uint8', 'NPY_UINT8'),
'int8': (int, 'npy_int8', 'NPY_INT8'),
'uint16': (int, 'npy_uint16', 'NPY_UINT16'),
......
theano/gradient.py
浏览文件 @ bf9413c8
......@@ -477,7 +477,7 @@ def grad(cost, wrt, consider_constant=None,
# function, sure, but nonetheless one we can and should support.
# So before we try to cast it make sure it even has a dtype
if (hasattr(g_cost.type, 'dtype') and
cost.type.dtype not in tensor.discrete_dtypes):
cost.type.dtype in tensor.continuous_dtypes):
# Here we enforce the constraint that floating point variables
# have the same dtype as their gradient.
g_cost = g_cost.astype(cost.type.dtype)
......@@ -485,7 +485,7 @@ def grad(cost, wrt, consider_constant=None,
# This is to be enforced by the Op.grad method for the
# Op that outputs cost.
if hasattr(g_cost.type, 'dtype'):
assert g_cost.type.dtype not in tensor.discrete_dtypes
assert g_cost.type.dtype in tensor.continuous_dtypes
grad_dict[cost] = g_cost
......@@ -1335,12 +1335,11 @@ def _float_ones_like(x):
""" Like ones_like, but forces the object to have a
floating point dtype """
rval = tensor.ones_like(x)
dtype = x.type.dtype
if dtype not in tensor.float_dtypes:
dtype = theano.config.floatX
if rval.type.dtype.find('float') != -1:
return rval
return rval.astype(theano.config.floatX)
return tensor.ones_like(x, dtype=dtype)
class numeric_grad(object):
......
theano/sandbox/cuda/opt.py
浏览文件 @ bf9413c8
......@@ -237,8 +237,11 @@ optdb['canonicalize'].register('local_cut_gpu_host_gpu',
# 'float64', 'complex128' and 'complex64' are not supported in elemwise
# on the gpu.
elemwise_cuda_dtype_supported = ['float32', 'uint8', 'int8', 'uint16', 'int16',
'uint32', 'int32', 'uint64', 'int64']
elemwise_cuda_dtype_supported = ['float32', 'bool',
'uint8', 'int8',
'uint16', 'int16',
'uint32', 'int32',
'uint64', 'int64']
def dtype_in_elemwise_supported(op):
......@@ -298,8 +301,8 @@ def local_gpu_elemwise_0(node):
return False
# first establish that float32 can store all inputs
upcastable = set(['float32', 'int8', 'int16', 'uint8',
'uint16'])
upcastable = set(['float32', 'bool', 'int8', 'int16',
'uint8', 'uint16'])
# case 1 - all inputs are already float32
if all([i.type.dtype == 'float32' for i in node.inputs]):
# TODO: change this when fusion makes Elemwise with
......
theano/sandbox/cuda/rng_curand.py
浏览文件 @ bf9413c8
......@@ -28,7 +28,7 @@ class CURAND_Base(GpuOp):
CURAND. This Op uses a generic-typed shared variable to point to a CObject
that encapsulates this opaque reference.
Each random variable is created with a generator of False.
Each random variable is created with a generator of None.
The actual random number generator is allocated from the seed, on the first
call to allocate random numbers (see c_code).
......@@ -210,7 +210,7 @@ class CURAND_Base(GpuOp):
%(fail)s;
}
%(o_generator)s = PyCObject_FromVoidPtr(gen, &free_generator);
assert (%(i_generator)s == Py_False);
assert (%(i_generator)s == Py_None);
}
else if (%(destructive)s)
{
......@@ -244,7 +244,7 @@ class CURAND_Base(GpuOp):
return code
def c_code_cache_version(self):
return (4,)
return (5,)
class CURAND_Normal(CURAND_Base):
......@@ -328,7 +328,7 @@ class CURAND_RandomStreams(object):
else:
msg = "size must be a tuple of int or a Theano variable"
assert isinstance(size, Variable) and size.ndim == 1, msg
generator = theano.shared(False) # makes a generic
generator = theano.shared(None) # makes a generic
s_size = theano.tensor.as_tensor_variable(size)
u = CURAND_Uniform.new_auto_update(generator, ndim, dtype, s_size,
self.next_seed())
......@@ -360,7 +360,7 @@ class CURAND_RandomStreams(object):
else:
msg = "size must be a tuple of int or a Theano variable"
assert isinstance(size, Variable) and size.ndim == 1, msg
generator = theano.shared(False) # makes a generic
generator = theano.shared(None) # makes a generic
s_size = theano.tensor.as_tensor_variable(size)
u = CURAND_Normal.new_auto_update(generator, ndim, dtype, s_size,
self.next_seed())
......
theano/scalar/basic.py
浏览文件 @ bf9413c8
......@@ -34,6 +34,7 @@ from theano.gradient import grad_undefined
from theano.printing import pprint
import collections
builtin_bool = bool
builtin_complex = complex
builtin_int = int
builtin_float = float
......@@ -161,7 +162,7 @@ class Scalar(Type):
TODO: refactor to be named ScalarType for consistency with TensorType.
"""
__props__ = ('dtype',)
ndim = 0
def __init__(self, dtype):
......@@ -200,6 +201,8 @@ class Scalar(Type):
def values_eq_approx(self, a, b, tolerance=1e-4):
# The addition have risk of overflow especially with [u]int8
if self.dtype == 'bool':
return a == b
diff = a - b
if diff == 0:
return True
......@@ -227,12 +230,6 @@ class Scalar(Type):
else:
return []
def __eq__(self, other):
return type(self) == type(other) and other.dtype == self.dtype
def __hash__(self):
return hash('theano.scalar.Scalar') ^ hash(self.dtype)
def dtype_specs(self):
try:
# To help debug dtype/typenum problem, here is code to get
......@@ -244,7 +241,8 @@ class Scalar(Type):
# now, as Theano always expect the exact typenum that
# correspond to our supported dtype.
"""
for dtype in ['int8', 'uint8', 'short', 'ushort', 'intc', 'uintc',
for dtype in ['bool', 'int8', 'uint8', 'short', 'ushort', 'intc',
'uintc',
'longlong', 'ulonglong', 'single', 'double',
'longdouble', 'csingle', 'cdouble', 'clongdouble',
'float32', 'float64', 'int8', 'int16', 'int32',
......@@ -260,6 +258,7 @@ class Scalar(Type):
'complex128': (numpy.complex128, 'theano_complex128',
'Complex128'),
'complex64': (numpy.complex64, 'theano_complex64', 'Complex64'),
'bool': (numpy.bool_, 'npy_bool', 'Bool'),
'uint8': (numpy.uint8, 'npy_uint8', 'UInt8'),
'int8': (numpy.int8, 'npy_int8', 'Int8'),
'uint16': (numpy.uint16, 'npy_uint16', 'UInt16'),
......@@ -288,12 +287,13 @@ class Scalar(Type):
def c_literal(self, data):
if 'complex' in self.dtype:
raise NotImplementedError("No literal for complex values.")
if self.dtype == 'bool':
return '1' if data else '0'
return str(data)
def c_declare(self, name, sub, check_input=True):
if(check_input):
pre = """
typedef %(dtype)s %(name)s_dtype; // Deprecated use dtype_%(name)s instead.
typedef %(dtype)s dtype_%(name)s;
""" % dict(name=name, dtype=self.dtype_specs()[1])
else:
......@@ -309,6 +309,7 @@ class Scalar(Type):
def c_extract(self, name, sub, check_input=True):
if self.dtype == 'float16':
# This doesn't work at the numpy level
raise NotImplementedError('float16')
specs = self.dtype_specs()
if(check_input):
......@@ -517,6 +518,7 @@ theano.compile.register_view_op_c_code(
1)
bool = get_scalar_type('bool')
int8 = get_scalar_type('int8')
int16 = get_scalar_type('int16')
int32 = get_scalar_type('int32')
......@@ -536,7 +538,8 @@ uint_types = uint8, uint16, uint32, uint64
float_types = float16, float32, float64
complex_types = complex64, complex128
discrete_types = int_types + uint_types
integer_types = int_types + uint_types
discrete_types = (bool,) + integer_types
continuous_types = float_types + complex_types
all_types = discrete_types + continuous_types
......@@ -681,38 +684,57 @@ complexs64 = _multi(complex64)
complexs128 = _multi(complex128)
# Using a class instead of a function makes it possible to deep-copy it in
# Python 2.4.
# Note that currently only a few functions use this mechanism, because it is
# enough to make the test-suite pass with Python 2.4. However, it may prove
# necessary to use this same mechanism in other places as well in the future.
class upcast_out(object):
def __new__(self, *types):
dtype = Scalar.upcast(*types)
return get_scalar_type(dtype),
def upcast_out(*types):
dtype = Scalar.upcast(*types)
return get_scalar_type(dtype),
class upgrade_to_float(object):
def __new__(self, *types):
"""
Upgrade any int types to float32 or float64 to avoid losing precision.
def upcast_out_nobool(*types):
type = upcast_out(*types)
if type[0] == bool:
raise TypeError("bool output not supported")
return type
"""
conv = {int8: float32,
int16: float32,
int32: float64,
int64: float64,
uint8: float32,
uint16: float32,
uint32: float64,
uint64: float64}
return get_scalar_type(Scalar.upcast(*[conv.get(type, type)
for type in types])),
def upcast_out_min8(*types):
type = upcast_out(*types)
if type[0] == bool:
return int8,
return type
class same_out(object):
def __new__(self, type):
return type,
def upgrade_to_float(*types):
"""
Upgrade any int types to float32 or float64 to avoid losing precision.
"""
conv = {bool: float32,
int8: float32,
int16: float32,
int32: float64,
int64: float64,
uint8: float32,
uint16: float32,
uint32: float64,
uint64: float64}
return get_scalar_type(Scalar.upcast(*[conv.get(type, type)
for type in types])),
def same_out(type):
return type,
def same_out_nobool(type):
if type == bool:
raise TypeError("bool input not supported")
return type,
def same_out_min8(type):
if type == bool:
return int8,
return type,
def upcast_out_no_complex(*types):
......@@ -728,6 +750,8 @@ def same_out_float_only(type):
class transfer_type(gof.utils.object2):
__props__ = ('transfer',)
def __init__(self, *transfer):
assert all(type(x) in [int, str] or x is None for x in transfer)
self.transfer = transfer
......@@ -748,26 +772,16 @@ class transfer_type(gof.utils.object2):
return retval
# return [upcast if i is None else types[i] for i in self.transfer]
def __eq__(self, other):
return type(self) == type(other) and self.transfer == other.transfer
def __hash__(self):
return hash(self.transfer)
class specific_out(gof.utils.object2):
__props__ = ('spec',)
def __init__(self, *spec):
self.spec = spec
def __call__(self, *types):
return self.spec
def __eq__(self, other):
return type(self) == type(other) and self.spec == other.spec
def __hash__(self):
return hash(self.spec)
def int_out(*types):
return int64,
......@@ -914,15 +928,15 @@ class ScalarOp(Op):
return test
def __hash__(self):
return hash(type(self).__name__) ^ hash(
getattr(self, 'output_types_preference', 0))
return hash((type(self),
getattr(self, 'output_types_preference', 0)))
def __str__(self):
if hasattr(self, 'name') and self.name:
return self.name
else:
param = [(k, v) for k, v in self.__dict__.items()
if k not in ["name", "_op_use_c_code",
if k not in ["name", "_op_use_c_code", "bool",
"output_types_preference"]]
if param:
return "%s{%s}" % (self.__class__.__name__,
......@@ -1006,31 +1020,65 @@ class BinaryScalarOp(ScalarOp):
###############
class LogicalComparison(BinaryScalarOp):
def __init__(self, *args, **kwargs):
BinaryScalarOp.__init__(self, *args, **kwargs)
# This is for compat with old pickles.
self.bool = True
def __eq__(self, other):
return (BinaryScalarOp.__eq__(self, other) and
getattr(self, 'bool', False) == getattr(self, 'bool', False))
def __hash__(self):
# bool should always be True
return BinaryScalarOp.__hash__(self)
def output_types(self, *input_dtypes):
return [int8]
return [bool] if getattr(self, 'bool', False) else [int8]
def grad(self, inputs, output_gradients):
x, y = inputs
out = self(x, y)
assert str(out.type.dtype).find('int') != -1
assert out.type == bool
return [x.zeros_like().astype(theano.config.floatX),
y.zeros_like().astype(theano.config.floatX)]
def c_code_cache_version(self):
super_version = super(LogicalComparison, self).c_code_cache_version()
return super_version + (0,)
class FixedLogicalComparison(UnaryScalarOp):
"""
Comparison to a fixed value.
"""
def __init__(self, *args, **kwargs):
UnaryScalarOp.__init__(self, *args, **kwargs)
# This is for compat with old pickles
self.bool = True
def __eq__(self, other):
return (UnaryScalarOp.__eq__(self, other) and
getattr(self, 'bool', False) == getattr(self, 'bool', False))
def __hash__(self):
# bool should always be True
return UnaryScalarOp.__hash__(self)
def output_types(self, *input_dtypes):
return [int8]
return [bool] if getattr(self, 'bool', False) else [int8]
def grad(self, inputs, output_gradients):
x, = inputs
out = self(x)
assert str(out.type.dtype).find('int') != -1
assert out.type == bool
return [x.zeros_like().astype(theano.config.floatX)]
def c_code_cache_version(self):
super_version = super(FixedLogicalComparison, self).c_code_cache_version()
return super_version + (0,)
class LT(LogicalComparison):
identity = False
......@@ -1202,21 +1250,10 @@ class InRange(LogicalComparison):
def c_code(self, node, name, inputs, outputs, sub):
(x, low, hi) = inputs
(z,) = outputs
if self.openlow:
cmp1 = '>'
else:
cmp1 = '>='
# backport
# cmp1 = '>' if self.openlow else '>='
if self.openhi:
cmp2 = '<'
else:
cmp2 = '<='
cmp1 = '>' if self.openlow else '>='
cmp2 = '<' if self.openhi else '<='
# backport
# cmp2 = '<' if self.openhi else '<='
return ("%(z)s = %(x)s %(cmp1)s %(low)s &&"
" %(x)s %(cmp2)s %(hi)s;" % locals())
......@@ -1247,13 +1284,8 @@ class Switch(ScalarOp):
nfunc_spec = ('where', 3, 1)
def impl(self, cond, ift, iff):
if cond:
return ift
else:
return iff
return ift if cond else iff
# backport
# return ift if cond else iff
def c_code(self, node, name, inputs, outputs, sub):
(cond, ift, iff) = inputs
(z,) = outputs
......@@ -1290,9 +1322,9 @@ switch = Switch()
class UnaryBitOp(UnaryScalarOp):
def output_types(self, *input_types):
for i in input_types[0]:
if i not in (int8, int16, int32, int64):
raise TypeError('input to a BitOp must have type int8,'
' int16, int32 or int64... not %s' % i)
if i not in discrete_types:
raise TypeError('input to a BitOp must have type (u)int8, '
'(u)int16, (u)int32 or (u)int64 or bool not %s' % i)
return upcast_out(*input_types[0])
def grad(self, inputs, output_gradients):
......@@ -1302,10 +1334,13 @@ class UnaryBitOp(UnaryScalarOp):
class BinaryBitOp(BinaryScalarOp):
def output_types(self, *input_types):
t0, t1 = input_types[0]
if t0 == bool and t1 == bool:
return [bool]
for i in input_types[0]:
if i not in (int8, int16, int32, int64):
raise TypeError('input to a BitOp must have type int8,'
' int16, int32 or int64... not %s' % i)
if i not in integer_types:
raise TypeError('input to a BitOp must have type (u)int8, '
'(u)int16, (u)int32 or (u)int64 or '
'be all bools not %s' % i)
return upcast_out(*input_types[0])
def grad(self, inputs, output_gradients):
......@@ -1371,6 +1406,8 @@ class Invert(UnaryBitOp):
def c_code(self, node, name, inputs, outputs, sub):
(x,) = inputs
(z,) = outputs
if node.outputs[0].type == bool:
return "%(z)s = (!%(x)s);" % locals()
return "%(z)s = (~%(x)s);" % locals()
invert = Invert()
......@@ -1463,10 +1500,13 @@ class Add(ScalarOp):
def c_code(self, node, name, inputs, outputs, sub):
(z,) = outputs
op = " + "
if node.outputs[0].type == bool:
op = " || "
if not inputs:
return z + " = 0;"
else:
return z + " = " + " + ".join(inputs) + ";"
return z + " = " + op.join(inputs) + ";"
def grad(self, inputs, gout):
(gz,) = gout
......@@ -1502,10 +1542,13 @@ class Mul(ScalarOp):
def c_code(self, node, name, inputs, outputs, sub):
(z,) = outputs
op = " * "
if node.outputs[0].type == bool:
op = " && "
if not inputs:
return z + " = 1;"
else:
return z + " = " + " * ".join(inputs) + ";"
return z + " = " + op.join(inputs) + ";"
def grad(self, inputs, gout):
(gz,) = gout
......@@ -1571,7 +1614,7 @@ class Sub(BinaryScalarOp):
second_part = -gz
return first_part, second_part
sub = Sub(upcast_out, name='sub')
sub = Sub(upcast_out_nobool, name='sub')
def int_or_true_div(x_discrete, y_discrete):
......@@ -1937,7 +1980,7 @@ class Pow(BinaryScalarOp):
raise theano.gof.utils.MethodNotDefined()
pow = Pow(upcast_out, name='pow')
pow = Pow(upcast_out_min8, name='pow')
class Clip(ScalarOp):
......@@ -2062,6 +2105,8 @@ class Cast(UnaryScalarOp):
def c_code(self, node, name, inputs, outputs, sub):
(x,) = inputs
(z,) = outputs
if node.outputs[0].type == bool:
return "%s = (%s) ? 1 : 0;" % (z, x)
return "%s = (%s)%s;" % (z, node.outputs[0].type.dtype_specs()[1], x)
def grad(self, inputs, gout):
......@@ -2075,10 +2120,11 @@ class Cast(UnaryScalarOp):
def c_code_cache_version(self):
s = super(Cast, self).c_code_cache_version()
if s:
return (3,) + s
return (4,) + s
else:
return s
convert_to_bool = Cast(bool, name='convert_to_bool')
convert_to_int8 = Cast(int8, name='convert_to_int8')
convert_to_int16 = Cast(int16, name='convert_to_int16')
convert_to_int32 = Cast(int32, name='convert_to_int32')
......@@ -2094,6 +2140,7 @@ convert_to_complex64 = Cast(complex64, name='convert_to_complex64')
convert_to_complex128 = Cast(complex128, name='convert_to_complex128')
_cast_mapping = {
'bool': convert_to_bool,
'int8': convert_to_int8,
'int16': convert_to_int16,
'int32': convert_to_int32,
......@@ -2173,6 +2220,12 @@ abs_ = Abs(same_out)
class Sgn(UnaryScalarOp):
nfunc_spec = ('sign', 1, 1)
@staticmethod
def output_types_preference(x):
if x == bool:
raise TypeError(x)
return same_out_nocomplex(x)
def impl(self, x):
# casting to output type is handled by filter
return numpy.sign(x)
......@@ -2205,7 +2258,7 @@ class Sgn(UnaryScalarOp):
return (4,) + s
else: # if parent is unversioned, we are too
return s
sgn = Sgn(same_out_nocomplex, name='sgn')
sgn = Sgn(name='sgn')
class Ceil(UnaryScalarOp):
......@@ -2228,7 +2281,7 @@ class Ceil(UnaryScalarOp):
(x,) = inputs
(z,) = outputs
return "%(z)s = ceil(%(x)s);" % locals()
ceil = Ceil(same_out_nocomplex, name='ceil')
ceil = Ceil(upgrade_to_float_no_complex, name='ceil')
class Floor(UnaryScalarOp):
......@@ -2251,7 +2304,7 @@ class Floor(UnaryScalarOp):
(x,) = inputs
(z,) = outputs
return "%(z)s = floor(%(x)s);" % locals()
floor = Floor(same_out_nocomplex, name='floor')
floor = Floor(upgrade_to_float_no_complex, name='floor')
class Trunc(UnaryScalarOp):
......@@ -2269,7 +2322,7 @@ class Trunc(UnaryScalarOp):
(x,) = inputs
(z,) = outputs
return "%(z)s = %(x)s >= 0? floor(%(x)s): -floor(-%(x)s);" % locals()
trunc = Trunc(same_out_nocomplex, name='trunc')
trunc = Trunc(upgrade_to_float_no_complex, name='trunc')
class RoundHalfToEven(UnaryScalarOp):
......@@ -2409,13 +2462,6 @@ class Neg(UnaryScalarOp):
nfunc_spec = ('negative', 1, 1)
def impl(self, x):
# We have to make sure x is not a numpy.bool_, because
# `-numpy.bool_(True)` is `False` (we want 0), and
# `-numpy.bool_(False)` is `True` (we want 1).
# This happens for Composite, as the intermediate results are not
# casted in the dtype of the intermediate variable in general.
if isinstance(x, numpy.bool_):
x = numpy.int8(x)
return -x
def grad(self, inputs, gout):
......@@ -2433,7 +2479,7 @@ class Neg(UnaryScalarOp):
(x,) = inputs
(z,) = outputs
return "%(z)s = -%(x)s;" % locals()
neg = Neg(same_out, name='neg')
neg = Neg(same_out_nobool, name='neg')
pprint.assign(add, printing.OperatorPrinter('+', -2, 'either'))
pprint.assign(mul, printing.OperatorPrinter('*', -1, 'either'))
......@@ -3423,7 +3469,7 @@ class Conj(UnaryScalarOp):
def impl(self, x):
return numpy.conj(x)
conj = Conj(same_out, name='conj')
conj = Conj(same_out_min8, name='conj')
class ComplexFromPolar(BinaryScalarOp):
......
theano/scalar/tests/test_basic.py
浏览文件 @ bf9413c8
......@@ -155,17 +155,6 @@ class test_composite(unittest.TestCase):
si3 = theano.scalar.float32()
sop.make_node(si0 * si3, si1, si2)
def test_composite_neg_bool(self):
# Check that taking the negation of a Boolean intermediate value
# works correctly with Python code. It used to be an issue because
# `-numpy.bool_(True)` is False and `-numpy.bool_(False)` is True.
x = floats('x')
y = - (x > 0)
z = Composite([x], [y]).make_node(x).outputs[0]
f = theano.function([x], z, mode=theano.Mode(linker='py'))
for inp, out in zip([-1, 0, 1], [0, 0, -1]):
self.assertTrue(f(inp) == out)
class test_logical(unittest.TestCase):
def test_gt(self):
......
theano/tensor/basic.py
浏览文件 @ bf9413c8
......@@ -52,6 +52,7 @@ python_all = all
complex_dtypes = list(map(str, scal.complex_types))
continuous_dtypes = list(map(str, scal.continuous_types))
float_dtypes = list(map(str, scal.float_types))
integer_dtypes = list(map(str, scal.integer_types))
discrete_dtypes = list(map(str, scal.discrete_types))
all_dtypes = list(map(str, scal.all_types))
int_dtypes = list(map(str, scal.int_types))
......@@ -302,7 +303,7 @@ class NumpyAutocaster(object):
# returns either an exact x_==x, or the last cast x_
return x_
autocast_int = NumpyAutocaster(('int16', 'int32', 'int64'))
autocast_int = NumpyAutocaster(('int8', 'int16', 'int32', 'int64'))
autocast_float = NumpyAutocaster(('float16', 'float32', 'float64'))
......@@ -379,16 +380,10 @@ def constant_or_value(x, rtype, name=None, ndim=None, dtype=None):
x_ = autocast_float(x)
elif isinstance(x, numpy.ndarray):
x_ = x
# Currently we do not have a bool dtype in Theano.
# So we upcast it to uint8 to avoid breaking our interface for
# constant.
if x.dtype == 'bool':
x_ = numpy.asarray(x_, dtype='uint8')
else:
# Here x is probably a list or a tuple. If it contains a long,
# we will behave like the current NumPy version: 1.7 and below,
# it will only work if the long fits in int64. For NumPy 1.7.1+,
# it will work if the long fits in int64 or uint64.
# Here x is probably a list or a tuple. If it contains a
# long, we will behave like the current NumPy version: it
# will work if the long fits in int64 or uint64.
x_ = numpy.asarray(x)
if x_.size == 0 and not hasattr(x, 'dtype'):
x_ = numpy.asarray(x, dtype=config.floatX)
......@@ -521,11 +516,6 @@ def _allclose(a, b, rtol=None, atol=None):
if atol is not None:
atol_ = atol
# Work around bug in Numpy, see
# http://projects.scipy.org/numpy/ticket/1684
if str(b.dtype) in int_dtypes and (numpy.absolute(b) < 0).any():
b = theano._asarray(b, dtype='float64')
return numpy.allclose(a, b, atol=atol_, rtol=rtol_)
......@@ -1247,6 +1237,10 @@ def _conversion(real_value, name):
# what types you are casting to what. That logic is implemented by the
# `cast()` function below.
_convert_to_bool = _conversion(
elemwise.Elemwise(scal.convert_to_bool), 'bool')
"""Cast to boolean"""
_convert_to_int8 = _conversion(
elemwise.Elemwise(scal.convert_to_int8), 'int8')
"""Cast to 8-bit integer"""
......@@ -1300,6 +1294,7 @@ _convert_to_complex128 = _conversion(
"""Cast to double-precision complex"""
_cast_mapping = {
'bool': _convert_to_bool,
'int8': _convert_to_int8,
'int16': _convert_to_int16,
'int32': _convert_to_int32,
......@@ -3191,6 +3186,10 @@ def mean(input, axis=None, dtype=None, op=False, keepdims=False,
for i in axis:
s = true_div(s, shp[i])
# This can happen when axis is an empty list/tuple
if s.dtype != shp.dtype and s.dtype in discrete_dtypes:
s = cast(s, shp.dtype)
if dtype == 'float16' or (dtype is None and input.dtype == 'float16'):
s = cast(s, 'float16')
s.name = 'mean'
......
theano/tensor/blas.py
浏览文件 @ bf9413c8
......@@ -1189,7 +1189,7 @@ def _gemm_canonicalize(r, scale, rval, maxclients):
def scaled(thing):
if scale == 1:
return thing
if scale == -1:
if scale == -1 and thing.type.dtype != 'bool':
return -thing
else:
return scale * thing
......
theano/tensor/elemwise.py
浏览文件 @ bf9413c8
......@@ -20,12 +20,6 @@ from theano.tensor import elemwise_cgen as cgen
config = theano.config
# We cannot import discrete_dtypes or float_dtypes from tensor.basic yet,
# so we redefine them here
discrete_dtypes = list(map(str, scalar.discrete_types))
float_dtypes = list(map(str, scalar.float_types))
int_dtypes = list(map(str, scalar.int_types))
# tensor depends on elemwise to provide definitions for several ops
# but elemwise needs to make TensorType instances, so we have these as
......@@ -818,9 +812,9 @@ second dimension
# NumPy 1.10.1 raise an error when giving the signature
# when the input is complex. So add it only when inputs is int.
out_dtype = node.outputs[0].dtype
if (out_dtype in float_dtypes and
if (out_dtype in theano.tensor.float_dtypes and
isinstance(self.nfunc, numpy.ufunc) and
node.inputs[0].dtype in discrete_dtypes):
node.inputs[0].dtype in theano.tensor.discrete_dtypes):
char = numpy.sctype2char(out_dtype)
sig = char * node.nin + '->' + char * node.nout
node.tag.sig = sig
......@@ -1076,7 +1070,7 @@ second dimension
# We loop over the "aliased" outputs, i.e., those that are
# inplace (overwrite the contents of one of the inputs) and
# make the output pointers point to theur corresponding input
# make the output pointers point to their corresponding input
# pointers.
for output, oname in izip(aliased_outputs, aliased_onames):
olv_index = inputs.index(dmap[output][0])
......@@ -1641,10 +1635,10 @@ for(int i=0;i<PyArray_NDIM(%(iname)s);i++){
task1_code = self.scalar_op.c_code(
Apply(self.scalar_op,
[get_scalar_type(dtype=input.type.dtype).make_variable()
for input in (node.inputs * 2)],
[get_scalar_type(dtype=output.type.dtype).make_variable()
for input in node.outputs]),
[get_scalar_type(dtype=iv.type.dtype).make_variable()
for iv in (node.inputs * 2)],
[get_scalar_type(dtype=ov.type.dtype).make_variable()
for ov in node.outputs]),
None,
["%s_i" % aname, "%s_i" % inames[0]],
["%s_i" % aname],
......@@ -1708,18 +1702,16 @@ for(int i=0;i<PyArray_NDIM(%(iname)s);i++){
class All(CAReduce):
""" Applies `bitwise and` to all the values of a tensor along the
""" Applies `logical and` to all the values of a tensor along the
specified axis(es).
Equivalent to `CAReduce(scalar.and\_, axis=axis)`.
"""
def __init__(self, axis=None):
CAReduce.__init__(self, scalar.and_, axis)
def _output_dtype(self, idtype):
return "int8"
return "bool"
def __str__(self):
if self.axis is None:
......@@ -1729,7 +1721,7 @@ class All(CAReduce):
def make_node(self, input):
input = as_tensor_variable(input)
if input.dtype not in ["int8", "uint8"]:
if input.dtype != "bool":
input = theano.tensor.neq(input, 0)
ret = super(All, self).make_node(input)
return ret
......@@ -1743,15 +1735,13 @@ class Any(CAReduce):
""" Applies `bitwise or` to all the values of a tensor along the
specified axis(es).
Equivalent to `CAReduce(scalar.or\_, axis=axis)`.
"""
def __init__(self, axis=None):
CAReduce.__init__(self, scalar.or_, axis)
def _output_dtype(self, idtype):
return "int8"
return "bool"
def __str__(self):
if self.axis is None:
......@@ -1761,7 +1751,7 @@ class Any(CAReduce):
def make_node(self, input):
input = as_tensor_variable(input)
if input.dtype not in ["int8", "uint8"]:
if input.dtype != "bool":
input = theano.tensor.neq(input, 0)
ret = super(Any, self).make_node(input)
return ret
......@@ -1863,6 +1853,7 @@ class CAReduceDtype(CAReduce):
if dtype is None:
# If input has a discrete dtype, upcast it to 64
return dict(
bool='int64',
int8='int64',
int16='int64',
int32='int64',
......@@ -1878,6 +1869,7 @@ class CAReduceDtype(CAReduce):
acc_dtype = self.acc_dtype
if acc_dtype is None:
return dict(
bool='int64',
int8='int64',
int16='int64',
int32='int64',
......@@ -1990,7 +1982,7 @@ class Sum(CAReduceDtype):
out = self(*inp)
if out.dtype.find('int') != -1:
if out.dtype not in theano.tensor.continuous_dtypes:
return [x.zeros_like(dtype=theano.config.floatX)]
gz, = grads
......@@ -2101,8 +2093,8 @@ class Prod(CAReduceDtype):
out = self(*inp)
if (out.dtype in discrete_dtypes or
self.acc_dtype in discrete_dtypes):
if (out.dtype in theano.tensor.discrete_dtypes or
self.acc_dtype in theano.tensor.discrete_dtypes):
# There is an int conversion in the way
return [prod_in.zeros_like(dtype=theano.config.floatX)]
......
theano/tensor/extra_ops.py
浏览文件 @ bf9413c8
from __future__ import absolute_import, print_function, division
import numpy as np
import numpy
import warnings
from six.moves import xrange
import theano
......@@ -561,103 +560,6 @@ def diff(x, n=1, axis=-1):
return DiffOp(n=n, axis=axis)(x)
class BinCountOp(theano.Op):
"""
.. note:: Deprecated
Use bincount() instead.
See function bincount for docstring.
"""
compatible_type = ('int8', 'int16', 'int32', 'int64',
'uint8', 'uint16', 'uint32', 'uint64')
"""Tuple of all compatible dtype for the parameter of this op."""
__props__ = ("minlength",)
def __init__(self, minlength=None):
self.minlength = minlength
if minlength is not None:
numpy_ver = [int(n) for n in numpy.__version__.split('.')[:2]]
if not bool(numpy_ver >= [1, 6]):
raise NotImplementedError(
"BinCountOp with minlength attribute"
" requires NumPy 1.6 or higher.")
def make_node(self, x, weights):
warnings.warn((
"Tile op is deprecated, use tile function instead."),
stacklevel=3)
x = basic.as_tensor_variable(x)
if x.dtype not in BinCountOp.compatible_type:
raise TypeError("Inputs dtype must be an integer.")
# Some dtypes are not supported by numpy's implementation of bincount.
# Until another one is available, we should fail at graph construction
# time, not wait for execution.
int_bitwidth = theano.configdefaults.python_int_bitwidth()
if int_bitwidth == 64:
numpy_unsupported_dtypes = ('uint64',)
if int_bitwidth == 32:
numpy_unsupported_dtypes = ('uint32', 'int64', 'uint64')
intp_bitwidth = theano.configdefaults.local_bitwidth()
if intp_bitwidth == 32:
out_type = basic.ivector()
elif intp_bitwidth == 64:
out_type = basic.lvector()
if x.dtype in numpy_unsupported_dtypes:
raise TypeError(
("Input dtypes %s are not supported by numpy.bincount, "
% numpy_unsupported_dtypes), x.dtype)
if x.ndim != 1:
raise TypeError("Inputs must be of dimension 1.")
if weights is None:
weights = theano.gof.Constant(theano.gof.Generic(), None)
else:
weights = basic.as_tensor_variable(weights)
out_type = basic.dvector()
if weights.ndim != 1:
raise TypeError("Weights cannot have a number of"
"dimension different of 1.")
return theano.Apply(self, [x, weights], [out_type])
def perform(self, node, inputs, output_storage):
x = inputs[0]
weights = inputs[1]
z = output_storage[0]
if weights is not None and weights.shape != x.shape:
raise TypeError("All inputs must have the same shape.")
# Needed for numpy 1.4.1 compatibility
if self.minlength:
out = np.bincount(x, weights=weights, minlength=self.minlength)
else:
out = np.bincount(x, weights=weights)
z[0] = theano._asarray(out, dtype=node.outputs[0].dtype)
def grad(self, inputs, outputs_gradients):
output = self(*inputs)
if output.dtype.find('int') != -1:
return [inp.zeros_like().astype(theano.config.floatX)
for inp in inputs]
raise NotImplementedError()
def infer_shape(self, node, ins_shapes):
x = node.inputs[0]
m = basic.max(x) + 1
if self.minlength is not None:
m = basic.maximum(m, self.minlength)
return [[m]]
def bincount(x, weights=None, minlength=None, assert_nonneg=False):
"""Count number of occurrences of each value in array of ints.
......@@ -773,7 +675,7 @@ class RepeatOp(theano.Op):
x = basic.as_tensor_variable(x)
repeats = basic.as_tensor_variable(repeats)
if repeats.dtype not in tensor.discrete_dtypes:
if repeats.dtype not in tensor.integer_dtypes:
raise TypeError("repeats.dtype must be an integer.")
# Some dtypes are not supported by numpy's implementation of repeat.
......
theano/tensor/nnet/sigm.py
浏览文件 @ bf9413c8
......@@ -75,7 +75,7 @@ class ScalarSigmoid(scalar.UnaryScalarOp):
# float16 limits: -11.0, 7.0f
# We use the float32 limits for float16 for now as the
# computation will happend in float32 anyway.
# computation will happen in float32 anyway.
if (node.inputs[0].type == scalar.float32 or
node.inputs[0].type == scalar.float16):
return """%(z)s = %(x)s < -88.0f ? 0.0 : %(x)s > 15.0f ? 1.0f : 1.0f /(1.0f + exp(-%(x)s));""" % locals()
......
theano/tensor/opt.py
浏览文件 @ bf9413c8
......@@ -5906,7 +5906,10 @@ def local_mul_specialize(node):
if new_inputs:
if len(new_inputs) == 1:
if neg:
rval = -new_inputs[0]
if new_inputs[0].dtype in (T.uint_dtypes + ['bool']):
return
else:
rval = -new_inputs[0]
else:
rval = new_inputs[0]
else:
......
theano/tensor/tests/test_basic.py
浏览文件 @ bf9413c8
......@@ -1206,9 +1206,7 @@ IntDivInplaceTester = makeBroadcastTester(
CeilTester = makeBroadcastTester(op=tensor.ceil,
expected=lambda a: numpy.asarray(
numpy.ceil(a),
a.dtype),
expected=upcast_float16_ufunc(numpy.ceil),
good=_good_broadcast_unary_normal_no_complex,
grad=copymod(_grad_broadcast_unary_normal,
without=['corner_case'],
......@@ -1217,7 +1215,7 @@ CeilTester = makeBroadcastTester(op=tensor.ceil,
dtype=floatX)]))
CeilInplaceTester = makeBroadcastTester(op=inplace.ceil_inplace,
expected=lambda a: numpy.asarray(numpy.ceil(a), a.dtype),
expected=upcast_float16_ufunc(numpy.ceil),
good=_good_broadcast_unary_normal_no_complex,
# corner cases includes a lot of integers: points where Ceil is not
# continuous (not differentiable)
......@@ -1229,7 +1227,7 @@ CeilInplaceTester = makeBroadcastTester(op=inplace.ceil_inplace,
inplace=True)
FloorTester = makeBroadcastTester(op=tensor.floor,
expected=lambda a: numpy.asarray(numpy.floor(a), a.dtype),
expected=upcast_float16_ufunc(numpy.floor),
good=_good_broadcast_unary_normal_no_complex,
# XXX: why does grad of floor not give huge values at
# the integer points in the 'corner_case' in
......@@ -1238,20 +1236,20 @@ FloorTester = makeBroadcastTester(op=tensor.floor,
grad=_grad_broadcast_unary_normal)
FloorInplaceTester = makeBroadcastTester(op=inplace.floor_inplace,
expected=lambda a: numpy.asarray(numpy.floor(a), a.dtype),
expected=upcast_float16_ufunc(numpy.floor),
good=_good_broadcast_unary_normal_no_complex,
grad=_grad_broadcast_unary_normal,
inplace=True)
TruncInplaceTester = makeBroadcastTester(
op=inplace.trunc_inplace,
expected=lambda a: numpy.asarray(numpy.trunc(a), a.dtype),
expected=upcast_float16_ufunc(numpy.trunc),
good=_good_broadcast_unary_normal_no_complex,
inplace=True)
TruncTester = makeBroadcastTester(
op=tensor.trunc,
expected=lambda a: numpy.asarray(numpy.trunc(a), a.dtype),
expected=upcast_float16_ufunc(numpy.trunc),
good=_good_broadcast_unary_normal_no_complex)
RoundHalfToEvenTester = makeBroadcastTester(
......@@ -5005,7 +5003,7 @@ class T_scalarfromtensor(unittest.TestCase):
self.assertTrue(v == 56, v)
if config.cast_policy == 'custom':
self.assertTrue(isinstance(v, numpy.int16))
self.assertTrue(isinstance(v, numpy.int8))
elif config.cast_policy in ('numpy', 'numpy+floatX'):
self.assertTrue(isinstance(
v, getattr(numpy, str(numpy.asarray(56).dtype))))
......@@ -7120,7 +7118,7 @@ class T_as_tensor_variable(unittest.TestCase):
def test_ndarray_bool(self):
ten = as_tensor_variable(numpy.array([True, False, False, True, True]))
assert ten.type.dtype == 'uint8'
assert ten.type.dtype == 'bool'
def test_memmap(self):
inp = numpy.random.rand(4, 3)
......@@ -8192,25 +8190,3 @@ def test_symbolic_slice():
a, b = x.shape[:2]
output = a.eval({x: numpy.zeros((5, 4, 3, 2), dtype=theano.config.floatX)})
assert output == numpy.array(5)
def test_composite_neg_bool():
# Check that taking the negation of a Boolean intermediate value
# works correctly with Python code. It used to be an issue because
# `-numpy.bool_(True)` is False and `-numpy.bool_(False)` is True.
x = theano.tensor.vector()
f = theano.function([x], - (x > 0), mode=theano.Mode(linker='py'))
utt.assert_allclose(f([-1, 0, 1]), [0, 0, -1])
"""
if __name__ == '__main__':
if 0:
unittest.main()
else:
testcase = FloorInplaceTester
suite = unittest.TestLoader()
suite = suite.loadTestsFromTestCase(testcase)
unittest.TextTestRunner(verbosity=2).run(suite)
"""
theano/tensor/tests/test_elemwise.py
浏览文件 @ bf9413c8
......@@ -113,42 +113,42 @@ class test_reduce_axes(unittest.TestCase):
[numpy.array(0), numpy.array(1)]]
for a in axes:
x = tensor.matrix()
m = x.sum(a)
x.sum(a)
def test_mean_axes(self):
axes = [None, 0, 1, [0, 1], numpy.array(1),
[numpy.array(0), numpy.array(1)]]
for a in axes:
x = tensor.matrix()
m = x.mean(a)
x.mean(a)
def test_max_axes(self):
axes = [None, 0, 1, [0, 1], numpy.array(1),
[numpy.array(0), numpy.array(1)]]
for a in axes:
x = tensor.matrix()
m = x.max(a)
x.max(a)
def test_min_axes(self):
axes = [None, 0, 1, [0, 1], numpy.array(1),
[numpy.array(0), numpy.array(1)]]
for a in axes:
x = tensor.matrix()
m = x.min(a)
x.min(a)
def test_argmax_axes(self):
axes = [None, 0, 1, [0, 1], numpy.array(1),
[numpy.array(0), numpy.array(1)]]
for a in axes:
x = tensor.matrix()
m = x.argmax(a)
x.argmax(a)
def test_var_axes(self):
axes = [None, 0, 1, [0, 1], numpy.array(1),
[numpy.array(0), numpy.array(1)]]
for a in axes:
x = tensor.matrix()
m = x.var(a)
x.var(a)
class test_Broadcast(unittest.TestCase):
......@@ -159,7 +159,7 @@ class test_Broadcast(unittest.TestCase):
ctype = TensorType
cop = Elemwise
openmp_minsize = 2*config.openmp_elemwise_minsize
openmp_minsize = 2 * config.openmp_elemwise_minsize
openmp_minsize_sqrt = int(math.ceil(math.sqrt(openmp_minsize)))
# The order is important if you change them.
......@@ -346,8 +346,7 @@ class test_CAReduce(unittest_tools.InferShapeTester):
((5, 0), (1, )),
((5, 0), ()),
((), None),
((), ())
]
((), ())]
type = TensorType
def with_linker(self, linker, scalar_op=scalar.add, dtype="floatX",
......@@ -371,7 +370,7 @@ class test_CAReduce(unittest_tools.InferShapeTester):
f = copy(linker).accept(FunctionGraph([x], [e])).make_function()
xv = numpy.asarray(numpy.random.rand(*xsh))
if not "int" in dtype:
if "int" not in dtype:
xv = numpy.asarray(xv, dtype=dtype)
else:
xv = numpy.asarray(xv < 0.5, dtype=dtype)
......@@ -452,10 +451,6 @@ class test_CAReduce(unittest_tools.InferShapeTester):
else:
self.fail()
else:
# numpy.{all,any} return bool type,
# but theano ops return an int8 array instead
if scalar_op in [scalar.and_, scalar.or_]:
zv = numpy.asarray(zv, dtype='int8')
if test_nan:
try:
self.assertTrue(
......@@ -614,18 +609,20 @@ class test_Prod(unittest.TestCase):
x = theano.tensor.dmatrix()
# sanity check
x2 = theano.tensor.dmatrix()
p = Prod(axis=1)(x)
p2 = Prod(axis=1)(x2)
fn = theano.function([x, x2], [p - p2], mode=self.mode)
# print "hand computed diff for each row"
x2_val = numpy.asarray([[1., 2., 3.003], [0.003, 5., 6], [
0., 0., 9.01]])
# print fn(x_val, x2_val)
fn2 = theano.function([x], [theano.tensor.grad(p.sum(), x)],
mode=self.mode)
# print "real grad"
# print fn2(x_val)
# Uncomment this for debugging if needed
# x2 = theano.tensor.dmatrix()
# p2 = Prod(axis=1)(x2)
# fn = theano.function([x, x2], [p - p2], mode=self.mode)
# print("hand computed diff for each row")
# x2_val = numpy.asarray([[1., 2., 3.003], [0.003, 5., 6], [
# 0., 0., 9.01]])
# print(fn(x_val, x2_val))
# fn2 = theano.function([x], [theano.tensor.grad(p.sum(), x)],
# mode=self.mode)
# print("real grad")
# print(fn2(x_val))
fn3 = theano.function([x], [p], mode=self.mode)
assert numpy.allclose(fn3(x_val), [6., 0., 0.])
......@@ -637,14 +634,14 @@ class test_Prod(unittest.TestCase):
# def fn5(x5):
# return theano.tensor.sqr(Prod(axis=1)(x5))
#x4 = theano.tensor.dmatrix()
#p4 = theano.tensor.sqr(Prod(axis=1)(x4))
#fn4 = theano.function([x4], p4)
# print "with sqr"
# print fn4(x_val)
# print fn4(x2_val)
# x4 = theano.tensor.dmatrix()
# p4 = theano.tensor.sqr(Prod(axis=1)(x4))
# fn4 = theano.function([x4], p4)
# print("with sqr")
# print(fn4(x_val))
# print(fn4(x2_val))
#unittest_tools.verify_grad(fn5, [x_val])
# unittest_tools.verify_grad(fn5, [x_val])
@attr('slow')
def test_prod_no_zeros_in_input(self):
......@@ -695,30 +692,33 @@ class test_Prod(unittest.TestCase):
x = theano.tensor.dmatrix()
pwz_a1 = ProdWithoutZeros(axis=0)(x)
pwz_grad = theano.grad(theano.tensor.sum(pwz_a1), x)
fn_a1 = theano.function([x], pwz_grad, mode=self.mode)
theano.function([x], pwz_grad, mode=self.mode)
@attr('slow')
def test_other_grad_tests(self):
x = theano.tensor.dmatrix()
x_val1 = numpy.array([[1, 2, 3], [0, 5, 6], [0, 0, 9]],
dtype='float32')
dtype='float32')
x_val2 = numpy.array([[1, 2, 0], [0, 5, 6], [7, 8, 9], [9, 10, 0]],
dtype='float32')
dtype='float32')
rng = rng = numpy.random.RandomState(43)
p = Prod(axis=1)
grad_p = theano.tensor.grad(p(x).sum(), x)
grad_fn = theano.function([x], grad_p, mode=self.mode)
assert numpy.allclose(grad_fn(x_val1), [[6., 3., 2.], [30., 0.,
0.], [0., 0., 0.]])
assert numpy.allclose(grad_fn(x_val2), [[0., 0., 2.], [30.,
0., 0.], [72., 63., 56.], [0., 0., 90.]])
assert numpy.allclose(
grad_fn(x_val1),
[[6., 3., 2.], [30., 0., 0.], [0., 0., 0.]])
assert numpy.allclose(
grad_fn(x_val2),
[[0., 0., 2.], [30., 0., 0.], [72., 63., 56.], [0., 0., 90.]])
p_axis0 = Prod(axis=0)
grad_p_axis0 = theano.tensor.grad(p_axis0(x).sum(), x)
grad_fn_axis0 = theano.function([x], grad_p_axis0, mode=self.mode)
assert numpy.allclose(grad_fn_axis0(x_val2), [[0., 400.,
0.], [63., 160., 0.], [0., 100., 0.], [0., 80., 0.]])
assert numpy.allclose(
grad_fn_axis0(x_val2),
[[0., 400., 0.], [63., 160., 0.], [0., 100., 0.], [0., 80., 0.]])
tensor.verify_grad(p, [x_val1], rng=rng, mode=self.mode)
......@@ -768,7 +768,7 @@ class test_IsInf_IsNan(unittest.TestCase):
numpy_isfunc = getattr(numpy, isfunc)
for x in self.test_vals:
if ((x.ndim == 0 and input is not self.scalar) or
(x.ndim == 1 and input is not self.vector)):
(x.ndim == 1 and input is not self.vector)):
# We only test with the appropriate input type.
continue
t_out = theano_isfunc(x)
......@@ -788,12 +788,10 @@ class T_reduce_dtype(unittest.TestCase):
op = CAReduce
axes = [None, 0, 1, [], [0], [1], [0, 1]]
methods = ['sum', 'prod']
dtypes = imap(str, theano.scalar.all_types)
dtypes = list(imap(str, theano.scalar.all_types))
# Test the default dtype of a method().
def test_reduce_default_dtype(self):
"""
Test the default dtype of a method().
"""
# We try multiple axis combinations even though axis should not matter.
for method in self.methods:
for idx, dtype in enumerate(self.dtypes):
......@@ -801,6 +799,7 @@ class T_reduce_dtype(unittest.TestCase):
x = tensor.matrix(dtype=dtype)
s = getattr(x, method)(axis=axis)
assert s.dtype == dict(
bool='int64',
int8='int64',
int16='int64',
int32='int64',
......@@ -818,6 +817,7 @@ class T_reduce_dtype(unittest.TestCase):
def test_reduce_default_acc_dtype(self):
# Test the default acc_dtype of a reduce().
# We try multiple axis combinations even though axis should not matter.
for method in self.methods:
for idx, dtype in enumerate(self.dtypes):
......@@ -825,12 +825,14 @@ class T_reduce_dtype(unittest.TestCase):
x = tensor.matrix(dtype=dtype)
s = getattr(x, method)(axis=axis)
assert s.owner.op.acc_dtype == dict(
bool='int64',
int8='int64',
int16='int64',
int32='int64',
uint8='uint64',
uint16='uint64',
uint32='uint64',
float16='float32',
float32='float64',
complex64='complex128',
).get(dtype, dtype)
......@@ -844,20 +846,18 @@ class T_reduce_dtype(unittest.TestCase):
@attr('slow')
def test_reduce_custom_dtype(self):
"""
Test the ability to provide your own output dtype for a reduce.
"""
# Test the ability to provide your own output dtype for a reduce.
# We try multiple axis combinations even though axis should not matter.
idx = 0
for method in self.methods:
for input_dtype in self.dtypes:
x = tensor.matrix(dtype=input_dtype)
for output_dtype in self.dtypes:
# If the output is a complex, the gradient of the reduce will
# cast the complex to the input dtype. We can't call the normal
# cast on a complex to a not complex as this is ambiguous.
if (not input_dtype.startswith('complex') and
output_dtype.startswith('complex')):
# Only tests case where both input and output are complex.
icomplex = input_dtype.startswith('complex')
ocomplex = output_dtype.startswith('complex')
if icomplex != ocomplex:
continue
axis = self.axes[idx % len(self.axes)]
......@@ -866,8 +866,8 @@ class T_reduce_dtype(unittest.TestCase):
f = theano.function([x], var, mode=self.mode)
topo = f.maker.fgraph.toposort()
assert [n for n in topo if isinstance(n.op, self.op)], (topo,
dtype)
assert [n for n in topo if isinstance(n.op, self.op)], \
(topo, output_dtype)
data = numpy.random.rand(3, 4) * 10
data = data.astype(input_dtype)
f(data)
......@@ -879,30 +879,28 @@ class T_reduce_dtype(unittest.TestCase):
idx += 1
def test_reduce_custom_acc_dtype(self):
"""
Test the ability to provide your own accumulator dtype for a reduce.
"""
# Test the ability to provide your own accumulator dtype for a reduce.
# We try multiple axis combinations even though axis should not matter.
idx = 0
for method in self.methods:
for input_dtype in self.dtypes:
x = tensor.matrix(dtype=input_dtype)
for acc_dtype in self.dtypes:
# If the accumulator is a complex, the gradient of the reduce will
# cast the complex to the input dtype. We can't call the normal
# cast on a complex to a not complex as this is ambiguous.
# If the accumulator is a complex, the gradient of the reduce will
# cast the complex to the input dtype. We can't call the normal
# cast on a complex to a not complex as this is ambiguous.
if (not input_dtype.startswith('complex') and
acc_dtype.startswith('complex')):
acc_dtype.startswith('complex')):
continue
axis = self.axes[idx % len(self.axes)]
# If output_dtype would force a downcast, we expect a TypeError
# We always allow int/uint inputs with float/complex outputs.
# If output_dtype would force a downcast, we expect a TypeError
# We always allow int/uint inputs with float/complex outputs.
upcasted_dtype = scalar.upcast(input_dtype, acc_dtype)
if (acc_dtype == upcasted_dtype or
(input_dtype in tensor.discrete_dtypes and
acc_dtype in tensor.continuous_dtypes)
):
acc_dtype in tensor.continuous_dtypes)):
var = getattr(x, method)(acc_dtype=acc_dtype,
axis=axis)
assert var.owner.op.acc_dtype == acc_dtype
......@@ -927,8 +925,7 @@ class T_reduce_dtype(unittest.TestCase):
s = getattr(x, method)()
f = theano.function([], s, mode=self.mode)
topo = f.maker.fgraph.toposort()
assert [n for n in topo if isinstance(n.op, self.op)], (topo,
dtype)
assert [n for n in topo if isinstance(n.op, self.op)], topo
s_val = f()
# Use extra precision in NumPy to compute the good answer.
ret = getattr(numpy.asarray([1e8, 1, -1e8], dtype='float64'),
......@@ -938,16 +935,15 @@ class T_reduce_dtype(unittest.TestCase):
class T_mean_dtype(unittest.TestCase):
def test_mean_default_dtype(self):
"""
Test the default dtype of a mean().
"""
# Test the default dtype of a mean().
# We try multiple axis combinations even though axis should not matter.
axes = [None, 0, 1, [], [0], [1], [0, 1]]
for idx, dtype in enumerate(imap(str, theano.scalar.all_types)):
axis = axes[idx % len(axes)]
x = tensor.matrix(dtype=dtype)
m = x.mean(axis=axis)
if dtype in tensor.discrete_dtypes and axis != []:
if dtype in tensor.discrete_dtypes:
assert m.dtype == 'float64'
else:
assert m.dtype == dtype, (m, m.dtype, dtype)
......@@ -958,9 +954,8 @@ class T_mean_dtype(unittest.TestCase):
@attr('slow')
def test_mean_custom_dtype(self):
"""
Test the ability to provide your own output dtype for a mean.
"""
# Test the ability to provide your own output dtype for a mean.
# We try multiple axis combinations even though axis should not matter.
axes = [None, 0, 1, [], [0], [1], [0, 1]]
idx = 0
......@@ -976,7 +971,7 @@ class T_mean_dtype(unittest.TestCase):
pass
else:
# Executed if no TypeError was raised
if sum_dtype in tensor.discrete_dtypes and axis != []:
if sum_dtype in tensor.discrete_dtypes:
assert mean_var.dtype == 'float64', (
(mean_var.dtype, sum_dtype))
else:
......@@ -984,7 +979,7 @@ class T_mean_dtype(unittest.TestCase):
(mean_var.dtype, sum_dtype))
if (('complex' in input_dtype or
'complex' in sum_dtype) and
input_dtype != sum_dtype):
input_dtype != sum_dtype):
continue
f = theano.function([x], mean_var)
data = numpy.random.rand(3, 4) * 10
......@@ -1017,15 +1012,15 @@ class T_mean_dtype(unittest.TestCase):
class T_prod_without_zeros_dtype(unittest.TestCase):
def test_prod_without_zeros_default_dtype(self):
"""
Test the default dtype of a ProdWithoutZeros().
"""
# Test the default dtype of a ProdWithoutZeros().
# We try multiple axis combinations even though axis should not matter.
axes = [None, 0, 1, [], [0], [1], [0, 1]]
for idx, dtype in enumerate(imap(str, theano.scalar.all_types)):
axis = axes[idx % len(axes)]
x = ProdWithoutZeros(axis=axis)(tensor.matrix(dtype=dtype))
assert x.dtype == dict(
bool='int64',
int8='int64',
int16='int64',
int32='int64',
......@@ -1035,9 +1030,8 @@ class T_prod_without_zeros_dtype(unittest.TestCase):
).get(dtype, dtype)
def test_prod_without_zeros_default_acc_dtype(self):
"""
Test the default dtype of a ProdWithoutZeros().
"""
# Test the default dtype of a ProdWithoutZeros().
# We try multiple axis combinations even though axis should not matter.
axes = [None, 0, 1, [], [0], [1], [0, 1]]
for idx, dtype in enumerate(imap(str, theano.scalar.all_types)):
......@@ -1045,16 +1039,17 @@ class T_prod_without_zeros_dtype(unittest.TestCase):
x = tensor.matrix(dtype=dtype)
p = ProdWithoutZeros(axis=axis)(x)
assert p.owner.op.acc_dtype == dict(
int8='int64',
int16='int64',
int32='int64',
uint8='uint64',
uint16='uint64',
uint32='uint64',
float16='float32',
float32='float64',
complex64='complex128'
).get(dtype, dtype)
bool='int64',
int8='int64',
int16='int64',
int32='int64',
uint8='uint64',
uint16='uint64',
uint32='uint64',
float16='float32',
float32='float64',
complex64='complex128'
).get(dtype, dtype)
if 'complex' in dtype:
continue
......@@ -1065,9 +1060,8 @@ class T_prod_without_zeros_dtype(unittest.TestCase):
@attr('slow')
def test_prod_without_zeros_custom_dtype(self):
"""
Test ability to provide your own output dtype for a ProdWithoutZeros().
"""
# Test ability to provide your own output dtype for a ProdWithoutZeros().
# We try multiple axis combinations even though axis should not matter.
axes = [None, 0, 1, [], [0], [1], [0, 1]]
idx = 0
......@@ -1076,11 +1070,11 @@ class T_prod_without_zeros_dtype(unittest.TestCase):
for output_dtype in imap(str, theano.scalar.all_types):
axis = axes[idx % len(axes)]
prod_woz_var = ProdWithoutZeros(
axis=axis, dtype=output_dtype)(x)
axis=axis, dtype=output_dtype)(x)
assert prod_woz_var.dtype == output_dtype
idx += 1
if ('complex' in output_dtype or
'complex' in input_dtype):
'complex' in input_dtype):
continue
f = theano.function([x], prod_woz_var)
data = numpy.random.rand(2, 3) * 3
......@@ -1089,9 +1083,8 @@ class T_prod_without_zeros_dtype(unittest.TestCase):
@attr('slow')
def test_prod_without_zeros_custom_acc_dtype(self):
"""
Test ability to provide your own acc_dtype for a ProdWithoutZeros().
"""
# Test ability to provide your own acc_dtype for a ProdWithoutZeros().
# We try multiple axis combinations even though axis should not matter.
axes = [None, 0, 1, [], [0], [1], [0, 1]]
idx = 0
......@@ -1104,23 +1097,23 @@ class T_prod_without_zeros_dtype(unittest.TestCase):
upcasted_dtype = scalar.upcast(input_dtype, acc_dtype)
if (acc_dtype == upcasted_dtype or
(input_dtype in tensor.discrete_dtypes and
acc_dtype in tensor.continuous_dtypes)
):
acc_dtype in tensor.continuous_dtypes)):
prod_woz_var = ProdWithoutZeros(
axis=axis, acc_dtype=acc_dtype)(x)
axis=axis, acc_dtype=acc_dtype)(x)
assert prod_woz_var.owner.op.acc_dtype == acc_dtype
if (acc_dtype.startswith('complex') and
input_dtype != acc_dtype):
input_dtype != acc_dtype):
continue
f = theano.function([x], prod_woz_var)
data = numpy.random.rand(2, 3) * 3
data = data.astype(input_dtype)
f(data)
else:
self.assertRaises(TypeError,
ProdWithoutZeros(axis=axis, acc_dtype=acc_dtype),
x)
self.assertRaises(
TypeError,
ProdWithoutZeros(axis=axis, acc_dtype=acc_dtype),
x)
idx += 1
......@@ -1160,24 +1153,26 @@ class TestElemwise(unittest_tools.InferShapeTester):
def test_infer_shape(self):
for s_left, s_right in [((5, 6), (5, 6)),
((5, 6), (5, 1)),
((5, 6), (1, 6)),
((5, 1), (5, 6)),
((1, 6), (5, 6)),
((2, 3, 4, 5), (2, 3, 4, 5)),
((2, 3, 4, 5), (2, 3, 1, 5)),
((2, 3, 4, 5), (1, 3, 4, 5)),
((2, 1, 4, 5), (2, 3, 4, 5)),
((2, 3, 4, 1), (2, 3, 4, 5))]:
for s_left, s_right in [
((5, 6), (5, 6)),
((5, 6), (5, 1)),
((5, 6), (1, 6)),
((5, 1), (5, 6)),
((1, 6), (5, 6)),
((2, 3, 4, 5), (2, 3, 4, 5)),
((2, 3, 4, 5), (2, 3, 1, 5)),
((2, 3, 4, 5), (1, 3, 4, 5)),
((2, 1, 4, 5), (2, 3, 4, 5)),
((2, 3, 4, 1), (2, 3, 4, 5))]:
dtype = theano.config.floatX
t_left = TensorType(dtype, [(entry == 1) for entry in s_left])()
t_right = TensorType(dtype, [(entry == 1) for entry in s_right])()
t_left_val = numpy.zeros(s_left, dtype=dtype)
t_right_val = numpy.zeros(s_right, dtype=dtype)
self._compile_and_check([t_left, t_right],
[Elemwise(scalar.add)(t_left, t_right)],
[t_left_val, t_right_val], Elemwise)
self._compile_and_check(
[t_left, t_right],
[Elemwise(scalar.add)(t_left, t_right)],
[t_left_val, t_right_val], Elemwise)
def test_input_dimensions_overflow(self):
# Elemwise.perform used to compute the product
......@@ -1186,7 +1181,7 @@ class TestElemwise(unittest_tools.InferShapeTester):
a, b, c, d, e, f = tensor.vectors('abcdef')
s = a + b + c + d + e + f
g = theano.function([a, b, c, d, e, f], s,
mode=theano.compile.Mode(linker='py'))
mode=theano.compile.Mode(linker='py'))
g(*[numpy.zeros(2 ** 11, config.floatX) for i in xrange(6)])
......@@ -1204,7 +1199,7 @@ def test_gt_grad():
input_ = T.vector(dtype=floatX)
random_values = numpy.random.RandomState(1234).uniform(
low=-1, high=1, size=(2, 2))
low=-1, high=1, size=(2, 2))
W_values = numpy.asarray(random_values, dtype=floatX)
W = theano.shared(value=W_values, name='weights')
correct_score = T.dot(input_, W)
......@@ -1235,7 +1230,7 @@ def test_clip_grad():
# use an x value less than y, an x value between y and z, and an x value
# greater than z
unittest_tools.verify_grad(func,
[numpy.asarray([-1., 0.5, 2.]), 0., 1.])
[numpy.asarray([-1., 0.5, 2.]), 0., 1.])
def test_clip_grad_int():
......
theano/tensor/tests/test_extra_ops.py
浏览文件 @ bf9413c8
......@@ -8,7 +8,7 @@ from theano.tests import unittest_tools as utt
from theano.tensor.extra_ops import (SearchsortedOp, searchsorted,
CumsumOp, cumsum, CumprodOp, cumprod,
CpuContiguous, cpu_contiguous, BinCountOp,
CpuContiguous, cpu_contiguous,
bincount, DiffOp, diff, squeeze, compress,
RepeatOp, repeat, Bartlett, bartlett,
FillDiagonal, fill_diagonal,
......@@ -18,9 +18,6 @@ from theano import tensor as T
from theano import config, tensor, function
from theano.tests.unittest_tools import attr
numpy_ver = [int(n) for n in numpy.__version__.split('.')[:2]]
numpy_16 = bool(numpy_ver >= [1, 6])
def test_cpu_contiguous():
a = T.fmatrix('a')
......@@ -218,12 +215,7 @@ class TestCumprodOp(utt.InferShapeTester):
utt.verify_grad(self.op_class(axis=axis), [a])
class TestBinCountOp(utt.InferShapeTester):
def setUp(self):
super(TestBinCountOp, self).setUp()
self.op_class = BinCountOp
self.op = BinCountOp()
class TestBinCount(utt.InferShapeTester):
def test_bincountFn(self):
w = T.vector('w')
......@@ -263,84 +255,6 @@ class TestBinCountOp(utt.InferShapeTester):
f5 = theano.function([x], bincount(x, assert_nonneg=True))
self.assertRaises(AssertionError, f5, a)
def test_bincountOp(self):
w = T.vector('w')
for dtype in ('int8', 'int16', 'int32', 'int64',
'uint8', 'uint16', 'uint32', 'uint64'):
# uint64 always fails
# int64 and uint32 also fail if python int are 32-bit
int_bitwidth = theano.configdefaults.python_int_bitwidth()
if int_bitwidth == 64:
numpy_unsupported_dtypes = ('uint64',)
if int_bitwidth == 32:
numpy_unsupported_dtypes = ('uint32', 'int64', 'uint64')
x = T.vector('x', dtype=dtype)
if dtype in numpy_unsupported_dtypes:
self.assertRaises(TypeError, BinCountOp(), x)
else:
a = np.random.randint(1, 51, size=(25)).astype(dtype)
weights = np.random.random((25,)).astype(config.floatX)
f1 = theano.function([x], BinCountOp()(x, weights=None))
f2 = theano.function([x, w], BinCountOp()(x, weights=w))
assert (np.bincount(a) == f1(a)).all()
assert np.allclose(np.bincount(a, weights=weights),
f2(a, weights))
if not numpy_16:
continue
f3 = theano.function([x], BinCountOp(minlength=23)(x, weights=None))
f4 = theano.function([x], BinCountOp(minlength=5)(x, weights=None))
assert (np.bincount(a, minlength=23) == f3(a)).all()
assert (np.bincount(a, minlength=5) == f4(a)).all()
@attr('slow')
def test_infer_shape(self):
for dtype in tensor.discrete_dtypes:
# uint64 always fails
# int64 and uint32 also fail if python int are 32-bit
int_bitwidth = theano.configdefaults.python_int_bitwidth()
if int_bitwidth == 64:
numpy_unsupported_dtypes = ('uint64',)
if int_bitwidth == 32:
numpy_unsupported_dtypes = ('uint32', 'int64', 'uint64')
x = T.vector('x', dtype=dtype)
if dtype in numpy_unsupported_dtypes:
self.assertRaises(TypeError, BinCountOp(), x)
else:
self._compile_and_check([x],
[BinCountOp()(x, None)],
[np.random.randint(
1, 51, size=(25,)).astype(dtype)],
self.op_class)
weights = np.random.random((25,)).astype(config.floatX)
self._compile_and_check([x],
[BinCountOp()(x, weights=weights)],
[np.random.randint(
1, 51, size=(25,)).astype(dtype)],
self.op_class)
if not numpy_16:
continue
self._compile_and_check([x],
[BinCountOp(minlength=60)(x, weights=weights)],
[np.random.randint(
1, 51, size=(25,)).astype(dtype)],
self.op_class)
self._compile_and_check([x],
[BinCountOp(minlength=5)(x, weights=weights)],
[np.random.randint(
1, 51, size=(25,)).astype(dtype)],
self.op_class)
class TestDiffOp(utt.InferShapeTester):
nb = 10 # Number of time iterating for n
......@@ -510,7 +424,7 @@ class TestRepeatOp(utt.InferShapeTester):
a = np.random.random((10, ) * ndim).astype(config.floatX)
for axis in self._possible_axis(ndim):
for dtype in tensor.discrete_dtypes:
for dtype in tensor.integer_dtypes:
r_var = T.scalar(dtype=dtype)
r = numpy.asarray(3, dtype=dtype)
if (dtype == 'uint64' or
......@@ -569,7 +483,7 @@ class TestRepeatOp(utt.InferShapeTester):
a = np.random.random(shp).astype(config.floatX)
for axis in self._possible_axis(ndim):
for dtype in tensor.discrete_dtypes:
for dtype in tensor.integer_dtypes:
r_var = T.scalar(dtype=dtype)
r = numpy.asarray(3, dtype=dtype)
if dtype in self.numpy_unsupported_dtypes:
......@@ -798,13 +712,11 @@ class test_Unique(utt.InferShapeTester):
self.ops = [Unique(),
Unique(True),
Unique(False, True),
Unique(True, True)]
if bool(numpy_ver >= [1, 9]):
self.ops.extend([
Unique(False, False, True),
Unique(True, False, True),
Unique(False, True, True),
Unique(True, True, True)])
Unique(True, True),
Unique(False, False, True),
Unique(True, False, True),
Unique(False, True, True),
Unique(True, True, True)]
def test_basic_vector(self):
"""
......@@ -816,13 +728,11 @@ class test_Unique(utt.InferShapeTester):
list_outs_expected = [[np.unique(inp)],
np.unique(inp, True),
np.unique(inp, False, True),
np.unique(inp, True, True)]
if bool(numpy_ver >= [1, 9]):
list_outs_expected.extend([
np.unique(inp, False, False, True),
np.unique(inp, True, False, True),
np.unique(inp, False, True, True),
np.unique(inp, True, True, True)])
np.unique(inp, True, True),
np.unique(inp, False, False, True),
np.unique(inp, True, False, True),
np.unique(inp, False, True, True),
np.unique(inp, True, True, True)]
for op, outs_expected in zip(self.ops, list_outs_expected):
f = theano.function(inputs=[x], outputs=op(x, return_list=True))
outs = f(inp)
......@@ -839,13 +749,11 @@ class test_Unique(utt.InferShapeTester):
list_outs_expected = [[np.unique(inp)],
np.unique(inp, True),
np.unique(inp, False, True),
np.unique(inp, True, True)]
if bool(numpy_ver >= [1, 9]):
list_outs_expected.extend([
np.unique(inp, False, False, True),
np.unique(inp, True, False, True),
np.unique(inp, False, True, True),
np.unique(inp, True, True, True)])
np.unique(inp, True, True),
np.unique(inp, False, False, True),
np.unique(inp, True, False, True),
np.unique(inp, False, True, True),
np.unique(inp, True, True, True)]
for op, outs_expected in zip(self.ops, list_outs_expected):
f = theano.function(inputs=[x], outputs=op(x, return_list=True))
outs = f(inp)
......
theano/tensor/type.py
浏览文件 @ bf9413c8
......@@ -255,6 +255,7 @@ class TensorType(Type):
'float16': (float, 'npy_float16', 'NPY_FLOAT16'),
'float32': (float, 'npy_float32', 'NPY_FLOAT32'),
'float64': (float, 'npy_float64', 'NPY_FLOAT64'),
'bool': (bool, 'npy_bool', 'NPY_BOOL'),
'uint8': (int, 'npy_uint8', 'NPY_UINT8'),
'int8': (int, 'npy_int8', 'NPY_INT8'),
'uint16': (int, 'npy_uint16', 'NPY_UINT16'),
......@@ -340,15 +341,9 @@ class TensorType(Type):
return False
if a.dtype != b.dtype:
return False
if 'int' in str(a.dtype):
if str(a.dtype) not in theano.tensor.continuous_dtypes:
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 = theano.tensor.basic._allclose(a, b, rtol=rtol, atol=atol)
if cmp:
# Numpy claims they are close, this is good enough for us.
......
theano/tensor/var.py
浏览文件 @ bf9413c8
......@@ -470,19 +470,17 @@ class _tensor_py_operators(object):
def check_bool(args_el):
try:
if isinstance(args_el, (numpy.bool_, bool)) or \
args_el.dtype == 'int8' or args_el.dtype == 'uint8':
raise TypeError(('TensorType does not support boolean '
'mask for indexing such as tensor[x==0]. '
'Instead you can use non_zeros() such as '
'tensor[(x == 0).nonzeros()]. '
'If you are indexing on purpose with an '
'int8, please cast it to int16.'))
if (isinstance(args_el, (numpy.bool_, bool)) or
args_el.dtype == 'bool'):
raise TypeError('TensorType does not support boolean '
'mask for indexing such as tensor[x==0]. '
'Instead you can use non_zeros() such as '
'tensor[(x == 0).nonzeros()]. ')
except AttributeError:
pass
if not isinstance(args_el, theano.tensor.Variable) and \
isinstance(args_el, collections.Iterable):
if (not isinstance(args_el, theano.tensor.Variable) and
isinstance(args_el, collections.Iterable)):
for el in args_el:
check_bool(el)
......
theano/tests/test_flake8.py
浏览文件 @ bf9413c8
......@@ -56,7 +56,6 @@ whitelist_flake8 = [
"tensor/tests/test_opt.py",
"tensor/tests/test_basic.py",
"tensor/tests/test_blas.py",
"tensor/tests/test_elemwise.py",
"tensor/tests/test_merge.py",
"tensor/tests/test_gc.py",
"tensor/tests/test_complex.py",
......
Markdown 格式
0% 或
您添加了 0 人 到此讨论。请谨慎行事。
请先完成此评论的编辑！
请 注册 或者 后发表评论