def setUp(self):
super(Base, self).setUp()
self.x_size = 7
self.channel_size = 3
self.z_size = 4
self.probs_size = 11
tensor = math_ops.range(
0, self.x_size * self.channel_size * self.z_size * self.probs_size)
tensor = array_ops.reshape(
tensor,
[self.x_size, self.channel_size, self.z_size, self.probs_size])
a0 = ('x', range(self.x_size))
a1 = ('channel', ['red', 'green', 'blue'])
a2 = 'z'
a3 = ('probs', np.linspace(0.0, 1.0, self.probs_size))
self.tensor = tensor
self.a0 = a0
self.a1 = a1
self.a2 = a2
self.a2_resolved = ('z', self.z_size)
self.a3 = a3
self.original_lt = core.LabeledTensor(tensor, [a0, a1, a2, a3])
self.x_probs_lt = core.slice_function(self.original_lt, {'z': 0})
self.x_probs_lt = ops.select(self.x_probs_lt, {'channel': 'red'})
self.channel_probs_lt = core.slice_function(self.original_lt, {
'x': 3,
'z': 0
})
def setUp(self):
super(Base, self).setUp()
self.x_size = 7
self.channel_size = 3
self.z_size = 4
self.probs_size = 11
tensor = math_ops.range(0, self.x_size * self.channel_size * self.z_size *
self.probs_size)
tensor = array_ops.reshape(
tensor, [self.x_size, self.channel_size, self.z_size, self.probs_size])
a0 = ('x', range(self.x_size))
a1 = ('channel', ['red', 'green', 'blue'])
a2 = 'z'
a3 = ('probs', np.linspace(0.0, 1.0, self.probs_size))
self.tensor = tensor
self.a0 = a0
self.a1 = a1
self.a2 = a2
self.a2_resolved = ('z', self.z_size)
self.a3 = a3
self.original_lt = core.LabeledTensor(tensor, [a0, a1, a2, a3])
self.x_probs_lt = core.slice_function(self.original_lt, {'z': 0})
self.x_probs_lt = ops.select(self.x_probs_lt, {'channel': 'red'})
self.channel_probs_lt = core.slice_function(self.original_lt,
{'x': 3,
'z': 0})
def setUp(self):
super(SqueezeTest, self).setUp()
self.squeezable_lt = core.slice_function(self.original_lt, {
'channel': slice(0, 1),
'probs': slice(0, 1)
})
def test_slice(self):
select_lt = core.slice_function(self.original_lt, {'channel': slice(0, 2)})
a1_sliced = ('channel', ['red', 'green'])
golden_lt = core.LabeledTensor(self.tensor[:, :2, :, :],
[self.a0, a1_sliced, self.a2, self.a3])
self.assertLabeledTensorsEqual(select_lt, golden_lt)
def test_slice_unlabeled(self):
select_lt = core.slice_function(self.original_lt, {'z': slice(1, 3)})
a2_sliced = 'z'
golden_lt = core.LabeledTensor(self.tensor[:, :, 1:3, :],
[self.a0, self.a1, a2_sliced, self.a3])
self.assertLabeledTensorsEqual(select_lt, golden_lt)
def test_slice(self):
select_lt = core.slice_function(self.original_lt, {'channel': slice(0, 2)})
a1_sliced = ('channel', ['red', 'green'])
golden_lt = core.LabeledTensor(self.tensor[:, :2, :, :],
[self.a0, a1_sliced, self.a2, self.a3])
self.assertLabeledTensorsEqual(select_lt, golden_lt)
def test_slice_unlabeled(self):
select_lt = core.slice_function(self.original_lt, {'z': slice(1, 3)})
a2_sliced = 'z'
golden_lt = core.LabeledTensor(self.tensor[:, :, 1:3, :],
[self.a0, self.a1, a2_sliced, self.a3])
self.assertLabeledTensorsEqual(select_lt, golden_lt)
def test_slices(self):
select_lt = core.slice_function(
self.original_lt, {'x': slice(1, 5),
'channel': slice(1, None)})
a0_sliced = ('x', range(1, 5))
a1_sliced = ('channel', ['green', 'blue'])
golden_lt = core.LabeledTensor(self.tensor[1:5, 1:, :, :],
[a0_sliced, a1_sliced, self.a2, self.a3])
self.assertLabeledTensorsEqual(select_lt, golden_lt)
def test_slices(self):
select_lt = core.slice_function(self.original_lt, {'x': slice(1, 5),
'channel': slice(1,
None)})
a0_sliced = ('x', range(1, 5))
a1_sliced = ('channel', ['green', 'blue'])
golden_lt = core.LabeledTensor(self.tensor[1:5, 1:, :, :],
[a0_sliced, a1_sliced, self.a2, self.a3])
self.assertLabeledTensorsEqual(select_lt, golden_lt)
def test_scalar(self):
select_lt = core.slice_function(self.original_lt, {'channel': 1})
golden_lt = core.LabeledTensor(self.tensor[:, 1, :, :],
[self.a0, self.a2, self.a3])
self.assertLabeledTensorsEqual(select_lt, golden_lt)
def test_name(self):
select_lt = core.slice_function(self.original_lt, {'channel': 1})
self.assertIn('lt_slice', select_lt.name)
def test(self):
squeeze_lt = ops.squeeze(self.squeezable_lt, ['probs'])
golden_lt = core.slice_function(self.squeezable_lt, {'probs': 0})
self.assertLabeledTensorsEqual(squeeze_lt, golden_lt)
def setUp(self):
super(SqueezeTest, self).setUp()
self.squeezable_lt = core.slice_function(
self.original_lt, {'channel': slice(0, 1),
'probs': slice(0, 1)})
def test_slice(self):
map_lt = ops.map_fn(lambda t: core.slice_function(t, {'channel': 1}),
self.original_lt)
slice_lt = core.slice_function(self.original_lt, {'channel': 1})
self.assertLabeledTensorsEqual(map_lt, slice_lt)
def select(labeled_tensor, selection, name=None):
"""Slice out a subset of the tensor.
Args:
labeled_tensor: The input tensor.
selection: A dictionary mapping an axis name to a scalar, slice or list of
values to select. Currently supports two types of selections:
(a) Any number of scalar and/or slice selections.
(b) Exactly one list selection, without any scalars or slices.
name: Optional op name.
Returns:
The selection as a `LabeledTensor`.
Raises:
ValueError: If the tensor doesn't have an axis in the selection or if
that axis lacks labels.
KeyError: If any labels in a selection are not found in the original axis.
NotImplementedError: If you attempt to combine a list selection with
scalar selection or another list selection.
"""
with ops.name_scope(name, 'lt_select', [labeled_tensor]) as scope:
labeled_tensor = core.convert_to_labeled_tensor(labeled_tensor)
slices = {}
indexers = {}
for axis_name, value in selection.items():
if axis_name not in labeled_tensor.axes:
raise ValueError(
'The tensor does not have an axis named %s. Its axes are: %r' %
(axis_name, labeled_tensor.axes.keys()))
axis = labeled_tensor.axes[axis_name]
if axis.labels is None:
raise ValueError(
'The axis named %s does not have labels. The axis is: %r' %
(axis_name, axis))
if isinstance(value, slice):
# TODO(shoyer): consider deprecating using slices in favor of lists
if value.start is None:
start = None
else:
start = axis.index(value.start)
if value.stop is None:
stop = None
else:
# For now, follow the pandas convention of making labeled slices
# inclusive of both bounds.
stop = axis.index(value.stop) + 1
if value.step is not None:
raise NotImplementedError('slicing with a step is not yet supported')
slices[axis_name] = slice(start, stop)
else:
# We're allowing anything NumPy treats as a scalar or 1D array.
value = np.asarray(value)
if value.ndim == 0:
slices[axis_name] = axis.index(value.item())
elif value.ndim == 1:
if indexers:
raise NotImplementedError(
'select does not yet support more than one list selection at '
'the same time')
indexer = [axis.index(v) for v in value.tolist()]
indexers[axis_name] = ops.convert_to_tensor(
indexer, dtype=dtypes.int64)
else:
raise NotImplementedError(
'select does not yet support selections with more than one '
'dimension: %s on axis %r' % (value, axis_name))
if indexers and slices:
raise NotImplementedError(
'select does not yet support combined scalar and list selection')
# For now, handle array selection separately, because tf.gather_nd does
# not support gradients yet. Later, using gather_nd will let us combine
# these paths.
if indexers:
(axis_name, indexer), = indexers.items()
axis = core.Axis(axis_name, selection[axis_name])
return _gather_1d_on_axis(labeled_tensor, indexer, axis, name=scope)
else:
return core.slice_function(labeled_tensor, slices, name=scope)
def test_slice_unknown_shape(self):
lt = core.LabeledTensor(tf.placeholder(tf.float32, [None, 1]), ['x', 'y'])
sliced_lt = core.slice_function(lt, {'y': 0})
self.assertEqual(list(sliced_lt.axes.values()), [lt.axes['x']])
def test_slice_unknown_shape(self):
lt = core.LabeledTensor(
array_ops.placeholder(dtypes.float32, [None, 1]), ['x', 'y'])
sliced_lt = core.slice_function(lt, {'y': 0})
self.assertEqual(list(sliced_lt.axes.values()), [lt.axes['x']])
def test(self):
squeeze_lt = ops.squeeze(self.squeezable_lt, ['probs'])
golden_lt = core.slice_function(self.squeezable_lt, {'probs': 0})
self.assertLabeledTensorsEqual(squeeze_lt, golden_lt)
def test_name(self):
select_lt = core.slice_function(self.original_lt, {'channel': 1})
self.assertIn('lt_slice', select_lt.name)
def test_name(self):
foldl_lt = ops.foldl(core.add, self.original_lt,
core.slice_function(self.original_lt, {'x': 0}))
self.assertIn('lt_foldl', foldl_lt.name)
def test_slice(self):
map_lt = ops.map_fn(lambda t: core.slice_function(t, {'channel': 1}),
self.original_lt)
slice_lt = core.slice_function(self.original_lt, {'channel': 1})
self.assertLabeledTensorsEqual(map_lt, slice_lt)
def test_scalar(self):
select_lt = core.slice_function(self.original_lt, {'channel': 1})
golden_lt = core.LabeledTensor(self.tensor[:, 1, :, :], [self.a0, self.a2,
self.a3])
self.assertLabeledTensorsEqual(select_lt, golden_lt)
def select(labeled_tensor, selection, name=None):
"""Slice out a subset of the tensor.
Args:
labeled_tensor: The input tensor.
selection: A dictionary mapping an axis name to a scalar, slice or list of
values to select. Currently supports two types of selections:
(a) Any number of scalar and/or slice selections.
(b) Exactly one list selection, without any scalars or slices.
name: Optional op name.
Returns:
The selection as a `LabeledTensor`.
Raises:
ValueError: If the tensor doesn't have an axis in the selection or if
that axis lacks labels.
KeyError: If any labels in a selection are not found in the original axis.
NotImplementedError: If you attempt to combine a list selection with
scalar selection or another list selection.
"""
with ops.name_scope(name, 'lt_select', [labeled_tensor]) as scope:
labeled_tensor = core.convert_to_labeled_tensor(labeled_tensor)
slices = {}
indexers = {}
for axis_name, value in selection.items():
if axis_name not in labeled_tensor.axes:
raise ValueError(
'The tensor does not have an axis named %s. Its axes are: %r'
% (axis_name, labeled_tensor.axes.keys()))
axis = labeled_tensor.axes[axis_name]
if axis.labels is None:
raise ValueError(
'The axis named %s does not have labels. The axis is: %r' %
(axis_name, axis))
if isinstance(value, slice):
# TODO(shoyer): consider deprecating using slices in favor of lists
if value.start is None:
start = None
else:
start = axis.index(value.start)
if value.stop is None:
stop = None
else:
# For now, follow the pandas convention of making labeled slices
# inclusive of both bounds.
stop = axis.index(value.stop) + 1
if value.step is not None:
raise NotImplementedError(
'slicing with a step is not yet supported')
slices[axis_name] = slice(start, stop)
# Needs to be after checking for slices, since slice objects claim to be
# instances of collections.Hashable but hash() on them fails.
elif isinstance(value, collections.Hashable):
slices[axis_name] = axis.index(value)
elif isinstance(value, list):
if indexers:
raise NotImplementedError(
'select does not yet support more than one list selection at '
'the same time')
indexer = [axis.index(v) for v in value]
indexers[axis_name] = ops.convert_to_tensor(indexer,
dtype=dtypes.int64)
else:
# If type checking is working properly, this shouldn't be possible.
raise TypeError('cannot handle arbitrary types')
if indexers and slices:
raise NotImplementedError(
'select does not yet support combined scalar and list selection'
)
# For now, handle array selection separately, because tf.gather_nd does
# not support gradients yet. Later, using gather_nd will let us combine
# these paths.
if indexers:
(axis_name, indexer), = indexers.items()
axis = core.Axis(axis_name, selection[axis_name])
return _gather_1d_on_axis(labeled_tensor,
indexer,
axis,
name=scope)
else:
return core.slice_function(labeled_tensor, slices, name=scope)
def test_name(self):
foldl_lt = ops.foldl(core.add, self.original_lt,
core.slice_function(self.original_lt, {'x': 0}))
self.assertIn('lt_foldl', foldl_lt.name)
