def local_max_and_argmax(node):
"""
If we don't use the argmax, change it to a max only.
"""
if isinstance(node.op, T.MaxAndArgmax):
axis = node.op.get_params(node)
if len(node.outputs[1].clients) == 0:
new = CAReduce(scal.maximum, axis)(node.inputs[0])
return [new, None]
if len(node.outputs[0].clients) == 0:
return [None, T._argmax(node.inputs[0], axis)]
def test_infer_shape(self):
for xsh, tosum in [((5, 6), None), ((5, 6), (0, 1)), ((5, 6), (0, )),
((5, 6), (1, )), ((5, 6), (-1, )), ((5, 6), (-2, )),
((2, 3, 4, 5), (0, 1, 3)), ((2, 3, 4, 5), (-2, -3)),
((5, 0), None), ((5, 0), (0, )), ((5, 0), (1, )),
((5, 6), ()), ((5, 0), ()), ((), None), ((), ())]:
dtype = theano.config.floatX
x = TensorType(dtype, [(entry == 1) for entry in xsh])('x')
if tosum is None:
tosum = range(len(xsh))
xv = numpy.asarray(numpy.random.rand(*xsh), dtype=dtype)
self._compile_and_check([x], [CAReduce(scalar.add, axis=tosum)(x)],
[xv], CAReduce,
["local_cut_useless_reduce"])
def local_max_and_argmax(node):
"""
If we don't use the argmax, change it to a max only.
"""
if node.op == T._max_and_argmax:
if len(node.outputs[1].clients) == 0:
#MaxAndArgmax support variable axis,
#but CAReduce support only constant axis.
try:
axis = get_scalar_constant_value(node.inputs[1])
except NotScalarConstantError:
return False
new = CAReduce(scal.maximum, axis)(node.inputs[0])
return [new, None]
def local_max_to_min(node):
"""
change -(max(-x)) to min
This is tested in tensor/tests/test_basic.py:test_min_max
:note: we don't need an opt that will do the reverse as by default
the interface put only MaxAndArgmax into the graph.
"""
if node.op == T.neg and node.inputs[0].owner:
max = node.inputs[0]
if (max.owner and
isinstance(max.owner.op, CAReduce)
and max.owner.op.scalar_op == scal.maximum):
neg = max.owner.inputs[0]
if neg.owner and neg.owner.op == T.neg:
return [CAReduce(scal.minimum,
max.owner.op.axis)(neg.owner.inputs[0])]
return False
def apply(self, fgraph):
did_something = True
while did_something:
nodelist = fgraph.toposort()
did_something = False
for node in nodelist:
if node.op == T._max_and_argmax:
if len(node.outputs[1].clients) == 0:
try:
axis = get_scalar_constant_value(node.inputs[1])
except NotScalarConstantError:
return False
new = CAReduce(scal.maximum, axis)(node.inputs[0])
try:
fgraph.replace_all_validate(
((node.outputs[0], new),),
reason=self.__class__.__name__)
did_something = True
break
except InconsistencyError, e:
pass
def local_max_and_argmax(node):
"""
If we don't use the argmax, change it to a max only.
"""
if node.op == T._max_and_argmax:
if len(node.outputs[1].clients) == 0:
# MaxAndArgmax support variable axis,
# but CAReduce support only constant axis.
if node.inputs[1].data is None:
axis = None
else:
try:
axis = get_scalar_constant_value(node.inputs[1])
except NotScalarConstantError:
axis = node.inputs[1]
if not isinstance(axis, T.TensorConstant):
return False
axis = axis.data
new = CAReduce(scal.maximum, axis)(node.inputs[0])
return [new, None]
if len(node.outputs[0].clients) == 0:
return [None, T._argmax(node.inputs[0], node.inputs[1])]
def with_linker(self, linker, scalar_op=scalar.add, dtype="floatX",
test_nan=False, tensor_op=None):
for xsh, tosum in [((5, 6), None),
((5, 6), (0, 1)),
((5, 6), (0, )),
((5, 6), (1, )),
((5, 6), (-1, )),
((5, 6), (-2, )),
((5, 6), ()),
((2, 3, 4, 5), (0, 1, 3)),
((2, 3, 4, 5), (-2, -3)),
((5, 0), None),
((5, 0), (0, )),
((5, 0), (1, )),
((5, 0), ()),
((), None),
((), ())]:
if dtype == "floatX":
dtype = theano.config.floatX
x = TensorType(dtype, [(entry == 1) for entry in xsh])('x')
if tensor_op is None:
e = CAReduce(scalar_op, axis=tosum)(x)
else:
e = tensor_op(x, axis=tosum)
if tosum is None:
tosum = range(len(xsh))
f = copy(linker).accept(FunctionGraph([x], [e])).make_function()
xv = numpy.asarray(numpy.random.rand(*xsh))
if not "int" in dtype:
xv = numpy.asarray(xv, dtype=dtype)
else:
xv = numpy.asarray(xv < 0.5, dtype=dtype)
if test_nan and xv.size > 0:
if len(xsh) > 0:
xv = xv.flatten()
xv[0] = numpy.nan
xv = xv.reshape(*xsh)
else:
xv = numpy.asarray(numpy.nan, dtype=dtype)
zv = xv
numpy_raised = False
if len(tosum) > 1 and any([a < 0 for a in tosum]):
#In that case, we need to use the good order of axis
#in the reduction.
axis2 = []
for a in tosum:
if a < 0:
axis2.append(a + len(xsh))
else:
axis2.append(a)
assert len(axis2) == len(tosum)
tosum = tuple(axis2)
if tensor_op == tensor.all:
for axis in reversed(sorted(tosum)):
zv = numpy.all(zv, axis)
if len(tosum) == 0:
zv = zv != 0
elif tensor_op == tensor.any:
for axis in reversed(sorted(tosum)):
zv = numpy.any(zv, axis)
if len(tosum) == 0:
zv = zv != 0
elif scalar_op == scalar.add:
for axis in reversed(sorted(tosum)):
zv = numpy.add.reduce(zv, axis)
elif scalar_op == scalar.mul:
for axis in reversed(sorted(tosum)):
zv = numpy.multiply.reduce(zv, axis)
elif scalar_op == scalar.maximum:
try:
for axis in reversed(sorted(tosum)):
zv = numpy.maximum.reduce(zv, axis)
except ValueError:
numpy_raised = True
elif scalar_op == scalar.minimum:
try:
for axis in reversed(sorted(tosum)):
zv = numpy.minimum.reduce(zv, axis)
except ValueError:
numpy_raised = True
elif scalar_op == scalar.or_:
for axis in reversed(sorted(tosum)):
zv = numpy.bitwise_or.reduce(zv, axis)
elif scalar_op == scalar.and_:
for axis in reversed(sorted(tosum)):
zv = numpy.bitwise_and.reduce(zv, axis)
elif scalar_op == scalar.xor:
# There is no identity value for the xor function
# So we can't support shape of dimensions 0.
if numpy.prod(zv.shape) == 0:
continue
for axis in reversed(sorted(tosum)):
zv = numpy.bitwise_xor.reduce(zv, axis)
else:
raise Exception(
"Test for CAReduce with scalar_op %s not implemented" %
str(scalar_op))
if scalar_op in [scalar.maximum, scalar.minimum] and numpy_raised:
try:
out = f(xv)
assert out.dtype == dtype
except ValueError:
pass
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:
self.assertTrue(theano.tensor.TensorType.values_eq(f(xv),
zv),
(f(xv), zv))
else:
f_xv = f(xv)
self.assertTrue((f_xv.shape == zv.shape), (f_xv, zv))
self.assertTrue(numpy.allclose(f_xv, zv), (f_xv, zv))
#test CAReduce.infer_shape
#the Shape op don't implement c_code!
if isinstance(linker, gof.PerformLinker):
x = TensorType(dtype, [(entry == 1) for entry in xsh])('x')
if tensor_op is None:
e = CAReduce(scalar_op, axis=tosum)(x)
else:
e = tensor_op(x, axis=tosum)
if tosum is None:
tosum = range(len(xsh))
f = copy(linker).accept(FunctionGraph([x],
[e.shape])).make_function()
if not(scalar_op in [scalar.maximum, scalar.minimum] and
((xsh == () or numpy.prod(xsh) == 0))):
assert all(f(xv) == zv.shape)