Skip to content

P
pytensor
提交 90c3833e authored 作者: nouiz's avatar nouiz
浏览文件
操作

Merge pull request #242 from nouiz/max_default

change the default of theano.{max,min,argmax,argmin,max_and_argmax} to t...
上级 c08a6f31 586c0e3c
隐藏空白字符变更
内嵌 并排
正在显示 包含 366 行增加 和 552 行删除
theano/compile/sharedvalue.py
浏览文件 @ 90c3833e
...@@ -18,13 +18,6 @@ from theano.gof import Container, Variable, generic ...@@ -18,13 +18,6 @@ from theano.gof import Container, Variable, generic
_logger = logging.getLogger('theano.compile.sharedvalue') _logger = logging.getLogger('theano.compile.sharedvalue')
AddConfigVar('shared.value_borrows',
("DEPRECATED. You should not use the 'value' property of shared"
" variables, but use the .get_value() and .set_value() methods."
" False: shared variables 'value' property is guaranteed to not"
" alias theano-managed memory. True: no guarantee, but faster."),
BoolParam(True),
in_c_key=False)
class SharedVariable(Variable): class SharedVariable(Variable):
""" """
...@@ -125,29 +118,6 @@ class SharedVariable(Variable): ...@@ -125,29 +118,6 @@ class SharedVariable(Variable):
cp.tag = copy.copy(self.tag) cp.tag = copy.copy(self.tag)
return cp return cp
def _value_get(self):
warnings.warn(("The .value property of shared variables is deprecated."
" You should use the .get_value() method instead."),
stacklevel=2)
return self.get_value(borrow=config.shared.value_borrows, return_internal_type=False)
def _value_set(self, new_value):
warnings.warn(("The .value property of shared variables is deprecated."
" You should use the .set_value() method instead."),
stacklevel=2)
return self.set_value(new_value, borrow=config.shared.value_borrows)
#TODO: USE A CONFIG VARIABLE TO set these get/set methods to the non-borrowing versions
# Semantically things are clearer when using non-borrow versions. That should be the
# default. The default support transparently (if slowly) when the 'raw' value is in a
# different memory space (e.g. GPU or other machine).
value = property(_value_get, _value_set,
doc=("DEPRECATED. Shortcut for self.get_value() and "
"self.set_value(). "
"The `borrow` argument to these methods is read from "
"`theano.config.shared.value_borrows`. "
"You should call get_value() and set_value() directly."))
def filter_update(self, update): def filter_update(self, update):
""" """
When this shared variable is updated by a pfunc, the update value will be run through this function. When this shared variable is updated by a pfunc, the update value will be run through this function.
......
theano/tensor/basic.py
浏览文件 @ 90c3833e
...@@ -1795,6 +1795,7 @@ shape = Shape() ...@@ -1795,6 +1795,7 @@ shape = Shape()
_shape = shape #was used in the past, now use shape directly. _shape = shape #was used in the past, now use shape directly.
pprint.assign(_shape, printing.MemberPrinter('shape')) pprint.assign(_shape, printing.MemberPrinter('shape'))
class SpecifyShape(Op): class SpecifyShape(Op):
""" """
L{Op} put into the graph the user provided shape L{Op} put into the graph the user provided shape
...@@ -1808,14 +1809,18 @@ class SpecifyShape(Op): ...@@ -1808,14 +1809,18 @@ class SpecifyShape(Op):
@note: We currently don't support specifying partial shape information. @note: We currently don't support specifying partial shape information.
""" """
view_map = {0: [0]} view_map = {0: [0]}
def __hash__(self): def __hash__(self):
return hash(type(self)) return hash(type(self))
def __eq__(self, other): def __eq__(self, other):
return type(self) == type(other) return type(self) == type(other)
def __str__(self): def __str__(self):
return self.__class__.__name__ return self.__class__.__name__
def make_node(self, x, shape): def make_node(self, x, shape):
if not isinstance(x,Variable): if not isinstance(x, Variable):
x = as_tensor_variable(x) x = as_tensor_variable(x)
shape = as_tensor_variable(shape) shape = as_tensor_variable(shape)
return Apply(self, [x, shape], [x.type()]) return Apply(self, [x, shape], [x.type()])
...@@ -1823,22 +1828,22 @@ class SpecifyShape(Op): ...@@ -1823,22 +1828,22 @@ class SpecifyShape(Op):
def perform(self, node, inp, out_): def perform(self, node, inp, out_):
x, shape = inp x, shape = inp
out, = out_ out, = out_
assert numpy.all(x.shape==shape), ("got shape", x.shape, assert numpy.all(x.shape == shape), ("got shape", x.shape,
"expected", shape) "expected", shape)
out[0] = x out[0] = x
def infer_shape(self, node, shapes): def infer_shape(self, node, shapes):
xshape, sshape = shapes xshape, sshape = shapes
new_shape=[] new_shape = []
for dim in xrange(node.inputs[0].ndim): for dim in xrange(node.inputs[0].ndim):
try: try:
s=get_constant_value(node.inputs[1][dim]) s = get_constant_value(node.inputs[1][dim])
s=as_tensor_variable(s) s = as_tensor_variable(s)
new_shape.append(s) new_shape.append(s)
except TypeError, e: except TypeError, e:
new_shape.append(node.inputs[1][dim]) new_shape.append(node.inputs[1][dim])
assert len(new_shape)==len(xshape) assert len(new_shape) == len(xshape)
return [new_shape] return [new_shape]
def grad(self, inp, grads): def grad(self, inp, grads):
...@@ -1847,9 +1852,10 @@ class SpecifyShape(Op): ...@@ -1847,9 +1852,10 @@ class SpecifyShape(Op):
# Should I set an SpecifyShape on gz? I think so # Should I set an SpecifyShape on gz? I think so
# But I don't do it now as we need to make an optimization # But I don't do it now as we need to make an optimization
# to remove that op from the graph to don't block other optimization # to remove that op from the graph to don't block other optimization
# Should I do an optimizer that will remove the SpecifyShape? I think Yes # Should I do an optimizer that will remove the SpecifyShape?
# I think Yes
return [gz, None] return [gz, None]
return [specify_shape(gz,s), None] return [specify_shape(gz, s), None]
def R_op(self, inputs, eval_points): def R_op(self, inputs, eval_points):
if eval_points[0] is None: if eval_points[0] is None:
...@@ -1860,97 +1866,83 @@ class SpecifyShape(Op): ...@@ -1860,97 +1866,83 @@ class SpecifyShape(Op):
specify_shape = SpecifyShape() specify_shape = SpecifyShape()
class MaxAndArgmax(Op): class MaxAndArgmax(Op):
"""Calculate the max and argmax over a given axis. """Calculate the max and argmax over a given axis.
.. note::
If axis is None it means to calculate the max over the last dimension which is
DIFFERENT FROM NUMPY!!
To have the behavior of numpy do a flatten of the input before passing the data to this op.
If the input to flatten is not ccontiguous, this will make a copy to a contiguous version.
""" """
nin=2 # tensor, axis nin = 2 # tensor, axis
nout=2 # max val, max idx nout = 2 # max val, max idx
E_axis = 'invalid axis' E_axis = 'invalid axis'
def __eq__(self,other): def __eq__(self, other):
return type(self)==type(other) return type(self) == type(other)
def __hash__(self): def __hash__(self):
return hash(type(self)) return hash(type(self))
def make_node(self, x, axis='DEFAULT'): def make_node(self, x, axis=None):
x = _as_tensor_variable(x) x = _as_tensor_variable(x)
if x.type.ndim <= 1 and axis in ('DEFAULT', None): if isinstance(axis, int):
# The old and new behavior are not different.
axis = 0
if axis=='DEFAULT':
axis=x.type.ndim - 1
warnings.warn(("The default axis of MaxAndArgmax will change! "
"Now we return the max and the armax over the last dimensions. "
"It will change to be the same as numpy: the max and argmax over "
"all dimensions. To hide this warning and be compatible with the "
"future behavior, set axis to -1 to have the current behavior. "
"MaxAndArgmax currently support axis over only 1 dimensions, so "
"you must flatten the tensor to have the futur behavior."),
stacklevel=3)
elif axis is None:
axis = x.type.ndim - 1
warnings.warn(("The behavior of MaxAndArgmax when axis==None will "
"change! Now we return the max and argmax over the last "
"dimensions. It will change to the max and argmax over all "
"dimensions as numpy. To hide this warning and be compatible with "
"the future behavior, set axis to -1 to have the current behavior. "
"MaxAndArgmax currently support axis over only 1 dimensions, so "
"you must flatten the tensor to have the futur behavior."),
stacklevel=3)
if isinstance(axis,int):
axis = [axis] axis = [axis]
elif isinstance(axis,(tuple,list)): elif isinstance(axis, (tuple, list)):
assert len(axis)==1,"MaxAndArgmax don't support multiple axis. the max fct support it." if len(axis) != 1:
#we make the axis all positive to make the infer_shape work with negative axis list(axis)
if x.type.ndim>0: axis.sort()
for id,a in enumerate(axis): assert axis == range(x.type.ndim), (
if not isinstance(a, TensorVariable) and a<0: "MaxAndArgmax don't support multiple"
if -a>x.type.ndim: " axis. the max fct support it.")
# we make the axis all positive to make the infer_shape work
# with negative axis
if x.type.ndim > 0 and axis is not None:
for id, a in enumerate(axis):
if not isinstance(a, TensorVariable) and a < 0:
if -a > x.type.ndim:
raise ValueError('axis out of range') raise ValueError('axis out of range')
axis[id]=x.type.ndim+a axis[id] = x.type.ndim + a
axis = _as_tensor_variable(axis) if axis is None:
axis = _as_tensor_variable(range(x.type.ndim))
else:
axis = _as_tensor_variable(axis)
inputs = [x, axis] inputs = [x, axis]
#TODO: figure things out if axis is a constant broadcastable = [False] * (x.type.ndim - len(axis.data))
broadcastable = [False] * (x.type.ndim - 1) outputs = [tensor(x.type.dtype, broadcastable, name='max'),
outputs = [tensor(x.type.dtype, broadcastable,name='max'), tensor('int32', broadcastable, name='argmax')]
tensor('int32', broadcastable,name='argmax')]
return Apply(self, inputs, outputs) return Apply(self, inputs, outputs)
def perform(self, node, inp, outs): def perform(self, node, inp, outs):
x, axis = inp x, axis = inp
max, max_idx = outs max, max_idx = outs
if len(axis) == 0 or python_all(axis == range(x.ndim)):
axis = None
max[0] = numpy.asarray(numpy.max(x, axis)) max[0] = numpy.asarray(numpy.max(x, axis))
max_idx[0] = theano._asarray(numpy.argmax(x, axis), dtype='int32') max_idx[0] = theano._asarray(numpy.argmax(x, axis), dtype='int32')
def infer_shape(self, node, shapes): def infer_shape(self, node, shapes):
ishape, axis_shape = shapes ishape, axis_shape = shapes
axis=node.inputs[1] axis = node.inputs[1]
if axis is None: if python_all(axis.data == range(node.inputs[0].ndim)):
return [(),()] return [(), ()]
rval = tuple([ishape[i] for (i,b) in enumerate(node.inputs[0].type.broadcastable) if i !=axis.data]) rval = tuple([ishape[i] for (i, b) in enumerate(
return [rval,rval] node.inputs[0].type.broadcastable) if i != axis.data])
return [rval, rval]
def R_op(self, inputs, eval_points): def R_op(self, inputs, eval_points):
if eval_points[0] is None: if eval_points[0] is None:
return [None, None] return [None, None]
if not isinstance(inputs[1], theano.Constant): if not isinstance(inputs[1], theano.Constant):
raise ValueError( ('R_op supported for arg_max only for ' raise ValueError(('R_op supported for arg_max only for '
'constant axis!')) 'constant axis!'))
if inputs[1].data > 1: if inputs[1].data > 1:
raise ValueError( ('R_op supported for arg_max only when ' raise ValueError(('R_op supported for arg_max only when '
' axis is 0 or 1')) ' axis is 0 or 1'))
if inputs[0].ndim != 2: if inputs[0].ndim != 2:
raise ValueError( ('R_op supported for arg_max only when ' raise ValueError(('R_op supported for arg_max only when '
' input is a matrix')) ' input is a matrix'))
max_vals, max_pos = self.make_node(*inputs).outputs max_vals, max_pos = self.make_node(*inputs).outputs
if inputs[1].data == 0: if inputs[1].data == 0:
return [eval_points[0][max_pos, arange(eval_points[0].shape[1])], None] return [eval_points[0][max_pos,
arange(eval_points[0].shape[1])], None]
else: else:
return [eval_points[0][arange(eval_points[0].shape[0]), return [eval_points[0][arange(eval_points[0].shape[0]),
max_pos], None] max_pos], None]
...@@ -1987,112 +1979,48 @@ class MaxAndArgmax(Op): ...@@ -1987,112 +1979,48 @@ class MaxAndArgmax(Op):
def __str__(self): def __str__(self):
return self.__class__.__name__ return self.__class__.__name__
_max_and_argmax = MaxAndArgmax() _max_and_argmax = MaxAndArgmax()
@_redefine_asRoutine(_max_and_argmax) @_redefine_asRoutine(_max_and_argmax)
def max_and_argmax(a): def max_and_argmax(a):
pass pass
@constructor @constructor
def max(x, axis='DEFAULT'): def max(x, axis=None):
""" """
Return maximum elements obtained by iterating over given axis Return maximum elements obtained by iterating over given axis
Default axis is the last one. This will change. Default axis is None: max over all dimensions.
:note: we return an error as numpy when we reduce a dim with a shape of 0 :note: we return an error as numpy when we reduce a dim with a shape of 0
:note2: see MaxAndArgmax note for a difference between numpy and theano when axis==None
""" """
if x.type.ndim <= 1 and axis in ('DEFAULT', None): if isinstance(axis, (list, tuple)) and len(axis) > 1:
# The old and new behavior are not different. return CAReduce(scal.maximum, axis)(x)
axis = 0
elif axis=='DEFAULT':
axis = x.type.ndim - 1
warnings.warn(("The default axis of max will change! Now we return the "
"max over the last dimensions. It will change to be the same as numpy: "
"the max over all dimensions. To hide this warning and be compatible "
"with the future behavior, set axis to -1 to have the current "
"behavior. To have the futur behavior set axis to range(nb dim), but "
"this don't support the grad. To have the grad, you must flatten the "
"tensor before calling max()."),
stacklevel=2)
elif axis is None:
axis = x.type.ndim - 1
warnings.warn(("The behavior of max when axis==None will change! Now "
"we return the max over the last dimensions. It will change to the max "
"over all dimensions as numpy. To hide this warning and be compatible "
"with the future behavior, set axis to -1 to have the current "
"behavior. To have the futur behavior set axis to range(nb dim), but "
"this don't support the grad. To have the grad, you must flatten the "
"tensor before calling max()."),
stacklevel=2)
if isinstance(axis,(list,tuple)) and len(axis)>1:
return CAReduce(scal.maximum,axis)(x)
try: try:
const = get_constant_value(axis) const = get_constant_value(axis)
return CAReduce(scal.maximum,list(const))(x) return CAReduce(scal.maximum, list(const))(x)
except Exception: except Exception:
return max_and_argmax(x,axis)[0] return max_and_argmax(x, axis)[0]
@constructor @constructor
def argmax(x, axis='DEFAULT'): def argmax(x, axis=None):
""" """
Return indexes of maximum elements obtained by iterating over given axis Return indexes of maximum elements obtained by iterating over given axis
Default axis is the last one. This will change. When axis is None (the default value), the argmax is performed
over the flattened tensor.
""" """
if x.type.ndim <= 1 and axis in ('DEFAULT', None):
# The old and new behavior are not different.
axis = 0
elif axis=='DEFAULT':
axis = x.type.ndim - 1
warnings.warn(("The default axis of argmax will change! Now we return "
"the argmax over the last dimensions. It will change to be the same as "
"numpy: the argmax over all dimensions. To hide this warning and be "
"compatible with the future behavior, set axis to -1 to have the "
"current behavior. To have the futur behavior, you must flatten the "
"tensor before calling max()."),
stacklevel=2)
elif axis is None:
axis = x.type.ndim - 1
warnings.warn(("The behavior of argmax when axis==None will change! "
"Now we return the argmax over the last dimensions. It will change to "
"the argmax over all dimensions as numpy. To hide this warning and be "
"compatible with the future behavior, set axis to -1 to have the "
"current behavior. To have the futur behavior, you must flatten the "
"tensor before calling argmax()."),
stacklevel=2)
# In python (using MaxAndArgmax.perform()) this leads to an wasteful # In python (using MaxAndArgmax.perform()) this leads to an wasteful
# implementation that goes through the data twice instead of once # implementation that goes through the data twice instead of once
# but when Argmax.c_impl() is in place, it should be fine. # but when Argmax.c_impl() is in place, it should be fine.
return max_and_argmax(x,axis)[1] return max_and_argmax(x, axis)[1]
@constructor @constructor
def min(x, axis='DEFAULT'): def min(x, axis=None):
if x.type.ndim <= 1 and axis in ('DEFAULT', None):
# The old and new behavior are not different.
axis = 0
elif axis=='DEFAULT':
axis = x.type.ndim - 1
warnings.warn(("The default axis of min will change! Now we return the "
"min over the last dimensions. It will change to be the same as numpy: "
"the min over all dimensions. To hide this warning and be compatible "
"with the future behavior, set axis to -1 to have the current "
"behavior. To have the future behavior, set axis to range(x.ndim), but "
"this does not support the grad. To be able to get the grad, you must "
"flatten the tensor before calling min()."),
stacklevel=2)
elif axis is None:
axis = x.type.ndim - 1
warnings.warn(("The behavior of min when axis is None will change! Now "
"we return the min over the last dimensions. It will change to the min "
"over all dimensions as numpy. To hide this warning and be compatible "
"with the future behavior, set axis to -1 to have the current "
"behavior. To have the future behavior, set axis to range(x.ndim), but "
"this does not support the grad. To be able to get the grad, you must "
"flatten the tensor before calling min()."),
stacklevel=2)
str_x_type = str(x.dtype) str_x_type = str(x.dtype)
if str_x_type.startswith('float') or str_x_type in int_dtypes: if str_x_type.startswith('float') or str_x_type in int_dtypes:
return -max(-x, axis=axis) return -max(-x, axis=axis)
...@@ -2100,29 +2028,9 @@ def min(x, axis='DEFAULT'): ...@@ -2100,29 +2028,9 @@ def min(x, axis='DEFAULT'):
#Be careful about unsigned integers, complex #Be careful about unsigned integers, complex
raise NotImplementedError() raise NotImplementedError()
@constructor @constructor
def argmin(x, axis='DEFAULT'): def argmin(x, axis=None):
if x.type.ndim <= 1 and axis in ('DEFAULT', None):
# The old and new behavior are not different.
axis = 0
elif axis=='DEFAULT':
axis = x.type.ndim - 1
warnings.warn(("The default axis of argmin will change! Now we return "
"the argmin over the last dimensions. It will change to be the same as "
"numpy: the argmin over all dimensions. To hide this warning and be "
"compatible with the future behavior, set axis to -1 to have the "
"current behavior. To have the futur behavior, you must flatten the "
"axis before calling argmin."),
stacklevel=2)
elif axis is None:
axis = x.type.ndim - 1
warnings.warn(("The behavior of argmin when axis==None will change! "
"Now we return the argmin over the last dimensions. It will change to "
"the argmin over all dimensions as numpy. To hide this warning and be "
"compatible with the future behavior, set axis to -1 to have the "
"current behavior. To have the futur behavior, you must flatten the "
"axis before calling argmin."),
stacklevel=2)
str_x_type = str(x.dtype) str_x_type = str(x.dtype)
if str_x_type.startswith('float') or str_x_type in int_dtypes: if str_x_type.startswith('float') or str_x_type in int_dtypes:
return argmax(-x, axis=axis) return argmax(-x, axis=axis)
...@@ -2130,6 +2038,7 @@ def argmin(x, axis='DEFAULT'): ...@@ -2130,6 +2038,7 @@ def argmin(x, axis='DEFAULT'):
#Be careful about unsigned integers, complex #Be careful about unsigned integers, complex
raise NotImplementedError() raise NotImplementedError()
@constructor @constructor
def smallest(*args): def smallest(*args):
"""Return the [elementwise] smallest of a variable number of arguments (like python's min).""" """Return the [elementwise] smallest of a variable number of arguments (like python's min)."""
......
theano/tensor/opt_uncanonicalize.py
浏览文件 @ 90c3833e
...@@ -57,7 +57,7 @@ class MaxAndArgmaxOptimizer(Optimizer): ...@@ -57,7 +57,7 @@ class MaxAndArgmaxOptimizer(Optimizer):
if len(node.outputs[1].clients)==0: if len(node.outputs[1].clients)==0:
try: try:
axis=get_constant_value(node.inputs[1]) axis=get_constant_value(node.inputs[1])
except ValueError: except (ValueError, TypeError), e:
return False return False
new = CAReduce(scal.maximum,axis)(node.inputs[0]) new = CAReduce(scal.maximum,axis)(node.inputs[0])
......
theano/tensor/tests/test_basic.py
浏览文件 @ 90c3833e
...@@ -1449,6 +1449,7 @@ class T_Shape(unittest.TestCase): ...@@ -1449,6 +1449,7 @@ class T_Shape(unittest.TestCase):
s = shape(numpy.ones((5, 3, 10))) s = shape(numpy.ones((5, 3, 10)))
self.assertTrue((eval_outputs([s]) == [5, 3, 10]).all()) self.assertTrue((eval_outputs([s]) == [5, 3, 10]).all())
class T_max_and_argmax(unittest.TestCase): class T_max_and_argmax(unittest.TestCase):
def setUp(self): def setUp(self):
utt.seed_rng() utt.seed_rng()
...@@ -1456,108 +1457,91 @@ class T_max_and_argmax(unittest.TestCase): ...@@ -1456,108 +1457,91 @@ class T_max_and_argmax(unittest.TestCase):
def test0(self): def test0(self):
n = as_tensor_variable(5.0) n = as_tensor_variable(5.0)
v,i = eval_outputs(max_and_argmax(n)) v, i = eval_outputs(max_and_argmax(n))
self.assertTrue(v == 5.0) self.assertTrue(v == 5.0)
self.assertTrue(i == 0) self.assertTrue(i == 0)
v = eval_outputs(max_and_argmax(n)[0].shape) v = eval_outputs(max_and_argmax(n)[0].shape)
assert len(v)==0 assert len(v) == 0
v = eval_outputs(max_and_argmax(n)[1].shape) v = eval_outputs(max_and_argmax(n)[1].shape)
assert len(v)==0 assert len(v) == 0
def test1(self): def test1(self):
n = as_tensor_variable([1,2,3,2,-6]) n = as_tensor_variable([1, 2, 3, 2, -6])
v,i = eval_outputs(max_and_argmax(n)) v, i = eval_outputs(max_and_argmax(n))
self.assertTrue(v == 3) self.assertTrue(v == 3)
self.assertTrue(i == 2) self.assertTrue(i == 2)
v = eval_outputs(max_and_argmax(n)[0].shape) v = eval_outputs(max_and_argmax(n)[0].shape)
assert len(v)==0 assert len(v) == 0
def test2(self): def test2(self):
data = numpy.random.rand(2,3) data = numpy.random.rand(2, 3)
n = as_tensor_variable(data)
v,i = eval_outputs(max_and_argmax(n,-1))
self.assertTrue(numpy.all(v == numpy.max(data,-1)))
self.assertTrue(numpy.all(i == numpy.argmax(data,-1)))
v = eval_outputs(max_and_argmax(n,-1)[0].shape)
assert v==(2)
def test2b(self):
data = numpy.random.rand(2,3)
n = as_tensor_variable(data) n = as_tensor_variable(data)
v,i = eval_outputs(max_and_argmax(n,0)) for (axis, np_axis) in [(-1, -1), (0, 0), (1, 1), (None, None),
self.assertTrue(numpy.all(v == numpy.max(data,0))) ([0, 1], None), ([1, 0], None)]:
self.assertTrue(numpy.all(i == numpy.argmax(data,0))) v, i = eval_outputs(max_and_argmax(n, axis))
v = eval_outputs(max_and_argmax(n,0)[0].shape) self.assertTrue(numpy.all(v == numpy.max(data, np_axis)))
assert v==(3) self.assertTrue(numpy.all(i == numpy.argmax(data, np_axis)))
v = eval_outputs(max_and_argmax(n,1)[0].shape) v_shape = eval_outputs(max_and_argmax(n, axis)[0].shape)
assert v==(2) assert tuple(v_shape) == numpy.max(data, np_axis).shape
#currently not supported
#v = eval_outputs(max_and_argmax(n,[0,1])[0].shape)
#assert v.size==0
def test2_invalid(self): def test2_invalid(self):
n = as_tensor_variable(numpy.random.rand(2,3)) n = as_tensor_variable(numpy.random.rand(2, 3))
# Silence expected error messages # Silence expected error messages
_logger = logging.getLogger('theano.gof.opt') _logger = logging.getLogger('theano.gof.opt')
oldlevel = _logger.level oldlevel = _logger.level
_logger.setLevel(logging.CRITICAL) _logger.setLevel(logging.CRITICAL)
try: try:
try: try:
eval_outputs(max_and_argmax(n,3)) eval_outputs(max_and_argmax(n, 3))
assert False assert False
except ValueError, e: except ValueError, e:
pass pass
finally: finally:
_logger.setLevel(oldlevel) _logger.setLevel(oldlevel)
def test2_invalid_neg(self): def test2_invalid_neg(self):
n = as_tensor_variable(numpy.random.rand(2,3)) n = as_tensor_variable(numpy.random.rand(2, 3))
old_stderr = sys.stderr old_stderr = sys.stderr
sys.stderr = StringIO.StringIO() sys.stderr = StringIO.StringIO()
try: try:
try: try:
eval_outputs(max_and_argmax(n,-3)) eval_outputs(max_and_argmax(n, -3))
assert False assert False
except ValueError, e: except ValueError, e:
pass pass
finally: finally:
sys.stderr = old_stderr sys.stderr = old_stderr
def test2_valid_neg(self): def test2_valid_neg(self):
n = as_tensor_variable(numpy.random.rand(2,3)) n = as_tensor_variable(numpy.random.rand(2, 3))
v,i = eval_outputs(max_and_argmax(n,-1)) v, i = eval_outputs(max_and_argmax(n, -1))
self.assertTrue(v.shape == (2,)) self.assertTrue(v.shape == (2,))
self.assertTrue(i.shape == (2,)) self.assertTrue(i.shape == (2,))
self.assertTrue(numpy.all(v == numpy.max(n.value,-1))) self.assertTrue(numpy.all(v == numpy.max(n.value, -1)))
self.assertTrue(numpy.all(i == numpy.argmax(n.value,-1))) self.assertTrue(numpy.all(i == numpy.argmax(n.value, -1)))
v,i = eval_outputs(max_and_argmax(n,-2)) v, i = eval_outputs(max_and_argmax(n, -2))
self.assertTrue(v.shape == (3,)) self.assertTrue(v.shape == (3,))
self.assertTrue(i.shape == (3,)) self.assertTrue(i.shape == (3,))
self.assertTrue(numpy.all(v == numpy.max(n.value,-2))) self.assertTrue(numpy.all(v == numpy.max(n.value, -2)))
self.assertTrue(numpy.all(i == numpy.argmax(n.value,-2))) self.assertTrue(numpy.all(i == numpy.argmax(n.value, -2)))
v = eval_outputs(max_and_argmax(n,-1)[0].shape) v = eval_outputs(max_and_argmax(n, -1)[0].shape)
assert v==(2) assert v == (2)
v = eval_outputs(max_and_argmax(n,-2)[0].shape) v = eval_outputs(max_and_argmax(n, -2)[0].shape)
assert v==(3) assert v == (3)
def test3(self): def test3(self):
n = as_tensor_variable(numpy.random.rand(2,3,4)) data = numpy.random.rand(2, 3, 4)
v,i = eval_outputs(max_and_argmax(n,0)) n = as_tensor_variable(data)
self.assertTrue(v.shape == (3,4)) for (axis, np_axis) in [(-1, -1), (0, 0), (1, 1), (None, None),
self.assertTrue(i.shape == (3,4)) ([0, 1, 2], None), ([1, 2, 0], None)]:
v,i = eval_outputs(max_and_argmax(n,1)) v, i = eval_outputs(max_and_argmax(n, axis))
self.assertTrue(v.shape == (2,4)) self.assertTrue(numpy.all(v == numpy.max(data, np_axis)))
self.assertTrue(i.shape == (2,4)) self.assertTrue(numpy.all(i == numpy.argmax(data, np_axis)))
v,i = eval_outputs(max_and_argmax(n,2)) v = eval_outputs(max_and_argmax(n, axis)[0].shape)
self.assertTrue(v.shape == (2,3)) assert tuple(v) == numpy.max(data, np_axis).shape
self.assertTrue(i.shape == (2,3))
v = eval_outputs(max_and_argmax(n,0)[0].shape)
assert tuple(v)==(3,4)
v = eval_outputs(max_and_argmax(n,1)[0].shape)
assert tuple(v)==(2,4)
v = eval_outputs(max_and_argmax(n,2)[0].shape)
assert tuple(v)==(2,3)
def test_grad(self): def test_grad(self):
data = numpy.random.rand(2,3) data = numpy.random.rand(2, 3)
n = as_tensor_variable(data) n = as_tensor_variable(data)
def check_grad_max(data, max_grad_data, axis=None): def check_grad_max(data, max_grad_data, axis=None):
...@@ -1565,35 +1549,39 @@ class T_max_and_argmax(unittest.TestCase): ...@@ -1565,35 +1549,39 @@ class T_max_and_argmax(unittest.TestCase):
Why this is needed? verify_grad is not enought? Why this is needed? verify_grad is not enought?
""" """
#This work only for axis in [0,None] #This work only for axis in [0,None]
assert axis in [0,None] assert axis in [0, None]
z = numpy.zeros_like(data) z = numpy.zeros_like(data)
z = z.flatten() z = z.flatten()
argmax=numpy.argmax(data,axis=axis) argmax = numpy.argmax(data, axis=axis)
if argmax.ndim==0: if argmax.ndim == 0:
z[argmax]+=1 z[argmax] += 1
else: else:
for id,v in enumerate(argmax): for id, v in enumerate(argmax):
z[v*numpy.prod(data.shape[data.ndim-1:axis:-1])+id]+=1 z[v * numpy.prod(data.shape[data.ndim - 1:axis:-1])
+ id] += 1
z = z.reshape(data.shape) z = z.reshape(data.shape)
assert numpy.all(max_grad_data == z) assert numpy.all(max_grad_data == z)
#test grad of max #test grad of max
#axis is the last one #axis is the last one
utt.verify_grad(lambda v: max_and_argmax(v,axis=-1)[0], [data]) utt.verify_grad(lambda v: max_and_argmax(v, axis=-1)[0], [data])
utt.verify_grad(lambda v: max_and_argmax(v,axis=-1)[1], [data]) utt.verify_grad(lambda v: max_and_argmax(v, axis=-1)[1], [data])
utt.verify_grad(lambda v: max_and_argmax(v,axis=[0])[0], [data]) utt.verify_grad(lambda v: max_and_argmax(v, axis=[0])[0], [data])
utt.verify_grad(lambda v: max_and_argmax(v,axis=[0])[1], [data]) utt.verify_grad(lambda v: max_and_argmax(v, axis=[0])[1], [data])
check_grad_max(data,eval_outputs(grad(max_and_argmax(n,axis=0)[0].sum(),n)),axis=0) check_grad_max(data, eval_outputs(grad(
max_and_argmax(n, axis=0)[0].sum(), n)), axis=0)
utt.verify_grad(lambda v: max_and_argmax(v,axis=[1])[0], [data]) utt.verify_grad(lambda v: max_and_argmax(v, axis=[1])[0], [data])
utt.verify_grad(lambda v: max_and_argmax(v,axis=[1])[1], [data]) utt.verify_grad(lambda v: max_and_argmax(v, axis=[1])[1], [data])
#check_grad_max(data,eval_outputs(grad(max_and_argmax(n,axis=1)[0],n)),axis=1) #check_grad_max(data,eval_outputs(grad(
# max_and_argmax(n,axis=1)[0],n)),axis=1)
utt.verify_grad(lambda v: max_and_argmax(v.flatten())[0], [data]) utt.verify_grad(lambda v: max_and_argmax(v.flatten())[0], [data])
utt.verify_grad(lambda v: max_and_argmax(v.flatten())[1], [data]) utt.verify_grad(lambda v: max_and_argmax(v.flatten())[1], [data])
check_grad_max(data,eval_outputs(grad(max_and_argmax(n.flatten())[0],n))) check_grad_max(data, eval_outputs(grad(
max_and_argmax(n.flatten())[0], n)))
# Test 4d inner dimensions # Test 4d inner dimensions
data = numpy.random.rand(2, 3, 4, 5) data = numpy.random.rand(2, 3, 4, 5)
...@@ -1608,60 +1596,48 @@ class T_argmin_argmax(unittest.TestCase): ...@@ -1608,60 +1596,48 @@ class T_argmin_argmax(unittest.TestCase):
utt.seed_rng() utt.seed_rng()
MaxAndArgmax.debug = 0 MaxAndArgmax.debug = 0
def test0(self): def test_scalar(self):
for fct in [argmin,argmax]: for fct in [argmin, argmax]:
n = as_tensor_variable(5.0) n = as_tensor_variable(5.0)
i = eval_outputs(fct(n)) i = eval_outputs(fct(n))
self.assertTrue(i == 0) self.assertTrue(i == 0)
v = eval_outputs(fct(n).shape) v = eval_outputs(fct(n).shape)
assert len(v)==0 assert len(v) == 0
def test1(self): def test_list(self):
n = as_tensor_variable([1,2,3,2,-6]) n = as_tensor_variable([1, 2, 3, 2, -6])
i = eval_outputs(argmin(n)) i = eval_outputs(argmin(n))
self.assertTrue(i == 4) self.assertTrue(i == 4)
v = eval_outputs(argmin(n).shape) v = eval_outputs(argmin(n).shape)
assert len(v)==0 assert len(v) == 0
n = as_tensor_variable([1,2,3,2,-6]) n = as_tensor_variable([1, 2, 3, 2, -6])
i = eval_outputs(argmax(n)) i = eval_outputs(argmax(n))
self.assertTrue(i == 2) self.assertTrue(i == 2)
v = eval_outputs(argmax(n).shape) v = eval_outputs(argmax(n).shape)
assert len(v)==0 assert len(v) == 0
def test2(self): def test2(self):
for fct,nfct in [(argmax,numpy.argmax),(argmin,numpy.argmin)]: data = numpy.random.rand(2, 3)
data = numpy.random.rand(2,3) n = as_tensor_variable(data)
n = as_tensor_variable(data) for fct, nfct in [(argmax, numpy.argmax), (argmin, numpy.argmin)]:
i = eval_outputs(fct(n,-1)) for (axis, np_axis) in [(-1, -1), (0, 0), (1, 1), (None, None),
self.assertTrue(numpy.all(i == nfct(data,-1))) ([0, 1], None), ([1, 0], None)]:
v = eval_outputs(fct(n,-1).shape) v = eval_outputs(fct(n, axis))
assert v==(2) self.assertTrue(numpy.all(v == nfct(data, np_axis)))
v_shape = eval_outputs(fct(n, axis).shape)
def test2b(self): assert tuple(v_shape) == nfct(data, np_axis).shape
for fct,nfct in [(argmax,numpy.argmax),(argmin,numpy.argmin)]:
data = numpy.random.rand(2,3)
n = as_tensor_variable(data)
i = eval_outputs(fct(n,0))
self.assertTrue(numpy.all(i == nfct(data,0)))
v = eval_outputs(fct(n,0).shape)
assert v==(3)
v = eval_outputs(fct(n,1).shape)
assert v==(2)
#currently not supported
#v = eval_outputs(fct(n,[0,1]).shape)
#assert v.size==0
def test2_invalid(self): def test2_invalid(self):
for fct,nfct in [(argmax,numpy.argmax),(argmin,numpy.argmin)]: for fct, nfct in [(argmax, numpy.argmax), (argmin, numpy.argmin)]:
n = as_tensor_variable(numpy.random.rand(2,3)) n = as_tensor_variable(numpy.random.rand(2, 3))
# Silence expected error messages # Silence expected error messages
_logger = logging.getLogger('theano.gof.opt') _logger = logging.getLogger('theano.gof.opt')
oldlevel = _logger.level oldlevel = _logger.level
_logger.setLevel(logging.CRITICAL) _logger.setLevel(logging.CRITICAL)
try: try:
try: try:
eval_outputs(fct(n,3)) eval_outputs(fct(n, 3))
assert False assert False
except ValueError, e: except ValueError, e:
pass pass
...@@ -1669,13 +1645,13 @@ class T_argmin_argmax(unittest.TestCase): ...@@ -1669,13 +1645,13 @@ class T_argmin_argmax(unittest.TestCase):
_logger.setLevel(oldlevel) _logger.setLevel(oldlevel)
def test2_invalid_neg(self): def test2_invalid_neg(self):
for fct,nfct in [(argmax,numpy.argmax),(argmin,numpy.argmin)]: for fct, nfct in [(argmax, numpy.argmax), (argmin, numpy.argmin)]:
n = as_tensor_variable(numpy.random.rand(2,3)) n = as_tensor_variable(numpy.random.rand(2, 3))
old_stderr = sys.stderr old_stderr = sys.stderr
sys.stderr = StringIO.StringIO() sys.stderr = StringIO.StringIO()
try: try:
try: try:
eval_outputs(fct(n,-3)) eval_outputs(fct(n, -3))
assert False assert False
except ValueError, e: except ValueError, e:
pass pass
...@@ -1683,286 +1659,262 @@ class T_argmin_argmax(unittest.TestCase): ...@@ -1683,286 +1659,262 @@ class T_argmin_argmax(unittest.TestCase):
sys.stderr = old_stderr sys.stderr = old_stderr
def test2_valid_neg(self): def test2_valid_neg(self):
for fct,nfct in [(argmax,numpy.argmax),(argmin,numpy.argmin)]: for fct, nfct in [(argmax, numpy.argmax), (argmin, numpy.argmin)]:
n = as_tensor_variable(numpy.random.rand(2,3)) n = as_tensor_variable(numpy.random.rand(2, 3))
i = eval_outputs(fct(n,-1)) i = eval_outputs(fct(n, -1))
self.assertTrue(i.shape == (2,)) self.assertTrue(i.shape == (2,))
self.assertTrue(numpy.all(i == nfct(n.value,-1))) self.assertTrue(numpy.all(i == nfct(n.value, -1)))
i = eval_outputs(fct(n,-2)) i = eval_outputs(fct(n, -2))
self.assertTrue(i.shape == (3,)) self.assertTrue(i.shape == (3,))
self.assertTrue(numpy.all(i == nfct(n.value,-2))) self.assertTrue(numpy.all(i == nfct(n.value, -2)))
v = eval_outputs(fct(n,-1).shape) v = eval_outputs(fct(n, -1).shape)
assert v==(2) assert v == (2)
v = eval_outputs(fct(n,-2).shape) v = eval_outputs(fct(n, -2).shape)
assert v==(3) assert v == (3)
def test3(self): def test3(self):
for fct,nfct in [(argmax,numpy.argmax),(argmin,numpy.argmin)]: data = numpy.random.rand(2, 3, 4)
n = as_tensor_variable(numpy.random.rand(2,3,4)) n = as_tensor_variable(data)
i = eval_outputs(fct(n,0)) for fct, nfct in [(argmax, numpy.argmax), (argmin, numpy.argmin)]:
self.assertTrue(i.shape == (3,4)) for (axis, np_axis) in [(-1, -1), (0, 0), (1, 1), (2, 2),
self.assertTrue(numpy.all(i == nfct(n.value,0))) (None, None), ([0, 1, 2], None),
i = eval_outputs(fct(n,1)) ([1, 0, 2], None)]:
self.assertTrue(i.shape == (2,4)) v = eval_outputs(fct(n, axis))
self.assertTrue(numpy.all(i == nfct(n.value,1))) self.assertTrue(numpy.all(v == nfct(data, np_axis)))
i = eval_outputs(fct(n,2)) v_shape = eval_outputs(fct(n, axis).shape)
self.assertTrue(i.shape == (2,3)) assert tuple(v_shape) == nfct(data, np_axis).shape
self.assertTrue(numpy.all(i == nfct(n.value,2)))
v = eval_outputs(fct(n,0).shape)
assert tuple(v)==(3,4)
v = eval_outputs(fct(n,1).shape)
assert tuple(v)==(2,4)
v = eval_outputs(fct(n,2).shape)
assert tuple(v)==(2,3)
def test_grad_argmin(self): def test_grad_argmin(self):
data = numpy.random.rand(2,3) data = numpy.random.rand(2, 3)
n = as_tensor_variable(data) n = as_tensor_variable(data)
#test grad of argmin #test grad of argmin
utt.verify_grad(lambda v: argmin(v,axis=-1), [data]) utt.verify_grad(lambda v: argmin(v, axis=-1), [data])
utt.verify_grad(lambda v: argmin(v,axis=[0]), [data]) utt.verify_grad(lambda v: argmin(v, axis=[0]), [data])
utt.verify_grad(lambda v: argmin(v,axis=[1]), [data]) utt.verify_grad(lambda v: argmin(v, axis=[1]), [data])
utt.verify_grad(lambda v: argmin(v.flatten()), [data]) utt.verify_grad(lambda v: argmin(v.flatten()), [data])
try: try:
grad(argmin(n,axis=-1),n) grad(argmin(n, axis=-1), n)
raise Exception('Expected an error') raise Exception('Expected an error')
except TypeError: except TypeError:
pass pass
def test_grad_argmax(self): def test_grad_argmax(self):
data = numpy.random.rand(2,3) data = numpy.random.rand(2, 3)
n = as_tensor_variable(data) n = as_tensor_variable(data)
#test grad of argmax #test grad of argmax
utt.verify_grad(lambda v: argmax(v, axis=-1), [data]) utt.verify_grad(lambda v: argmax(v, axis=-1), [data])
utt.verify_grad(lambda v: argmax(v,axis=[0]), [data]) utt.verify_grad(lambda v: argmax(v, axis=[0]), [data])
utt.verify_grad(lambda v: argmax(v,axis=[1]), [data]) utt.verify_grad(lambda v: argmax(v, axis=[1]), [data])
utt.verify_grad(lambda v: argmax(v.flatten()), [data]) utt.verify_grad(lambda v: argmax(v.flatten()), [data])
try: try:
grad(argmax(n, axis=-1),n) grad(argmax(n, axis=-1), n)
raise Exception('Expected an error') raise Exception('Expected an error')
except TypeError: except TypeError:
pass pass
class T_min_max(unittest.TestCase): class T_min_max(unittest.TestCase):
def setUp(self): def setUp(self):
utt.seed_rng() utt.seed_rng()
MaxAndArgmax.debug = 0 MaxAndArgmax.debug = 0
def test0(self): def test_scalar(self):
for fct in [max,min]: for fct in [max, min]:
n = as_tensor_variable(5.0) n = as_tensor_variable(5.0)
v = eval_outputs(fct(n)) v = eval_outputs(fct(n))
self.assertTrue(v == 5.0) self.assertTrue(v == 5.0)
v = eval_outputs(fct(n).shape) v = eval_outputs(fct(n).shape)
assert len(v)==0 assert len(v) == 0
def test1(self): def test_list(self):
for fct,nfct in [(max,numpy.max),(min,numpy.min)]: for fct, nfct in [(max, numpy.max), (min, numpy.min)]:
n = as_tensor_variable([1,2,3,2,-6]) n = as_tensor_variable([1, 2, 3, 2, -6])
v = eval_outputs([fct(n)]) v = eval_outputs([fct(n)])
self.assertTrue(v == nfct(n.value)) self.assertTrue(v == nfct(n.value))
v = eval_outputs(fct(n).shape) v = eval_outputs(fct(n).shape)
assert len(v)==0 assert len(v) == 0
def test2(self): def test2(self):
for fct,nfct in [(max,numpy.max),(min,numpy.min)]: data = numpy.random.rand(2, 3)
data = numpy.random.rand(2,3) n = as_tensor_variable(data)
n = as_tensor_variable(data) for fct, nfct in [(max, numpy.max), (min, numpy.min)]:
v = eval_outputs(fct(n,-1)) for (axis, np_axis) in [(-1, -1), (0, 0), (1, 1), (None, None),
self.assertTrue(numpy.all(v == nfct(data,-1))) ([0, 1], None), ([1, 0], None)]:
v = eval_outputs(fct(n, axis))
v = eval_outputs(fct(n,-1).shape) self.assertTrue(numpy.all(v == nfct(data, np_axis)))
assert v==(2) v_shape = eval_outputs(fct(n, axis).shape)
assert tuple(v_shape) == nfct(data, np_axis).shape
def test2b(self):
for fct,nfct in [(max,numpy.max),(min,numpy.min)]:
data = numpy.random.rand(2,3)
n = as_tensor_variable(data)
v = eval_outputs(fct(n,0))
self.assertTrue(numpy.all(v == nfct(data,0)))
v = eval_outputs(fct(n,0).shape)
assert v==(3)
v = eval_outputs(fct(n,1).shape)
assert v==(2)
v = eval_outputs(fct(n,[0,1]).shape)
assert v.size==0
def test2_invalid(self): def test2_invalid(self):
for fct in [max,min]: for fct in [max, min]:
n = as_tensor_variable(numpy.random.rand(2,3)) n = as_tensor_variable(numpy.random.rand(2, 3))
# Silence expected error messages # Silence expected error messages
_logger = logging.getLogger('theano.gof.opt') _logger = logging.getLogger('theano.gof.opt')
oldlevel = _logger.level oldlevel = _logger.level
_logger.setLevel(logging.CRITICAL) _logger.setLevel(logging.CRITICAL)
try: try:
try: try:
eval_outputs(fct(n,3)) eval_outputs(fct(n, 3))
assert False assert False
except ValueError, e: except ValueError, e:
pass pass
finally: finally:
_logger.setLevel(oldlevel) _logger.setLevel(oldlevel)
def test2_invalid_neg(self): def test2_invalid_neg(self):
for fct in [max,min]: for fct in [max, min]:
n = as_tensor_variable(numpy.random.rand(2,3)) n = as_tensor_variable(numpy.random.rand(2, 3))
old_stderr = sys.stderr old_stderr = sys.stderr
sys.stderr = StringIO.StringIO() sys.stderr = StringIO.StringIO()
try: try:
try: try:
eval_outputs(fct(n,-3)) eval_outputs(fct(n, -3))
assert False assert False
except ValueError, e: except ValueError, e:
pass pass
finally: finally:
sys.stderr = old_stderr sys.stderr = old_stderr
def test2_valid_neg(self): def test2_valid_neg(self):
for fct,nfct in [(max,numpy.max),(min,numpy.min)]: for fct, nfct in [(max, numpy.max), (min, numpy.min)]:
n = as_tensor_variable(numpy.random.rand(2,3)) n = as_tensor_variable(numpy.random.rand(2, 3))
v = eval_outputs(fct(n,-1)) v = eval_outputs(fct(n, -1))
self.assertTrue(v.shape == (2,)) self.assertTrue(v.shape == (2,))
self.assertTrue(numpy.all(v == nfct(n.value,-1))) self.assertTrue(numpy.all(v == nfct(n.value, -1)))
v = eval_outputs(fct(n,-2)) v = eval_outputs(fct(n, -2))
self.assertTrue(v.shape == (3,)) self.assertTrue(v.shape == (3,))
self.assertTrue(numpy.all(v == nfct(n.value,-2))) self.assertTrue(numpy.all(v == nfct(n.value, -2)))
v = eval_outputs(fct(n,-1).shape) v = eval_outputs(fct(n, -1).shape)
assert v==(2) assert v == (2)
v = eval_outputs(fct(n,-2).shape) v = eval_outputs(fct(n, -2).shape)
assert v==(3) assert v == (3)
def test3(self): def test3(self):
for fct,nfct in [(max,numpy.max),(min,numpy.min)]: # Test with 1 axis or all axis out of 3 dims
n = as_tensor_variable(numpy.random.rand(2,3,4)) data = numpy.random.rand(2, 3, 4)
v = eval_outputs(fct(n,0)) n = as_tensor_variable(data)
self.assertTrue(v.shape == (3,4)) for fct, nfct in [(max, numpy.max), (min, numpy.min)]:
self.assertTrue(numpy.all(v == nfct(n.value,0))) for (axis, np_axis) in [(-1, -1), (0, 0), (1, 1), (2, 2),
v = eval_outputs(fct(n,1)) (None, None), ([0, 1, 2], None),
self.assertTrue(v.shape == (2,4)) ([1, 0, 2], None)]:
self.assertTrue(numpy.all(v == nfct(n.value,1))) v = eval_outputs(fct(n, axis))
v = eval_outputs(fct(n,2)) self.assertTrue(numpy.all(v == nfct(data, np_axis)))
self.assertTrue(v.shape == (2,3)) v_shape = eval_outputs(fct(n, axis).shape)
self.assertTrue(numpy.all(v == nfct(n.value,2))) assert tuple(v_shape) == nfct(data, np_axis).shape
v = eval_outputs(fct(n,[0,1]))
self.assertTrue(v.shape == (4,)) def test3b(self):
self.assertTrue(numpy.all(v == nfct(nfct(n.value,1),0))) # Test with 2 axis out of 3 dims
v = eval_outputs(fct(n,[0,2])) data = numpy.random.rand(2, 3, 4)
self.assertTrue(v.shape == (3,)) n = as_tensor_variable(data)
self.assertTrue(numpy.all(v == nfct(nfct(n.value,2),0))) for fct, nfct in [(max, numpy.max), (min, numpy.min)]:
v = eval_outputs(fct(n,[1,2])) for axis in [[0, 1], [1, 2], [0, 2]]:
self.assertTrue(v.shape == (2,)) v = eval_outputs(fct(n, axis))
self.assertTrue(numpy.all(v == nfct(nfct(n.value,2),1))) np_v = nfct(nfct(data, axis[1]), axis[0])
v = eval_outputs(fct(n,[0,1,2])) self.assertTrue(numpy.all(v == np_v))
self.assertTrue(v.shape == ()) v_shape = eval_outputs(fct(n, axis).shape)
assert tuple(v_shape) == np_v.shape
v = eval_outputs(fct(n,0).shape)
assert tuple(v)==(3,4)
v = eval_outputs(fct(n,1).shape)
assert tuple(v)==(2,4)
v = eval_outputs(fct(n,2).shape)
assert tuple(v)==(2,3)
v = eval_outputs(fct(n,[0,1]).shape)
self.assertTrue(v == (4,))
v = eval_outputs(fct(n,[0,2]).shape)
self.assertTrue(v == (3,))
v = eval_outputs(fct(n,[1,2]).shape)
self.assertTrue(v == (2,))
v = eval_outputs(fct(n,[0,1,2]).shape)
self.assertTrue(v.size == 0)
def test_grad_max(self): def test_grad_max(self):
data = numpy.random.rand(2,3) data = numpy.random.rand(2, 3)
n = as_tensor_variable(data) n = as_tensor_variable(data)
def check_grad_max(data, max_grad_data, axis=None): def check_grad_max(data, max_grad_data, axis=None):
#This work only for axis in [0,None] #This work only for axis in [0,None]
assert axis in [0,None] assert axis in [0, None]
z = numpy.zeros_like(data) z = numpy.zeros_like(data)
z = z.flatten() z = z.flatten()
argmax=numpy.argmax(data,axis=axis) argmax = numpy.argmax(data, axis=axis)
if argmax.ndim==0: if argmax.ndim == 0:
z[numpy.argmax(data,axis=axis)]+=1 z[numpy.argmax(data, axis=axis)] += 1
else: else:
for id,v in enumerate(argmax): for id, v in enumerate(argmax):
z[v*numpy.prod(data.shape[data.ndim-1:axis:-1])+id]+=1 z[v * numpy.prod(data.shape[data.ndim - 1:axis:-1])
+ id] += 1
z = z.reshape(data.shape) z = z.reshape(data.shape)
assert numpy.all(max_grad_data == z) assert numpy.all(max_grad_data == z)
#test grad of max #test grad of max
#axis is the last one #axis is the last one
utt.verify_grad(lambda v: max(v,axis=-1), [data]) utt.verify_grad(lambda v: max(v, axis=-1), [data])
utt.verify_grad(lambda v: max(v,axis=[0]), [data]) utt.verify_grad(lambda v: max(v, axis=[0]), [data])
check_grad_max(data,eval_outputs(grad(max(n,axis=0).sum(),n)),axis=0) check_grad_max(data, eval_outputs(grad(max(n, axis=0).sum(), n)),
axis=0)
utt.verify_grad(lambda v: max(v,axis=[1]), [data]) utt.verify_grad(lambda v: max(v, axis=[1]), [data])
#check_grad_max(data,eval_outputs(grad(max(n,axis=1),n)),axis=1) #check_grad_max(data,eval_outputs(grad(max(n,axis=1),n)),axis=1)
utt.verify_grad(lambda v: max(v.flatten()), [data]) utt.verify_grad(lambda v: max(v.flatten()), [data])
check_grad_max(data,eval_outputs(grad(max(n.flatten()),n))) check_grad_max(data, eval_outputs(grad(max(n.flatten()), n)))
def test_grad_min(self): def test_grad_min(self):
data = numpy.random.rand(2,3) data = numpy.random.rand(2, 3)
n = as_tensor_variable(data) n = as_tensor_variable(data)
def check_grad_min(data, min_grad_data, axis=None): def check_grad_min(data, min_grad_data, axis=None):
#This work only for axis in [0,None] #This work only for axis in [0, None]
assert axis in [0,None] assert axis in [0, None]
z = numpy.zeros_like(data) z = numpy.zeros_like(data)
z = z.flatten() z = z.flatten()
argmin=numpy.argmin(data,axis=axis) argmin = numpy.argmin(data, axis=axis)
if argmin.ndim==0: if argmin.ndim == 0:
z[numpy.argmin(data,axis=axis)]+=1 z[numpy.argmin(data, axis=axis)] += 1
else: else:
for id,v in enumerate(argmin): for id, v in enumerate(argmin):
z[v*numpy.prod(data.shape[data.ndim-1:axis:-1])+id]+=1 z[v * numpy.prod(data.shape[data.ndim - 1:axis:-1])
+ id] += 1
z = z.reshape(data.shape) z = z.reshape(data.shape)
assert numpy.all(min_grad_data == z) assert numpy.all(min_grad_data == z)
#test grad of min #test grad of min
#axis is the last one #axis is the last one
utt.verify_grad(lambda v: min(v,axis=-1), [data]) utt.verify_grad(lambda v: min(v, axis=-1), [data])
utt.verify_grad(lambda v: min(v,axis=[0]), [data]) utt.verify_grad(lambda v: min(v, axis=[0]), [data])
check_grad_min(data,eval_outputs(grad(min(n,axis=0).sum(),n)),axis=0) check_grad_min(data, eval_outputs(grad(min(n, axis=0).sum(), n)),
axis=0)
utt.verify_grad(lambda v: min(v,axis=[1]), [data]) utt.verify_grad(lambda v: min(v, axis=[1]), [data])
#check_grad_min(data,eval_outputs(grad(min(n,axis=1),n)),axis=1) #check_grad_min(data,eval_outputs(grad(min(n,axis=1),n)),axis=1)
utt.verify_grad(lambda v: min(v.flatten()), [data]) utt.verify_grad(lambda v: min(v.flatten()), [data])
check_grad_min(data,eval_outputs(grad(min(n.flatten()),n))) check_grad_min(data, eval_outputs(grad(min(n.flatten()), n)))
def _grad_list(self): def _grad_list(self):
""" """
Test the gradient when we have multiple axis at the same time. Test the gradient when we have multiple axis at the same time.
This not implemented, so we disable the test. See ticket: http://trac-hg.assembla.com/theano/ticket/511 This not implemented, so we disable the test. See ticket:
http://trac-hg.assembla.com/theano/ticket/511
""" """
data = numpy.random.rand(2,3) data = numpy.random.rand(2, 3)
n = as_tensor_variable(data) n = as_tensor_variable(data)
for fct in [max_and_argmax,max,min]: for fct in [max_and_argmax, max, min]:
utt.verify_grad(lambda v: fct(v,axis=[0,1]), [data]) utt.verify_grad(lambda v: fct(v, axis=[0, 1]), [data])
#check_grad_max(data,eval_outputs(grad(max_and_argmax(n,axis=1)[0],n)),axis=1) #check_grad_max(data, eval_outputs(grad(max_and_argmax(n,
#axis=1)[0], n)),axis=1)
class T_subtensor(unittest.TestCase): class T_subtensor(unittest.TestCase):
""" """
This is build in a way that allow to reuse it to test the equivalent gpu op. This is build in a way that allow to reuse it to test the
equivalent gpu op.
""" """
def __init__(self, name, shared=_shared, def __init__(self, name, shared=_shared,
sub=tensor.Subtensor, sub=tensor.Subtensor,
......
theano/tensor/tests/test_opt_uncanonicalize.py
浏览文件 @ 90c3833e
...@@ -13,91 +13,95 @@ from theano.tests import unittest_tools as utt ...@@ -13,91 +13,95 @@ from theano.tests import unittest_tools as utt
class T_max_and_argmax(unittest.TestCase): class T_max_and_argmax(unittest.TestCase):
def test_optimization(self): def test_optimization(self):
#If we use only the max output, we should replace this op with a faster one. #If we use only the max output, we should replace this op with
mode = theano.compile.mode.get_default_mode().including('canonicalize','fast_run') #a faster one.
mode = theano.compile.mode.get_default_mode().including(
'canonicalize', 'fast_run')
data = numpy.asarray(numpy.random.rand(2,3),dtype=config.floatX) for axis in [0, 1, -1]:
n = tensor.matrix() data = numpy.asarray(numpy.random.rand(2, 3), dtype=config.floatX)
n = tensor.matrix()
f = function([n], tensor.max_and_argmax(n,0)[0], mode=mode) f = function([n], tensor.max_and_argmax(n, axis)[0], mode=mode)
topo = f.maker.env.toposort() topo = f.maker.env.toposort()
assert len(topo)==1 assert len(topo) == 1
assert isinstance(topo[0].op, CAReduce) assert isinstance(topo[0].op, CAReduce)
f = function([n], tensor.max_and_argmax(n,0), mode=mode) f = function([n], tensor.max_and_argmax(n, axis), mode=mode)
topo = f.maker.env.toposort() topo = f.maker.env.toposort()
assert len(topo)==1 assert len(topo) == 1
assert isinstance(topo[0].op, tensor.MaxAndArgmax) assert isinstance(topo[0].op, tensor.MaxAndArgmax)
class T_min_max(unittest.TestCase): class T_min_max(unittest.TestCase):
def setUp(self): def setUp(self):
utt.seed_rng() utt.seed_rng()
self.mode = theano.compile.mode.get_default_mode().including('canonicalize','fast_run') self.mode = theano.compile.mode.get_default_mode().including(
'canonicalize', 'fast_run')
def test_optimization_max(self): def test_optimization_max(self):
data = numpy.asarray(numpy.random.rand(2,3),dtype=config.floatX) data = numpy.asarray(numpy.random.rand(2, 3), dtype=config.floatX)
n = tensor.matrix() n = tensor.matrix()
f = function([n],tensor.max(n,0), mode=self.mode) for axis in [0, 1, -1]:
topo = f.maker.env.toposort() f = function([n], tensor.max(n, axis), mode=self.mode)
assert len(topo)==1 topo = f.maker.env.toposort()
assert isinstance(topo[0].op,CAReduce) assert len(topo) == 1
f(data) assert isinstance(topo[0].op, CAReduce)
f(data)
f = function([n],tensor.max(-n,0), mode=self.mode) f = function([n], tensor.max(-n, axis), mode=self.mode)
topo = f.maker.env.toposort() topo = f.maker.env.toposort()
assert len(topo)==2 assert len(topo) == 2
assert isinstance(topo[0].op, Elemwise) assert isinstance(topo[0].op, Elemwise)
assert isinstance(topo[0].op.scalar_op, scalar.Neg) assert isinstance(topo[0].op.scalar_op, scalar.Neg)
assert isinstance(topo[1].op,CAReduce) assert isinstance(topo[1].op, CAReduce)
f(data) f(data)
f = function([n],-tensor.max(n,0), mode=self.mode) f = function([n], -tensor.max(n, axis), mode=self.mode)
topo = f.maker.env.toposort() topo = f.maker.env.toposort()
assert len(topo)==2 assert len(topo) == 2
assert isinstance(topo[0].op,CAReduce) assert isinstance(topo[0].op, CAReduce)
assert isinstance(topo[1].op, Elemwise) assert isinstance(topo[1].op, Elemwise)
assert isinstance(topo[1].op.scalar_op, scalar.Neg) assert isinstance(topo[1].op.scalar_op, scalar.Neg)
f(data) f(data)
f = function([n],-tensor.max(-n,0), mode=self.mode) f = function([n], -tensor.max(-n, axis), mode=self.mode)
topo = f.maker.env.toposort() topo = f.maker.env.toposort()
assert len(topo)==1 assert len(topo) == 1
assert isinstance(topo[0].op,CAReduce)#min assert isinstance(topo[0].op, CAReduce) # min
f(data) f(data)
def test_optimization_min(self): def test_optimization_min(self):
data = numpy.asarray(numpy.random.rand(2,3),dtype=config.floatX) data = numpy.asarray(numpy.random.rand(2, 3), dtype=config.floatX)
n = tensor.matrix() n = tensor.matrix()
f = function([n],tensor.min(n,0), mode=self.mode) for axis in [0, 1, -1]:
topo = f.maker.env.toposort() f = function([n], tensor.min(n, axis), mode=self.mode)
assert len(topo)==1 topo = f.maker.env.toposort()
assert isinstance(topo[0].op,CAReduce) assert len(topo) == 1
f(data) assert isinstance(topo[0].op, CAReduce)
f(data)
#test variant with neg to make sure we optimize correctly
f = function([n],tensor.min(-n,0), mode=self.mode) #test variant with neg to make sure we optimize correctly
topo = f.maker.env.toposort() f = function([n], tensor.min(-n, axis), mode=self.mode)
assert len(topo)==2 topo = f.maker.env.toposort()
assert isinstance(topo[0].op,CAReduce)#max assert len(topo) == 2
assert isinstance(topo[1].op, Elemwise) assert isinstance(topo[0].op, CAReduce) # max
assert isinstance(topo[1].op.scalar_op, scalar.Neg) assert isinstance(topo[1].op, Elemwise)
f(data) assert isinstance(topo[1].op.scalar_op, scalar.Neg)
f(data)
f = function([n],-tensor.min(n,0), mode=self.mode)
topo = f.maker.env.toposort() f = function([n], -tensor.min(n, axis), mode=self.mode)
assert len(topo)==2 topo = f.maker.env.toposort()
assert isinstance(topo[0].op, Elemwise) assert len(topo) == 2
assert isinstance(topo[0].op.scalar_op, scalar.Neg) assert isinstance(topo[0].op, Elemwise)
assert isinstance(topo[1].op,CAReduce)#max assert isinstance(topo[0].op.scalar_op, scalar.Neg)
f(data) assert isinstance(topo[1].op, CAReduce) # max
f(data)
f = function([n],-tensor.min(-n,0), mode=self.mode)
topo = f.maker.env.toposort() f = function([n], -tensor.min(-n, axis), mode=self.mode)
assert len(topo)==1 topo = f.maker.env.toposort()
assert isinstance(topo[0].op,CAReduce)#max assert len(topo) == 1
f(data) assert isinstance(topo[0].op, CAReduce) # max
f(data)
theano/tensor/tests/test_sharedvar.py
浏览文件 @ 90c3833e
...@@ -350,27 +350,6 @@ def makeSharedTester(shared_constructor_, ...@@ -350,27 +350,6 @@ def makeSharedTester(shared_constructor_,
assert may_share_memory(old_data, x_shared.container.storage[0]) assert may_share_memory(old_data, x_shared.container.storage[0])
x_shared.get_value(borrow=True) x_shared.get_value(borrow=True)
# Test by .value
# As we know that .value is deprecated, we filter out the warning
warnings.filterwarnings(
action='ignore',
message='The .value property of shared variables is deprecated.'
)
try:
nd += 1
old_data = x_shared.container.storage[0]
x_shared.value = nd
assert numpy.allclose(self.ref_fct(x_shared.value), self.ref_fct(self.cast_value(nd)))
assert may_share_memory(old_data, x_shared.container.storage[0]) == self.set_value_inplace
finally:
# Restore the default behavior.
# TODO There is a cleaner way to do this in Python 2.6, once
# Theano drops support of Python 2.4 and 2.5.
warnings.filterwarnings(
action='default',
message='The .value property of shared variables is deprecated.'
)
# Test by set_value with borrow=False # Test by set_value with borrow=False
nd += 1 nd += 1
old_data = x_shared.container.storage[0] old_data = x_shared.container.storage[0]
......
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