Skip to content

P
pytensor
提交 29d3f9e0 authored 作者: lamblin's avatar lamblin
浏览文件
操作

Merge pull request #425 from delallea/improved_set_subtensor

Fixed issues with advanced inc/set subtensor in some cases
上级 0f86ecd9 3a0b3dfb
隐藏空白字符变更
内嵌 并排
正在显示 包含 182 行增加 和 64 行删除
theano/tensor/basic.py
浏览文件 @ 29d3f9e0
......@@ -5097,6 +5097,7 @@ class AdvancedSubtensor1(Op):
def __hash__(self):
return hash(type(self))
def __eq__(self, other):
return type(self) == type(other)
......@@ -5115,7 +5116,7 @@ class AdvancedSubtensor1(Op):
x, i = inp
out, = out_
# Copy always implied by numpy advanced indexing semantic.
if out[0] is not None and out[0].shape==(len(i),)+x.shape[1:]:
if out[0] is not None and out[0].shape == (len(i),) + x.shape[1:]:
o = out[0]
else:
o = None
......@@ -5131,8 +5132,9 @@ class AdvancedSubtensor1(Op):
def grad(self, inputs, grads):
gz, = grads
assert len(inputs)==2
return [advanced_inc_subtensor1(zeros_like(inputs[0]),gz,inputs[1])]+[None]*(len(inputs)-1)
assert len(inputs) == 2
rval1 = [advanced_inc_subtensor1(zeros_like(inputs[0]), gz, inputs[1])]
return rval1 + [None] * (len(inputs) - 1)
def R_op(self, inputs, eval_points):
if eval_points[0] is None:
......@@ -5141,10 +5143,11 @@ class AdvancedSubtensor1(Op):
def infer_shape(self, node, ishapes):
x, ilist = ishapes
return [ilist+x[1:]]
return [ilist + x[1:]]
advanced_subtensor1 = AdvancedSubtensor1()
class AdvancedIncSubtensor1(Op):
"""Increments a subtensor using advanced slicing (list of index)"""
def __init__(self, inplace=False, set_instead_of_inc=False):
......@@ -5173,10 +5176,13 @@ class AdvancedIncSubtensor1(Op):
if x_.type.ndim == 0:
raise TypeError('cannot index into a scalar')
if y_.type.ndim > x_.type.ndim:
opname = 'increment'
if self.set_instead_of_inc:
opname = 'set'
else:
opname = 'increment'
raise TypeError('cannot %s x subtensor with ndim=%s'
' by y with ndim=%s to x subtensor with ndim=%s '%(
opname, x_.type.ndim, y_.type.ndim ))
' by y with ndim=%s to x subtensor with ndim=%s ' % (
opname, x_.type.ndim, y_.type.ndim))
return Apply(self, [x_, y_, ilist_], [x_.type()])
......@@ -5186,19 +5192,19 @@ class AdvancedIncSubtensor1(Op):
out, = out_
if not self.inplace:
x = x.copy()
# x[idx] += y don't work if the same index is present many times.
# It do it only once
# -- Numpy also behaves this way, is it a bug in numpy?
# In Numpy, x[idx] += y doesn't work if the same index is present
# many times: it does it only once. Is it a bug? In any case, for
# this reason we implement our own 'inc' iteration.
if self.set_instead_of_inc:
if y.ndim:
for (j,i) in enumerate(idx):
x[i] = y[j]
else:
for i in idx:
x[i] = y
x[idx] = y
else:
if y.ndim:
for (j,i) in enumerate(idx):
# If `y` has as many dimensions as `x`, then we want to iterate
# jointly on `x` and `y`. Otherwise, it means `y` should be
# broadcasted to fill all relevant rows of `x`.
assert y.ndim <= x.ndim # Should be guaranteed by `make_node`
if y.ndim == x.ndim:
assert len(y) == len(idx)
for (j, i) in enumerate(idx):
x[i] += y[j]
else:
for i in idx:
......@@ -5215,7 +5221,6 @@ class AdvancedIncSubtensor1(Op):
return self.make_node(eval_points[0], eval_points[1],
*inputs[2:]).outputs
def grad(self, inputs, grads):
g_output, = grads
x, y = inputs[:2]
......@@ -5228,6 +5233,7 @@ class AdvancedIncSubtensor1(Op):
advanced_inc_subtensor1 = AdvancedIncSubtensor1()
class AdvancedSubtensor(Op):
"""Return a subtensor copy, using advanced indexing.
"""
......@@ -5235,10 +5241,10 @@ class AdvancedSubtensor(Op):
# AdvancedSubtensor(args)(self, *args),
# if args contains and advanced indexing pattern
def __init__(self, args): #idx_list?
def __init__(self, args): # idx_list?
# For the moment, __init__ will be passed the whole list of arguments
#TODO: see what's the best solution
self.args = args #?
self.args = args # ?
#FIXME: do not store variables in the class instance
......@@ -5590,6 +5596,7 @@ class TensorDotGrad(Op):
tensordot_grad = TensorDotGrad
class TensorDot(Op):
"""Compute tensor-tensor products over the given axes.
See numpy documentation for details.
......@@ -5600,21 +5607,23 @@ class TensorDot(Op):
@classmethod
def parse_axes(cls, axes):
if not numpy.isscalar(axes) and len(axes)!=2:
raise ValueError("Axes should be scalar valued or a list/tuple of len 2.")
if not numpy.isscalar(axes) and len(axes) != 2:
raise ValueError("Axes should be scalar valued or a list/tuple of "
"len 2.")
if isinstance(axes,(list,tuple)):
if isinstance(axes, (list, tuple)):
axes_out = []
# cast axes[0] and axes[1] to tuples
for i,a in enumerate(axes):
for i, a in enumerate(axes):
if numpy.isscalar(a):
axes_out.append((a,))
else:
axes_out.append(tuple(a))
# these should be of same length
if len(axes_out[0])!=len(axes_out[1]):
raise ValueError("Elements of the axes list/tuple need to be of the same size.")
if len(axes_out[0]) != len(axes_out[1]):
raise ValueError("Elements of the axes list/tuple need to be "
"of the same size.")
axes = tuple(axes_out)
......@@ -5631,22 +5640,23 @@ class TensorDot(Op):
def make_node(self, x, y):
op = self
if isinstance(self.axes,int):
axes = [range(x.ndim-self.axes,x.ndim),range(self.axes)]
if isinstance(self.axes, int):
axes = [range(x.ndim - self.axes, x.ndim), range(self.axes)]
op = TensorDot(axes)
axesdim = numpy.size(op.axes)/2
axesdim = numpy.size(op.axes) / 2
x, y = map(as_tensor_variable, [x, y])
if axesdim > x.type.ndim or axesdim > y.type.ndim:
raise TypeError('Cannot sum over more dimensions than input. %i > %i,%i' %
axesdim, x.type.ndim, y.type.ndim)
raise TypeError('Cannot sum over more dimensions than input. '
'%i > %i,%i' %
(axesdim, x.type.ndim, y.type.ndim))
outdim = x.type.ndim + y.type.ndim - 2*axesdim
outdim = x.type.ndim + y.type.ndim - 2 * axesdim
output = tensor(dtype=scal.upcast(x.dtype, y.dtype),
broadcastable=[False]*outdim);
return Apply(op, inputs=[x,y], outputs=[output,])
broadcastable=[False] * outdim)
return Apply(op, inputs=[x, y], outputs=[output, ])
def perform(self, node, inp, out):
x, y = inp
......@@ -5654,7 +5664,8 @@ class TensorDot(Op):
try:
z[0] = numpy.asarray(numpy.tensordot(x, y, self.axes))
except ValueError, e:
# The error raised by numpy has no shape information, we mean to add that
# The error raised by numpy has no shape information, we mean to
# add that.
e.args = e.args + (x.shape, y.shape, self.axes)
raise
......@@ -5667,13 +5678,15 @@ class TensorDot(Op):
def __str__(self):
return "tensordot"
def tensordot(x, y=None, axes=2):
if y==None:
raise NotImplementedError('The interface to tensordot has changed from '\
'tensor.tensordot(axes)(x,y) to tensor.tensordot(x,y,axes). Please '\
'modify your code accordingly.')
if y is None:
raise NotImplementedError(
'The interface to tensordot has changed from '
'tensor.tensordot(axes)(x,y) to tensor.tensordot(x,y,axes). '
'Please modify your code accordingly.')
if x.ndim==0 or y.ndim==0:
if x.ndim == 0 or y.ndim == 0:
raise ValueError('Cannot perform tensordot of 0-d inputs.')
axes = TensorDot.parse_axes(axes)
......@@ -5682,16 +5695,16 @@ def tensordot(x, y=None, axes=2):
if numpy.isscalar(axes):
if axes >= x.ndim or axes >= y.ndim:
raise ValueError('axes should be smaller than the dimension of '\
'x and y (x.ndim=%i, y.ndim=%i)' % (x.ndim,y.ndim))
elif isinstance(axes, (list,tuple)):
'x and y (x.ndim=%i, y.ndim=%i)' % (x.ndim, y.ndim))
elif isinstance(axes, (list, tuple)):
if isinstance(axes[0],(list,tuple)) and \
if isinstance(axes[0], (list, tuple)) and \
(len(axes[0]) > x.ndim or (numpy.array(axes[0]) >= x.ndim).any()):
raise ValueError('axes[0] should be array_like, of length smaller'\
' than the dimension of x (x.ndim=%i, len(axes[0])=%i).' %
(x.ndim, len(axes[0])))
if isinstance(axes[1],(list,tuple)) and \
if isinstance(axes[1], (list, tuple)) and \
(len(axes[1]) > y.ndim or (numpy.array(axes[1]) >= y.ndim).any()):
raise ValueError('axes[1] should be array_like, of length smaller'\
'than the dimension of y (y.ndim=%i, len(axes[1])=%i).' %
......
theano/tensor/tests/test_basic.py
浏览文件 @ 29d3f9e0
......@@ -2049,6 +2049,7 @@ class T_subtensor(unittest.TestCase):
raise
finally:
_logger.setLevel(oldlevel)
def test1_err_subslice(self):
n = self.shared(numpy.ones(3, dtype=self.dtype))
try:
......@@ -2121,6 +2122,7 @@ class T_subtensor(unittest.TestCase):
tval = f()
self.assertTrue(tval.shape == ())
self.assertTrue(tval == 5.0)
def test1_ok_range_infinite(self):
#Subtensor.debug = True
n = self.shared(numpy.ones(3, dtype=self.dtype)*5)
......@@ -2185,6 +2187,7 @@ class T_subtensor(unittest.TestCase):
raise
finally:
sys.stderr = old_stderr
def test2_ok_elem(self):
n = self.shared(numpy.asarray(range(6), dtype=self.dtype).reshape((2,3)))
t = n[0,2]
......@@ -2192,6 +2195,7 @@ class T_subtensor(unittest.TestCase):
tval = self.eval_output_and_check(t)
self.assertTrue(tval.shape == ())
self.assertTrue(numpy.all(tval == 2))
def test2_ok_row(self):
n = self.shared(numpy.asarray(range(6), dtype=self.dtype).reshape((2,3)))
t = n[1]
......@@ -2404,7 +2408,6 @@ class T_subtensor(unittest.TestCase):
assert numpy.all(
f(start,stop,step) == v_data[start:stop:step].shape)
def test_slice_canonical_form_0(self):
start = tensor.iscalar('b')
stop = tensor.iscalar('e')
......@@ -2428,7 +2431,6 @@ class T_subtensor(unittest.TestCase):
assert numpy.all(t_out == v_out)
assert numpy.all(t_out.shape == v_out.shape)
def test_slice_canonical_form_1(self):
stop = tensor.iscalar('e')
step = tensor.iscalar('s')
......@@ -2672,7 +2674,109 @@ class T_subtensor(unittest.TestCase):
#single element
utt.verify_grad(
inc_slice(2, 1),
(numpy.asarray([[0, 1],[2, 3],[4, 5.]]), numpy.asarray(9.),))
(numpy.asarray([[0, 1], [2, 3], [4, 5.]]), numpy.asarray(9.),))
def test_advanced_inc_and_set(self):
"""
Test advanced increment and set.
"""
rng = numpy.random.RandomState(seed=utt.fetch_seed())
all_inputs_var = []
all_inputs_num = []
all_outputs_var = []
all_outputs_num = []
for set_instead_of_inc in (False, True):
for inplace in (False, True):
for data_shape in ((10,), (4, 5), (1, 2, 3), (4, 5, 6, 7)):
data_n_dims = len(data_shape)
# Symbolic variable to be incremented.
data_var = tensor.tensor(
broadcastable=[False] * data_n_dims,
dtype=self.dtype)
data_size = numpy.product(data_shape)
# Corresponding numeric variable.
data_num_init = numpy.arange(data_size, dtype=self.dtype)
data_num_init = data_num_init.reshape(data_shape)
inc_shapes = [data_shape[i:]
for i in xrange(0, len(data_shape) + 1)]
for inc_shape in inc_shapes:
inc_n_dims = len(inc_shape)
# We copy the numeric value to be 100% sure there is no
# risk of accidentally sharing it.
data_num = data_num_init.copy()
if inplace:
# We need to copy `data_var` as we do not want
# multiple in-place operations on it.
data_var = deepcopy(data_var)
# Symbolic variable with rows to be incremented.
idx_var = theano.tensor.vector(dtype='int64')
n_to_inc = rng.randint(data_shape[0])
# Corresponding numeric variable.
idx_num = rng.randint(0, data_shape[0], n_to_inc)
idx_num = idx_num.astype('int64')
# Symbolic variable with increment value.
inc_var = tensor.tensor(
broadcastable=[False] * inc_n_dims,
dtype=self.dtype)
# Trick for the case where `inc_shape` is the same as
# `data_shape`: what we actually want is the first
# shape element to be equal to the number of rows to
# increment.
if len(inc_shape) == len(data_shape):
inc_shape = (n_to_inc,) + inc_shape[1:]
inc_size = numpy.product(inc_shape)
# Corresponding numeric variable.
inc_num = rng.uniform(size=inc_size).astype(self.dtype)
inc_num = inc_num.reshape(inc_shape)
# Result of the incrementation.
# (i) Theano
if set_instead_of_inc:
op = set_subtensor
else:
op = inc_subtensor
output = op(data_var[idx_var], inc_var,
inplace=inplace)
# (ii) Numpy (note that Numpy increments only once
# duplicated indices, so we cannot directly use +=).
data_copy = data_num.copy()
for j, idx in enumerate(idx_num):
if len(inc_shape) == len(data_shape):
# Special case where there is no broadcasting.
if set_instead_of_inc:
data_copy[idx] = inc_num[j]
else:
data_copy[idx] += inc_num[j]
else:
if set_instead_of_inc:
data_copy[idx] = inc_num
else:
data_copy[idx] += inc_num
# Remember data for the Theano function (see below).
all_inputs_var += [data_var, idx_var, inc_var]
all_inputs_num += [data_num, idx_num, inc_num]
all_outputs_var.append(output)
all_outputs_num.append(data_copy)
if False: # Enable for debugging purpose.
f = theano.function([data_var, idx_var, inc_var],
output, accept_inplace=inplace)
if inplace:
# Ensure calling `f` will not alter `data_num`.
data_num = data_num.copy()
f_out = f(data_num.copy(), idx_num, inc_num)
assert numpy.allclose(f_out, data_copy)
if not inplace:
# Sanity check: `data_num` should be intact.
assert (data_num == data_num_init).all()
# Actual test (we compile a single Theano function to make it faster).
f = theano.function(all_inputs_var, all_outputs_var,
accept_inplace=True)
f_outs = f(*all_inputs_num)
assert len(f_outs) == len(all_outputs_num)
for f_out, output_num in izip(f_outs, all_outputs_num):
# NB: if this assert fails, it will probably be easier to debug if
# you enable the debug code above.
assert numpy.allclose(f_out, output_num)
class TestIncSubtensor1(unittest.TestCase):
......@@ -5151,7 +5255,7 @@ class test_broadcast(unittest.TestCase):
def test_len():
for shape in [(5,), (3, 4), (7, 4, 6)]:
x = tensor.tensor(dtype='floatX', broadcastable=(False,)*len(shape))
x = tensor.tensor(dtype='floatX', broadcastable=(False,) * len(shape))
try:
len(x)
assert False, "Expected an error"
......@@ -5166,12 +5270,12 @@ def test_mod():
as Python. That is what we want.
"""
x, y = fscalars('xy')
fn = gof.DualLinker().accept(gof.Env([x,y], [x%y])).make_function()
for a,b in ((0,1), (1,1), (0,-1), (1,-1), (-1,-1),
(1,2), (-1,2), (1,-2), (-1,-2),
(5,3), (-5,3), (5,-3), (-5,-3)
fn = gof.DualLinker().accept(gof.Env([x, y], [x % y])).make_function()
for a, b in ((0, 1), (1, 1), (0, -1), (1, -1), (-1, -1),
(1, 2), (-1, 2), (1, -2), (-1, -2),
(5, 3), (-5, 3), (5, -3), (-5, -3)
):
assert fn(a,b) == a%b, (a,)
assert fn(a, b) == a % b, (a,)
def test_mod_compile():
......@@ -5195,14 +5299,14 @@ def test_mod_compile():
shape = x.shape
out = tensor.switch(tensor.eq(3 % x.shape[0], 0), y, y[:-1])
f = theano.function([x,y],out)
f = theano.function([x, y], out)
def test_unalign():
if config.floatX == 'float64':
dtype="b1,f8"
dtype = "b1,f8"
else:
dtype="b1,f4"
dtype = "b1,f4"
a = numpy.empty(1e4, dtype=dtype)['f1']
b = numpy.empty(1e4, dtype=dtype)['f1']
......@@ -5210,24 +5314,25 @@ def test_unalign():
assert not b.flags.aligned
a[:] = rand(len(a))
b[:] = rand(len(b))
out_numpy = 2*a + 3*b
out_numpy = 2 * a + 3 * b
av,bv = tensor.vectors('ab')
f = theano.function([av,bv],2*av+3*bv)
av, bv = tensor.vectors('ab')
f = theano.function([av, bv], 2 * av + 3 * bv)
f.maker.env.toposort()
# FAST_COMPILE use the python code that support unaligned data
# The DebugMode make a copy of the inputs, so they will be aligned.
should_raise = theano.config.mode not in ["FAST_COMPILE","DebugMode", "DEBUG_MODE"]
should_raise = theano.config.mode not in ["FAST_COMPILE", "DebugMode",
"DEBUG_MODE"]
try:
out_theano = f(a,b)
out_theano = f(a, b)
assert not a.flags.aligned
assert not b.flags.aligned
assert numpy.allclose(out_numpy,out_theano)
assert numpy.allclose(out_numpy, out_theano)
if should_raise:
raise Exception("Expected an error from Theano!")
except NotImplementedError, e:
if not should_raise:
raise Exception("Theano raised an exception when none was expected")
raise Exception("Theano raised an unexpected exception")
def test_dimshuffle_duplicate():
......
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