def test_sequence_length_not_equal(self):
"""Tests that an error is raised when sequence lengths are not equal."""
# Input a with sequence_length = [2, 1]
sparse_input_a = sparse_tensor.SparseTensorValue(indices=((0, 0),
(0, 1), (1,
0)),
values=(0., 1., 10.),
dense_shape=(2, 2))
# Input b with sequence_length = [1, 1]
sparse_input_b = sparse_tensor.SparseTensorValue(indices=((0, 0), (1,
0)),
values=(1., 10.),
dense_shape=(2, 2))
numeric_column_a = sfc.sequence_numeric_column('aaa')
numeric_column_b = sfc.sequence_numeric_column('bbb')
_, sequence_length = sfc.sequence_input_layer(
features={
'aaa': sparse_input_a,
'bbb': sparse_input_b,
},
feature_columns=[numeric_column_a, numeric_column_b])
with monitored_session.MonitoredSession() as sess:
with self.assertRaisesRegexp(
errors.InvalidArgumentError,
r'\[Condition x == y did not hold element-wise:\] '
r'\[x \(sequence_input_layer/aaa/sequence_length:0\) = \] \[2 1\] '
r'\[y \(sequence_input_layer/bbb/sequence_length:0\) = \] \[1 1\]'
):
sess.run(sequence_length)
def test_sequence_example_into_input_layer(self):
examples = [_make_sequence_example().SerializeToString()] * 100
ctx_cols, seq_cols = self._build_feature_columns()
def _parse_example(example):
ctx, seq = parsing_ops.parse_single_sequence_example(
example,
context_features=fc.make_parse_example_spec(ctx_cols),
sequence_features=fc.make_parse_example_spec(seq_cols))
ctx.update(seq)
return ctx
ds = dataset_ops.Dataset.from_tensor_slices(examples)
ds = ds.map(_parse_example)
ds = ds.batch(20)
# Test on a single batch
features = ds.make_one_shot_iterator().get_next()
# Tile the context features across the sequence features
seq_layer, _ = sfc.sequence_input_layer(features, seq_cols)
ctx_layer = fc.input_layer(features, ctx_cols)
input_layer = sfc.concatenate_context_input(ctx_layer, seq_layer)
rnn_layer = recurrent.RNN(recurrent.SimpleRNNCell(10))
output = rnn_layer(input_layer)
with self.cached_session() as sess:
sess.run(variables.global_variables_initializer())
features_r = sess.run(features)
self.assertAllEqual(features_r['int_list'].dense_shape, [20, 3, 6])
output_r = sess.run(output)
self.assertAllEqual(output_r.shape, [20, 10])
def test_numeric_column_multi_dim(self):
"""Tests sequence_input_layer for multi-dimensional numeric_column."""
sparse_input = sparse_tensor.SparseTensorValue(
# example 0, values [[[0., 1.], [2., 3.]], [[4., 5.], [6., 7.]]]
# example 1, [[[10., 11.], [12., 13.]]]
indices=((0, 0), (0, 1), (0, 2), (0, 3), (0, 4), (0, 5), (0, 6),
(0, 7), (1, 0), (1, 1), (1, 2), (1, 3)),
values=(0., 1., 2., 3., 4., 5., 6., 7., 10., 11., 12., 13.),
dense_shape=(2, 8))
# The output of numeric_column._get_dense_tensor should be flattened.
expected_input_layer = [
[[0., 1., 2., 3.], [4., 5., 6., 7.]],
[[10., 11., 12., 13.], [0., 0., 0., 0.]],
]
expected_sequence_length = [2, 1]
numeric_column = sfc.sequence_numeric_column('aaa', shape=(2, 2))
input_layer, sequence_length = sfc.sequence_input_layer(
features={'aaa': sparse_input}, feature_columns=[numeric_column])
with monitored_session.MonitoredSession() as sess:
self.assertAllEqual(expected_input_layer,
input_layer.eval(session=sess))
self.assertAllEqual(expected_sequence_length,
sequence_length.eval(session=sess))
def test_sequence_length_not_equal(self):
"""Tests that an error is raised when sequence lengths are not equal."""
# Input a with sequence_length = [2, 1]
sparse_input_a = sparse_tensor.SparseTensorValue(
indices=((0, 0), (0, 1), (1, 0)),
values=(0., 1., 10.),
dense_shape=(2, 2))
# Input b with sequence_length = [1, 1]
sparse_input_b = sparse_tensor.SparseTensorValue(
indices=((0, 0), (1, 0)),
values=(1., 10.),
dense_shape=(2, 2))
numeric_column_a = sfc.sequence_numeric_column('aaa')
numeric_column_b = sfc.sequence_numeric_column('bbb')
_, sequence_length = sfc.sequence_input_layer(
features={
'aaa': sparse_input_a,
'bbb': sparse_input_b,
},
feature_columns=[numeric_column_a, numeric_column_b])
with monitored_session.MonitoredSession() as sess:
with self.assertRaisesRegexp(
errors.InvalidArgumentError,
r'\[Condition x == y did not hold element-wise:\] '
r'\[x \(sequence_input_layer/aaa/sequence_length:0\) = \] \[2 1\] '
r'\[y \(sequence_input_layer/bbb/sequence_length:0\) = \] \[1 1\]'):
sess.run(sequence_length)
def test_shared_embedding_column_with_non_sequence_categorical(self):
"""Tests that error is raised for non-sequence shared embedding column."""
vocabulary_size = 3
sparse_input_a = sparse_tensor.SparseTensorValue(
# example 0, ids [2]
# example 1, ids [0, 1]
indices=((0, 0), (1, 0), (1, 1)),
values=(2, 0, 1),
dense_shape=(2, 2))
sparse_input_b = sparse_tensor.SparseTensorValue(
# example 0, ids [2]
# example 1, ids [0, 1]
indices=((0, 0), (1, 0), (1, 1)),
values=(2, 0, 1),
dense_shape=(2, 2))
categorical_column_a = fc.categorical_column_with_identity(
key='aaa', num_buckets=vocabulary_size)
categorical_column_b = fc.categorical_column_with_identity(
key='bbb', num_buckets=vocabulary_size)
shared_embedding_columns = fc.shared_embedding_columns(
[categorical_column_a, categorical_column_b], dimension=2)
with self.assertRaisesRegexp(
ValueError,
r'In embedding_column: aaa_shared_embedding\. categorical_column must '
r'be of type _SequenceCategoricalColumn to use sequence_input_layer\.'):
_, _ = sfc.sequence_input_layer(
features={
'aaa': sparse_input_a,
'bbb': sparse_input_b
},
feature_columns=shared_embedding_columns)
def test_numeric_column_multi_dim(self):
"""Tests sequence_input_layer for multi-dimensional numeric_column."""
sparse_input = sparse_tensor.SparseTensorValue(
# example 0, values [[[0., 1.], [2., 3.]], [[4., 5.], [6., 7.]]]
# example 1, [[[10., 11.], [12., 13.]]]
indices=((0, 0), (0, 1), (0, 2), (0, 3), (0, 4), (0, 5), (0, 6), (0, 7),
(1, 0), (1, 1), (1, 2), (1, 3)),
values=(0., 1., 2., 3., 4., 5., 6., 7., 10., 11., 12., 13.),
dense_shape=(2, 8))
# The output of numeric_column._get_dense_tensor should be flattened.
expected_input_layer = [
[[0., 1., 2., 3.], [4., 5., 6., 7.]],
[[10., 11., 12., 13.], [0., 0., 0., 0.]],
]
expected_sequence_length = [2, 1]
numeric_column = sfc.sequence_numeric_column('aaa', shape=(2, 2))
input_layer, sequence_length = sfc.sequence_input_layer(
features={'aaa': sparse_input},
feature_columns=[numeric_column])
with monitored_session.MonitoredSession() as sess:
self.assertAllEqual(expected_input_layer, input_layer.eval(session=sess))
self.assertAllEqual(
expected_sequence_length, sequence_length.eval(session=sess))
def rnn_logit_fn(features, mode):
"""Recurrent Neural Network logit_fn.
Args:
features: This is the first item returned from the `input_fn`
passed to `train`, `evaluate`, and `predict`. This should be a
single `Tensor` or `dict` of same.
mode: Optional. Specifies if this training, evaluation or prediction. See
`ModeKeys`.
Returns:
A tuple of `Tensor` objects representing the logits and the sequence
length mask.
"""
# Can't import from tf.contrib at the module level, otherwise you
# can hit a circular import issue if tf_estimator.contrib is
# imported before tf.contrib.
from tensorflow.contrib.feature_column.python.feature_column import sequence_feature_column as seq_fc # pylint: disable=g-import-not-at-top
with variable_scope.variable_scope(
'sequence_input_layer',
values=tuple(six.itervalues(features)),
partitioner=input_layer_partitioner):
sequence_input, sequence_length = seq_fc.sequence_input_layer(
features=features, feature_columns=sequence_feature_columns)
summary.histogram('sequence_length', sequence_length)
if context_feature_columns:
context_input = feature_column_lib.input_layer(
features=features, feature_columns=context_feature_columns)
sequence_input = _concatenate_context_input(
sequence_input, context_input)
cell = rnn_cell_fn(mode)
# Ignore output state.
rnn_outputs, _ = rnn.dynamic_rnn(cell=cell,
inputs=sequence_input,
sequence_length=sequence_length,
dtype=dtypes.float32,
time_major=False)
if not return_sequences:
rnn_outputs = _select_last_activations(rnn_outputs,
sequence_length)
with variable_scope.variable_scope('logits', values=(rnn_outputs, )):
logits = core_layers.dense(
rnn_outputs,
units=output_units,
activation=None,
kernel_initializer=init_ops.glorot_uniform_initializer())
sequence_length_mask = array_ops.sequence_mask(sequence_length)
return logits, sequence_length_mask
def rnn_logit_fn(features, mode):
"""Recurrent Neural Network logit_fn.
Args:
features: This is the first item returned from the `input_fn`
passed to `train`, `evaluate`, and `predict`. This should be a
single `Tensor` or `dict` of same.
mode: Optional. Specifies if this training, evaluation or prediction. See
`ModeKeys`.
Returns:
A `Tensor` representing the logits.
"""
with variable_scope.variable_scope(
'sequence_input_layer',
values=tuple(six.itervalues(features)),
partitioner=input_layer_partitioner):
sequence_input, sequence_length = seq_fc.sequence_input_layer(
features=features, feature_columns=sequence_feature_columns)
summary.histogram('sequence_length', sequence_length)
if context_feature_columns:
context_input = feature_column_lib.input_layer(
features=features, feature_columns=context_feature_columns)
sequence_input = _concatenate_context_input(
sequence_input, context_input)
cell = rnn_cell_fn(mode)
# Ignore output state.
rnn_outputs, _ = rnn.dynamic_rnn(cell=cell,
inputs=sequence_input,
sequence_length=sequence_length,
dtype=dtypes.float32,
time_major=False)
if not return_sequences:
rnn_outputs = _select_last_activations(rnn_outputs,
sequence_length)
with variable_scope.variable_scope('logits', values=(rnn_outputs, )):
logits = core_layers.dense(
rnn_outputs,
units=output_units,
activation=None,
kernel_initializer=init_ops.glorot_uniform_initializer())
return logits
def rnn_logit_fn(features, mode):
"""Recurrent Neural Network logit_fn.
Args:
features: This is the first item returned from the `input_fn`
passed to `train`, `evaluate`, and `predict`. This should be a
single `Tensor` or `dict` of same.
mode: Optional. Specifies if this training, evaluation or prediction. See
`ModeKeys`.
Returns:
A `Tensor` representing the logits.
"""
with variable_scope.variable_scope(
'sequence_input_layer',
values=tuple(six.itervalues(features)),
partitioner=input_layer_partitioner):
sequence_input, sequence_length = seq_fc.sequence_input_layer(
features=features, feature_columns=sequence_feature_columns)
summary.histogram('sequence_length', sequence_length)
if context_feature_columns:
context_input = feature_column_lib.input_layer(
features=features,
feature_columns=context_feature_columns)
sequence_input = seq_fc.concatenate_context_input(
context_input, sequence_input)
cell = rnn_cell_fn(mode)
# Ignore output state.
rnn_outputs, _ = rnn.dynamic_rnn(
cell=cell,
inputs=sequence_input,
sequence_length=sequence_length,
dtype=dtypes.float32,
time_major=False)
last_activations = _select_last_activations(rnn_outputs, sequence_length)
with variable_scope.variable_scope('logits', values=(rnn_outputs,)):
logits = core_layers.dense(
last_activations,
units=output_units,
activation=None,
kernel_initializer=init_ops.glorot_uniform_initializer())
return logits
def test_indicator_column(self):
vocabulary_size_a = 3
sparse_input_a = sparse_tensor.SparseTensorValue(
# example 0, ids [2]
# example 1, ids [0, 1]
indices=((0, 0), (1, 0), (1, 1)),
values=(2, 0, 1),
dense_shape=(2, 2))
vocabulary_size_b = 2
sparse_input_b = sparse_tensor.SparseTensorValue(
# example 0, ids [1]
# example 1, ids [1, 0]
indices=((0, 0), (1, 0), (1, 1)),
values=(1, 1, 0),
dense_shape=(2, 2))
expected_input_layer = [
# example 0, ids_a [2], ids_b [1]
[[0., 0., 1., 0., 1.], [0., 0., 0., 0., 0.]],
# example 1, ids_a [0, 1], ids_b [1, 0]
[[1., 0., 0., 0., 1.], [0., 1., 0., 1., 0.]],
]
expected_sequence_length = [1, 2]
categorical_column_a = sfc.sequence_categorical_column_with_identity(
key='aaa', num_buckets=vocabulary_size_a)
indicator_column_a = sfc._sequence_indicator_column(
categorical_column_a)
categorical_column_b = sfc.sequence_categorical_column_with_identity(
key='bbb', num_buckets=vocabulary_size_b)
indicator_column_b = sfc._sequence_indicator_column(
categorical_column_b)
input_layer, sequence_length = sfc.sequence_input_layer(
features={
'aaa': sparse_input_a,
'bbb': sparse_input_b,
},
# Test that columns are reordered alphabetically.
feature_columns=[indicator_column_b, indicator_column_a])
with monitored_session.MonitoredSession() as sess:
self.assertAllEqual(expected_input_layer,
input_layer.eval(session=sess))
self.assertAllEqual(expected_sequence_length,
sequence_length.eval(session=sess))
def test_indicator_column(self):
vocabulary_size_a = 3
sparse_input_a = sparse_tensor.SparseTensorValue(
# example 0, ids [2]
# example 1, ids [0, 1]
indices=((0, 0), (1, 0), (1, 1)),
values=(2, 0, 1),
dense_shape=(2, 2))
vocabulary_size_b = 2
sparse_input_b = sparse_tensor.SparseTensorValue(
# example 0, ids [1]
# example 1, ids [1, 0]
indices=((0, 0), (1, 0), (1, 1)),
values=(1, 1, 0),
dense_shape=(2, 2))
expected_input_layer = [
# example 0, ids_a [2], ids_b [1]
[[0., 0., 1., 0., 1.], [0., 0., 0., 0., 0.]],
# example 1, ids_a [0, 1], ids_b [1, 0]
[[1., 0., 0., 0., 1.], [0., 1., 0., 1., 0.]],
]
expected_sequence_length = [1, 2]
categorical_column_a = sfc.sequence_categorical_column_with_identity(
key='aaa', num_buckets=vocabulary_size_a)
indicator_column_a = fc.indicator_column(categorical_column_a)
categorical_column_b = sfc.sequence_categorical_column_with_identity(
key='bbb', num_buckets=vocabulary_size_b)
indicator_column_b = fc.indicator_column(categorical_column_b)
input_layer, sequence_length = sfc.sequence_input_layer(
features={
'aaa': sparse_input_a,
'bbb': sparse_input_b,
},
# Test that columns are reordered alphabetically.
feature_columns=[indicator_column_b, indicator_column_a])
with monitored_session.MonitoredSession() as sess:
self.assertAllEqual(expected_input_layer, input_layer.eval(session=sess))
self.assertAllEqual(
expected_sequence_length, sequence_length.eval(session=sess))
def test_indicator_column_with_non_sequence_categorical(self):
"""Tests that error is raised for non-sequence categorical column."""
vocabulary_size = 3
sparse_input = sparse_tensor.SparseTensorValue(
# example 0, ids [2]
# example 1, ids [0, 1]
indices=((0, 0), (1, 0), (1, 1)),
values=(2, 0, 1),
dense_shape=(2, 2))
categorical_column_a = fc.categorical_column_with_identity(
key='aaa', num_buckets=vocabulary_size)
indicator_column_a = fc.indicator_column(categorical_column_a)
with self.assertRaisesRegexp(
ValueError,
r'In indicator_column: aaa_indicator\. categorical_column must be of '
r'type _SequenceCategoricalColumn to use sequence_input_layer\.'):
_, _ = sfc.sequence_input_layer(
features={'aaa': sparse_input},
feature_columns=[indicator_column_a])
def test_numeric_column(self):
sparse_input = sparse_tensor.SparseTensorValue(
# example 0, values [[0.], [1]]
# example 1, [[10.]]
indices=((0, 0), (0, 1), (1, 0)),
values=(0., 1., 10.),
dense_shape=(2, 2))
expected_input_layer = [
[[0.], [1.]],
[[10.], [0.]],
]
expected_sequence_length = [2, 1]
numeric_column = sfc.sequence_numeric_column('aaa')
input_layer, sequence_length = sfc.sequence_input_layer(
features={'aaa': sparse_input}, feature_columns=[numeric_column])
with monitored_session.MonitoredSession() as sess:
self.assertAllEqual(expected_input_layer,
input_layer.eval(session=sess))
self.assertAllEqual(expected_sequence_length,
sequence_length.eval(session=sess))
def test_numeric_column(self):
sparse_input = sparse_tensor.SparseTensorValue(
# example 0, values [[0.], [1]]
# example 1, [[10.]]
indices=((0, 0), (0, 1), (1, 0)),
values=(0., 1., 10.),
dense_shape=(2, 2))
expected_input_layer = [
[[0.], [1.]],
[[10.], [0.]],
]
expected_sequence_length = [2, 1]
numeric_column = sfc.sequence_numeric_column('aaa')
input_layer, sequence_length = sfc.sequence_input_layer(
features={'aaa': sparse_input},
feature_columns=[numeric_column])
with monitored_session.MonitoredSession() as sess:
self.assertAllEqual(expected_input_layer, input_layer.eval(session=sess))
self.assertAllEqual(
expected_sequence_length, sequence_length.eval(session=sess))
def test_embedding_column(self):
vocabulary_size = 3
sparse_input_a = sparse_tensor.SparseTensorValue(
# example 0, ids [2]
# example 1, ids [0, 1]
indices=((0, 0), (1, 0), (1, 1)),
values=(2, 0, 1),
dense_shape=(2, 2))
sparse_input_b = sparse_tensor.SparseTensorValue(
# example 0, ids [1]
# example 1, ids [2, 0]
indices=((0, 0), (1, 0), (1, 1)),
values=(1, 2, 0),
dense_shape=(2, 2))
embedding_dimension_a = 2
embedding_values_a = (
(1., 2.), # id 0
(3., 4.), # id 1
(5., 6.) # id 2
)
embedding_dimension_b = 3
embedding_values_b = (
(11., 12., 13.), # id 0
(14., 15., 16.), # id 1
(17., 18., 19.) # id 2
)
def _get_initializer(embedding_dimension, embedding_values):
def _initializer(shape, dtype, partition_info):
self.assertAllEqual((vocabulary_size, embedding_dimension),
shape)
self.assertEqual(dtypes.float32, dtype)
self.assertIsNone(partition_info)
return embedding_values
return _initializer
expected_input_layer = [
# example 0, ids_a [2], ids_b [1]
[[5., 6., 14., 15., 16.], [0., 0., 0., 0., 0.]],
# example 1, ids_a [0, 1], ids_b [2, 0]
[[1., 2., 17., 18., 19.], [3., 4., 11., 12., 13.]],
]
expected_sequence_length = [1, 2]
categorical_column_a = sfc.sequence_categorical_column_with_identity(
key='aaa', num_buckets=vocabulary_size)
embedding_column_a = sfc._sequence_embedding_column(
categorical_column_a,
dimension=embedding_dimension_a,
initializer=_get_initializer(embedding_dimension_a,
embedding_values_a))
categorical_column_b = sfc.sequence_categorical_column_with_identity(
key='bbb', num_buckets=vocabulary_size)
embedding_column_b = sfc._sequence_embedding_column(
categorical_column_b,
dimension=embedding_dimension_b,
initializer=_get_initializer(embedding_dimension_b,
embedding_values_b))
input_layer, sequence_length = sfc.sequence_input_layer(
features={
'aaa': sparse_input_a,
'bbb': sparse_input_b,
},
# Test that columns are reordered alphabetically.
feature_columns=[embedding_column_b, embedding_column_a])
global_vars = ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
self.assertItemsEqual(
('sequence_input_layer/aaa_embedding/embedding_weights:0',
'sequence_input_layer/bbb_embedding/embedding_weights:0'),
tuple([v.name for v in global_vars]))
with monitored_session.MonitoredSession() as sess:
self.assertAllEqual(embedding_values_a,
global_vars[0].eval(session=sess))
self.assertAllEqual(embedding_values_b,
global_vars[1].eval(session=sess))
self.assertAllEqual(expected_input_layer,
input_layer.eval(session=sess))
self.assertAllEqual(expected_sequence_length,
sequence_length.eval(session=sess))
def test_embedding_column(self):
vocabulary_size = 3
sparse_input_a = sparse_tensor.SparseTensorValue(
# example 0, ids [2]
# example 1, ids [0, 1]
indices=((0, 0), (1, 0), (1, 1)),
values=(2, 0, 1),
dense_shape=(2, 2))
sparse_input_b = sparse_tensor.SparseTensorValue(
# example 0, ids [1]
# example 1, ids [2, 0]
indices=((0, 0), (1, 0), (1, 1)),
values=(1, 2, 0),
dense_shape=(2, 2))
embedding_dimension_a = 2
embedding_values_a = (
(1., 2.), # id 0
(3., 4.), # id 1
(5., 6.) # id 2
)
embedding_dimension_b = 3
embedding_values_b = (
(11., 12., 13.), # id 0
(14., 15., 16.), # id 1
(17., 18., 19.) # id 2
)
def _get_initializer(embedding_dimension, embedding_values):
def _initializer(shape, dtype, partition_info):
self.assertAllEqual((vocabulary_size, embedding_dimension), shape)
self.assertEqual(dtypes.float32, dtype)
self.assertIsNone(partition_info)
return embedding_values
return _initializer
expected_input_layer = [
# example 0, ids_a [2], ids_b [1]
[[5., 6., 14., 15., 16.], [0., 0., 0., 0., 0.]],
# example 1, ids_a [0, 1], ids_b [2, 0]
[[1., 2., 17., 18., 19.], [3., 4., 11., 12., 13.]],
]
expected_sequence_length = [1, 2]
categorical_column_a = sfc.sequence_categorical_column_with_identity(
key='aaa', num_buckets=vocabulary_size)
embedding_column_a = fc.embedding_column(
categorical_column_a, dimension=embedding_dimension_a,
initializer=_get_initializer(embedding_dimension_a, embedding_values_a))
categorical_column_b = sfc.sequence_categorical_column_with_identity(
key='bbb', num_buckets=vocabulary_size)
embedding_column_b = fc.embedding_column(
categorical_column_b, dimension=embedding_dimension_b,
initializer=_get_initializer(embedding_dimension_b, embedding_values_b))
input_layer, sequence_length = sfc.sequence_input_layer(
features={
'aaa': sparse_input_a,
'bbb': sparse_input_b,
},
# Test that columns are reordered alphabetically.
feature_columns=[embedding_column_b, embedding_column_a])
global_vars = ops.get_collection(ops.GraphKeys.GLOBAL_VARIABLES)
self.assertItemsEqual(
('sequence_input_layer/aaa_embedding/embedding_weights:0',
'sequence_input_layer/bbb_embedding/embedding_weights:0'),
tuple([v.name for v in global_vars]))
with monitored_session.MonitoredSession() as sess:
self.assertAllEqual(embedding_values_a, global_vars[0].eval(session=sess))
self.assertAllEqual(embedding_values_b, global_vars[1].eval(session=sess))
self.assertAllEqual(expected_input_layer, input_layer.eval(session=sess))
self.assertAllEqual(
expected_sequence_length, sequence_length.eval(session=sess))
