Python源码示例：syntaxnet.util.check.Le()
示例1
def testCheckLe(self):
check.Le(1, 2, 'foo')
check.Le(1, 1, 'foo')
with self.assertRaisesRegexp(ValueError, 'bar'):
check.Le(1, 0, 'bar')
with self.assertRaisesRegexp(RuntimeError, 'baz'):
check.Le(1, -1, 'baz', RuntimeError)
示例2
def testCheckLe(self):
check.Le(1, 2, 'foo')
check.Le(1, 1, 'foo')
with self.assertRaisesRegexp(ValueError, 'bar'):
check.Le(1, 0, 'bar')
with self.assertRaisesRegexp(RuntimeError, 'baz'):
check.Le(1, -1, 'baz', RuntimeError)
示例3
def testCheckLe(self):
check.Le(1, 2, 'foo')
check.Le(1, 1, 'foo')
with self.assertRaisesRegexp(ValueError, 'bar'):
check.Le(1, 0, 'bar')
with self.assertRaisesRegexp(RuntimeError, 'baz'):
check.Le(1, -1, 'baz', RuntimeError)
示例4
def testCheckLe(self):
check.Le(1, 2, 'foo')
check.Le(1, 1, 'foo')
with self.assertRaisesRegexp(ValueError, 'bar'):
check.Le(1, 0, 'bar')
with self.assertRaisesRegexp(RuntimeError, 'baz'):
check.Le(1, -1, 'baz', RuntimeError)
示例5
def testCheckLe(self):
check.Le(1, 2, 'foo')
check.Le(1, 1, 'foo')
with self.assertRaisesRegexp(ValueError, 'bar'):
check.Le(1, 0, 'bar')
with self.assertRaisesRegexp(RuntimeError, 'baz'):
check.Le(1, -1, 'baz', RuntimeError)
示例6
def testCheckLe(self):
check.Le(1, 2, 'foo')
check.Le(1, 1, 'foo')
with self.assertRaisesRegexp(ValueError, 'bar'):
check.Le(1, 0, 'bar')
with self.assertRaisesRegexp(RuntimeError, 'baz'):
check.Le(1, -1, 'baz', RuntimeError)
示例7
def testCheckLe(self):
check.Le(1, 2, 'foo')
check.Le(1, 1, 'foo')
with self.assertRaisesRegexp(ValueError, 'bar'):
check.Le(1, 0, 'bar')
with self.assertRaisesRegexp(RuntimeError, 'baz'):
check.Le(1, -1, 'baz', RuntimeError)
示例8
def testCheckLe(self):
check.Le(1, 2, 'foo')
check.Le(1, 1, 'foo')
with self.assertRaisesRegexp(ValueError, 'bar'):
check.Le(1, 0, 'bar')
with self.assertRaisesRegexp(RuntimeError, 'baz'):
check.Le(1, -1, 'baz', RuntimeError)
示例9
def testCheckLe(self):
check.Le(1, 2, 'foo')
check.Le(1, 1, 'foo')
with self.assertRaisesRegexp(ValueError, 'bar'):
check.Le(1, 0, 'bar')
with self.assertRaisesRegexp(RuntimeError, 'baz'):
check.Le(1, -1, 'baz', RuntimeError)
示例10
def maybe_apply_dropout(inputs, keep_prob, per_sequence, stride=None):
"""Applies dropout, if so configured, to an input tensor.
The input may be rank 2 or 3 depending on whether the stride (i.e., batch
size) has been incorporated into the shape.
Args:
inputs: [stride * num_steps, dim] or [stride, num_steps, dim] input tensor.
keep_prob: Scalar probability of keeping each input element. If >= 1.0, no
dropout is performed.
per_sequence: If true, sample the dropout mask once per sequence, instead of
once per step. Requires |stride| when true.
stride: Scalar batch size. Optional if |per_sequence| is false.
Returns:
[stride * num_steps, dim] or [stride, num_steps, dim] tensor, matching the
shape of |inputs|, containing the masked or original inputs, depending on
whether dropout was actually performed.
"""
check.Ge(inputs.get_shape().ndims, 2, 'inputs must be rank 2 or 3')
check.Le(inputs.get_shape().ndims, 3, 'inputs must be rank 2 or 3')
flat = (inputs.get_shape().ndims == 2)
if keep_prob >= 1.0:
return inputs
if not per_sequence:
return tf.nn.dropout(inputs, keep_prob)
check.NotNone(stride, 'per-sequence dropout requires stride')
dim = inputs.get_shape().as_list()[-1]
check.NotNone(dim, 'inputs must have static activation dimension, but have '
'static shape %s' % inputs.get_shape().as_list())
# If needed, restore the batch dimension to separate the sequences.
inputs_sxnxd = tf.reshape(inputs, [stride, -1, dim]) if flat else inputs
# Replace |num_steps| with 1 in |noise_shape|, so the dropout mask broadcasts
# to all steps for a particular sequence.
noise_shape = [stride, 1, dim]
masked_sxnxd = tf.nn.dropout(inputs_sxnxd, keep_prob, noise_shape)
# If needed, flatten out the batch dimension in the return value.
return tf.reshape(masked_sxnxd, [-1, dim]) if flat else masked_sxnxd
示例11
def maybe_apply_dropout(inputs, keep_prob, per_sequence, stride=None):
"""Applies dropout, if so configured, to an input tensor.
The input may be rank 2 or 3 depending on whether the stride (i.e., batch
size) has been incorporated into the shape.
Args:
inputs: [stride * num_steps, dim] or [stride, num_steps, dim] input tensor.
keep_prob: Scalar probability of keeping each input element. If >= 1.0, no
dropout is performed.
per_sequence: If true, sample the dropout mask once per sequence, instead of
once per step. Requires |stride| when true.
stride: Scalar batch size. Optional if |per_sequence| is false.
Returns:
[stride * num_steps, dim] or [stride, num_steps, dim] tensor, matching the
shape of |inputs|, containing the masked or original inputs, depending on
whether dropout was actually performed.
"""
if keep_prob >= 1.0:
return inputs
if not per_sequence:
return tf.nn.dropout(inputs, keep_prob)
# We only check the dims if we are applying per-sequence dropout
check.Ge(inputs.get_shape().ndims, 2, 'inputs must be rank 2 or 3')
check.Le(inputs.get_shape().ndims, 3, 'inputs must be rank 2 or 3')
flat = (inputs.get_shape().ndims == 2)
check.NotNone(stride, 'per-sequence dropout requires stride')
dim = inputs.get_shape().as_list()[-1]
check.NotNone(dim, 'inputs must have static activation dimension, but have '
'static shape %s' % inputs.get_shape().as_list())
# If needed, restore the batch dimension to separate the sequences.
inputs_sxnxd = tf.reshape(inputs, [stride, -1, dim]) if flat else inputs
# Replace |num_steps| with 1 in |noise_shape|, so the dropout mask broadcasts
# to all steps for a particular sequence.
noise_shape = [stride, 1, dim]
masked_sxnxd = tf.nn.dropout(inputs_sxnxd, keep_prob, noise_shape)
# If needed, flatten out the batch dimension in the return value.
return tf.reshape(masked_sxnxd, [-1, dim]) if flat else masked_sxnxd
示例12
def maybe_apply_dropout(inputs, keep_prob, per_sequence, stride=None):
"""Applies dropout, if so configured, to an input tensor.
The input may be rank 2 or 3 depending on whether the stride (i.e., batch
size) has been incorporated into the shape.
Args:
inputs: [stride * num_steps, dim] or [stride, num_steps, dim] input tensor.
keep_prob: Scalar probability of keeping each input element. If >= 1.0, no
dropout is performed.
per_sequence: If true, sample the dropout mask once per sequence, instead of
once per step. Requires |stride| when true.
stride: Scalar batch size. Optional if |per_sequence| is false.
Returns:
[stride * num_steps, dim] or [stride, num_steps, dim] tensor, matching the
shape of |inputs|, containing the masked or original inputs, depending on
whether dropout was actually performed.
"""
if keep_prob >= 1.0:
return inputs
if not per_sequence:
return tf.nn.dropout(inputs, keep_prob)
# We only check the dims if we are applying per-sequence dropout
check.Ge(inputs.get_shape().ndims, 2, 'inputs must be rank 2 or 3')
check.Le(inputs.get_shape().ndims, 3, 'inputs must be rank 2 or 3')
flat = (inputs.get_shape().ndims == 2)
check.NotNone(stride, 'per-sequence dropout requires stride')
dim = inputs.get_shape().as_list()[-1]
check.NotNone(dim, 'inputs must have static activation dimension, but have '
'static shape %s' % inputs.get_shape().as_list())
# If needed, restore the batch dimension to separate the sequences.
inputs_sxnxd = tf.reshape(inputs, [stride, -1, dim]) if flat else inputs
# Replace |num_steps| with 1 in |noise_shape|, so the dropout mask broadcasts
# to all steps for a particular sequence.
noise_shape = [stride, 1, dim]
masked_sxnxd = tf.nn.dropout(inputs_sxnxd, keep_prob, noise_shape)
# If needed, flatten out the batch dimension in the return value.
return tf.reshape(masked_sxnxd, [-1, dim]) if flat else masked_sxnxd
示例13
def maybe_apply_dropout(inputs, keep_prob, per_sequence, stride=None):
"""Applies dropout, if so configured, to an input tensor.
The input may be rank 2 or 3 depending on whether the stride (i.e., batch
size) has been incorporated into the shape.
Args:
inputs: [stride * num_steps, dim] or [stride, num_steps, dim] input tensor.
keep_prob: Scalar probability of keeping each input element. If >= 1.0, no
dropout is performed.
per_sequence: If true, sample the dropout mask once per sequence, instead of
once per step. Requires |stride| when true.
stride: Scalar batch size. Optional if |per_sequence| is false.
Returns:
[stride * num_steps, dim] or [stride, num_steps, dim] tensor, matching the
shape of |inputs|, containing the masked or original inputs, depending on
whether dropout was actually performed.
"""
check.Ge(inputs.get_shape().ndims, 2, 'inputs must be rank 2 or 3')
check.Le(inputs.get_shape().ndims, 3, 'inputs must be rank 2 or 3')
flat = (inputs.get_shape().ndims == 2)
if keep_prob >= 1.0:
return inputs
if not per_sequence:
return tf.nn.dropout(inputs, keep_prob)
check.NotNone(stride, 'per-sequence dropout requires stride')
dim = inputs.get_shape().as_list()[-1]
check.NotNone(dim, 'inputs must have static activation dimension, but have '
'static shape %s' % inputs.get_shape().as_list())
# If needed, restore the batch dimension to separate the sequences.
inputs_sxnxd = tf.reshape(inputs, [stride, -1, dim]) if flat else inputs
# Replace |num_steps| with 1 in |noise_shape|, so the dropout mask broadcasts
# to all steps for a particular sequence.
noise_shape = [stride, 1, dim]
masked_sxnxd = tf.nn.dropout(inputs_sxnxd, keep_prob, noise_shape)
# If needed, flatten out the batch dimension in the return value.
return tf.reshape(masked_sxnxd, [-1, dim]) if flat else masked_sxnxd
示例14
def maybe_apply_dropout(inputs, keep_prob, per_sequence, stride=None):
"""Applies dropout, if so configured, to an input tensor.
The input may be rank 2 or 3 depending on whether the stride (i.e., batch
size) has been incorporated into the shape.
Args:
inputs: [stride * num_steps, dim] or [stride, num_steps, dim] input tensor.
keep_prob: Scalar probability of keeping each input element. If >= 1.0, no
dropout is performed.
per_sequence: If true, sample the dropout mask once per sequence, instead of
once per step. Requires |stride| when true.
stride: Scalar batch size. Optional if |per_sequence| is false.
Returns:
[stride * num_steps, dim] or [stride, num_steps, dim] tensor, matching the
shape of |inputs|, containing the masked or original inputs, depending on
whether dropout was actually performed.
"""
check.Ge(inputs.get_shape().ndims, 2, 'inputs must be rank 2 or 3')
check.Le(inputs.get_shape().ndims, 3, 'inputs must be rank 2 or 3')
flat = (inputs.get_shape().ndims == 2)
if keep_prob >= 1.0:
return inputs
if not per_sequence:
return tf.nn.dropout(inputs, keep_prob)
check.NotNone(stride, 'per-sequence dropout requires stride')
dim = inputs.get_shape().as_list()[-1]
check.NotNone(dim, 'inputs must have static activation dimension, but have '
'static shape %s' % inputs.get_shape().as_list())
# If needed, restore the batch dimension to separate the sequences.
inputs_sxnxd = tf.reshape(inputs, [stride, -1, dim]) if flat else inputs
# Replace |num_steps| with 1 in |noise_shape|, so the dropout mask broadcasts
# to all steps for a particular sequence.
noise_shape = [stride, 1, dim]
masked_sxnxd = tf.nn.dropout(inputs_sxnxd, keep_prob, noise_shape)
# If needed, flatten out the batch dimension in the return value.
return tf.reshape(masked_sxnxd, [-1, dim]) if flat else masked_sxnxd
示例15
def maybe_apply_dropout(inputs, keep_prob, per_sequence, stride=None):
"""Applies dropout, if so configured, to an input tensor.
The input may be rank 2 or 3 depending on whether the stride (i.e., batch
size) has been incorporated into the shape.
Args:
inputs: [stride * num_steps, dim] or [stride, num_steps, dim] input tensor.
keep_prob: Scalar probability of keeping each input element. If >= 1.0, no
dropout is performed.
per_sequence: If true, sample the dropout mask once per sequence, instead of
once per step. Requires |stride| when true.
stride: Scalar batch size. Optional if |per_sequence| is false.
Returns:
[stride * num_steps, dim] or [stride, num_steps, dim] tensor, matching the
shape of |inputs|, containing the masked or original inputs, depending on
whether dropout was actually performed.
"""
if keep_prob >= 1.0:
return inputs
if not per_sequence:
return tf.nn.dropout(inputs, keep_prob)
# We only check the dims if we are applying per-sequence dropout
check.Ge(inputs.get_shape().ndims, 2, 'inputs must be rank 2 or 3')
check.Le(inputs.get_shape().ndims, 3, 'inputs must be rank 2 or 3')
flat = (inputs.get_shape().ndims == 2)
check.NotNone(stride, 'per-sequence dropout requires stride')
dim = inputs.get_shape().as_list()[-1]
check.NotNone(dim, 'inputs must have static activation dimension, but have '
'static shape %s' % inputs.get_shape().as_list())
# If needed, restore the batch dimension to separate the sequences.
inputs_sxnxd = tf.reshape(inputs, [stride, -1, dim]) if flat else inputs
# Replace |num_steps| with 1 in |noise_shape|, so the dropout mask broadcasts
# to all steps for a particular sequence.
noise_shape = [stride, 1, dim]
masked_sxnxd = tf.nn.dropout(inputs_sxnxd, keep_prob, noise_shape)
# If needed, flatten out the batch dimension in the return value.
return tf.reshape(masked_sxnxd, [-1, dim]) if flat else masked_sxnxd
示例16
def maybe_apply_dropout(inputs, keep_prob, per_sequence, stride=None):
"""Applies dropout, if so configured, to an input tensor.
The input may be rank 2 or 3 depending on whether the stride (i.e., batch
size) has been incorporated into the shape.
Args:
inputs: [stride * num_steps, dim] or [stride, num_steps, dim] input tensor.
keep_prob: Scalar probability of keeping each input element. If >= 1.0, no
dropout is performed.
per_sequence: If true, sample the dropout mask once per sequence, instead of
once per step. Requires |stride| when true.
stride: Scalar batch size. Optional if |per_sequence| is false.
Returns:
[stride * num_steps, dim] or [stride, num_steps, dim] tensor, matching the
shape of |inputs|, containing the masked or original inputs, depending on
whether dropout was actually performed.
"""
check.Ge(inputs.get_shape().ndims, 2, 'inputs must be rank 2 or 3')
check.Le(inputs.get_shape().ndims, 3, 'inputs must be rank 2 or 3')
flat = (inputs.get_shape().ndims == 2)
if keep_prob >= 1.0:
return inputs
if not per_sequence:
return tf.nn.dropout(inputs, keep_prob)
check.NotNone(stride, 'per-sequence dropout requires stride')
dim = inputs.get_shape().as_list()[-1]
check.NotNone(dim, 'inputs must have static activation dimension, but have '
'static shape %s' % inputs.get_shape().as_list())
# If needed, restore the batch dimension to separate the sequences.
inputs_sxnxd = tf.reshape(inputs, [stride, -1, dim]) if flat else inputs
# Replace |num_steps| with 1 in |noise_shape|, so the dropout mask broadcasts
# to all steps for a particular sequence.
noise_shape = [stride, 1, dim]
masked_sxnxd = tf.nn.dropout(inputs_sxnxd, keep_prob, noise_shape)
# If needed, flatten out the batch dimension in the return value.
return tf.reshape(masked_sxnxd, [-1, dim]) if flat else masked_sxnxd
示例17
def apply_feature_id_dropout(ids, weights, channel):
"""Randomly perturbs a vector of feature IDs.
Args:
ids: Vector of feature IDs.
weights: Vector of feature weights.
channel: FixedFeatureChannel that extracted the |ids|.
Returns:
Copy of |ids| and |weights| where each ID is randomly replaced with
|channel.dropout_id|, according to the probabilities in
|channel.dropout_keep_probabilities|. The weights of dropped features are
set to zero if |channel.dropped_id| equals |channel.vocabulary_size|.
"""
check.Gt(
len(channel.dropout_keep_probability), 0,
'Channel {} dropout_keep_probability is empty'.format(channel.name))
check.Le(
len(channel.dropout_keep_probability), channel.vocabulary_size,
'Channel {} dropout_keep_probability is too long'.format(channel.name))
# Channel fields, converted from proto to constant tensor.
dropout_id = tf.constant(
channel.dropout_id, name='dropout_id', dtype=tf.int64)
dropout_keep_probabilities = tf.constant(
list(channel.dropout_keep_probability),
name='dropout_keep_probability',
dtype=tf.float32,
shape=[channel.vocabulary_size])
# The keep probabilities for the current batch of feature IDs.
keep_probabilities = tf.gather(dropout_keep_probabilities, ids)
# Draw random values and determine which IDs should be kept.
shape = tf.shape(ids)
noise = tf.random_uniform(shape) # \in [0,1)^d
should_keep = noise < keep_probabilities
# Replace dropped IDs with the specified replacement ID.
dropout_ids = tf.fill(shape, dropout_id)
new_ids = tf.where(should_keep, ids, dropout_ids)
if channel.dropout_id == channel.vocabulary_size:
# Replace weights of dropped IDs with 0.
zeros = tf.zeros(shape, dtype=tf.float32)
new_weights = tf.where(should_keep, weights, zeros)
else:
new_weights = weights
return new_ids, new_weights
示例18
def maybe_apply_dropout(inputs,
keep_prob,
per_sequence,
stride=None,
dropout_mask=None,
name=None):
"""Applies dropout, if so configured, to an input tensor.
The input may be rank 2 or 3 depending on whether the stride (i.e., batch
size) has been incorporated into the shape.
Args:
inputs: [stride * num_steps, dim] or [stride, num_steps, dim] input tensor.
keep_prob: Scalar probability of keeping each input element. If >= 1.0, no
dropout is performed.
per_sequence: If true, sample the dropout mask once per sequence, instead of
once per step. Either |stride| or |dropout_mask| must be set when true.
stride: Scalar batch size. Optional if |per_sequence| is false, or if
|dropout_mask| is provided.
dropout_mask: Precomputed dropout mask to apply to the |inputs|; must be
broadcastable to |inputs|. Optional if |per_sequence| is false, or if
|stride| is provided.
name: Optional name for the dropout operation, if dropout is applied.
Returns:
[stride * num_steps, dim] or [stride, num_steps, dim] tensor, matching the
shape of |inputs|, containing the masked or original inputs, depending on
whether dropout was actually performed.
"""
if keep_prob >= 1.0:
return inputs
if not per_sequence:
return tf.nn.dropout(inputs, keep_prob, name=name)
if dropout_mask is not None:
return tf.multiply(inputs, dropout_mask, name=name)
# We only check the dims if we are applying per-sequence dropout
check.Ge(inputs.get_shape().ndims, 2, 'inputs must be rank 2 or 3')
check.Le(inputs.get_shape().ndims, 3, 'inputs must be rank 2 or 3')
flat = (inputs.get_shape().ndims == 2)
check.NotNone(stride, 'per-sequence dropout requires stride')
dim = inputs.get_shape().as_list()[-1]
check.NotNone(dim, 'inputs must have static activation dimension, but have '
'static shape %s' % inputs.get_shape().as_list())
# If needed, restore the batch dimension to separate the sequences.
inputs_sxnxd = tf.reshape(inputs, [stride, -1, dim]) if flat else inputs
# Replace |num_steps| with 1 in |noise_shape|, so the dropout mask broadcasts
# to all steps for a particular sequence.
noise_shape = [stride, 1, dim]
masked_sxnxd = tf.nn.dropout(inputs_sxnxd, keep_prob, noise_shape, name=name)
# If needed, flatten out the batch dimension in the return value.
return tf.reshape(masked_sxnxd, [-1, dim]) if flat else masked_sxnxd
示例19
def apply_feature_id_dropout(ids, weights, channel):
"""Randomly perturbs a vector of feature IDs.
Args:
ids: Vector of feature IDs.
weights: Vector of feature weights.
channel: FixedFeatureChannel that extracted the |ids|.
Returns:
Copy of |ids| and |weights| where each ID is randomly replaced with
|channel.dropout_id|, according to the probabilities in
|channel.dropout_keep_probabilities|. The weights of dropped features are
set to zero if |channel.dropped_id| equals |channel.vocabulary_size|.
"""
check.Gt(
len(channel.dropout_keep_probability), 0,
'Channel {} dropout_keep_probability is empty'.format(channel.name))
check.Le(
len(channel.dropout_keep_probability), channel.vocabulary_size,
'Channel {} dropout_keep_probability is too long'.format(channel.name))
# Channel fields, converted from proto to constant tensor.
dropout_id = tf.constant(
channel.dropout_id, name='dropout_id', dtype=tf.int64)
dropout_keep_probabilities = tf.constant(
list(channel.dropout_keep_probability),
name='dropout_keep_probability',
dtype=tf.float32,
shape=[channel.vocabulary_size])
# The keep probabilities for the current batch of feature IDs.
keep_probabilities = tf.gather(dropout_keep_probabilities, ids)
# Draw random values and determine which IDs should be kept.
shape = tf.shape(ids)
noise = tf.random_uniform(shape) # \in [0,1)^d
should_keep = noise < keep_probabilities
# Replace dropped IDs with the specified replacement ID.
dropout_ids = tf.fill(shape, dropout_id)
new_ids = tf.where(should_keep, ids, dropout_ids)
if channel.dropout_id == channel.vocabulary_size:
# Replace weights of dropped IDs with 0.
zeros = tf.zeros(shape, dtype=tf.float32)
new_weights = tf.where(should_keep, weights, zeros)
else:
new_weights = weights
return new_ids, new_weights
示例20
def maybe_apply_dropout(inputs,
keep_prob,
per_sequence,
stride=None,
dropout_mask=None,
name=None):
"""Applies dropout, if so configured, to an input tensor.
The input may be rank 2 or 3 depending on whether the stride (i.e., batch
size) has been incorporated into the shape.
Args:
inputs: [stride * num_steps, dim] or [stride, num_steps, dim] input tensor.
keep_prob: Scalar probability of keeping each input element. If >= 1.0, no
dropout is performed.
per_sequence: If true, sample the dropout mask once per sequence, instead of
once per step. Either |stride| or |dropout_mask| must be set when true.
stride: Scalar batch size. Optional if |per_sequence| is false, or if
|dropout_mask| is provided.
dropout_mask: Precomputed dropout mask to apply to the |inputs|; must be
broadcastable to |inputs|. Optional if |per_sequence| is false, or if
|stride| is provided.
name: Optional name for the dropout operation, if dropout is applied.
Returns:
[stride * num_steps, dim] or [stride, num_steps, dim] tensor, matching the
shape of |inputs|, containing the masked or original inputs, depending on
whether dropout was actually performed.
"""
if keep_prob >= 1.0:
return inputs
if not per_sequence:
return tf.nn.dropout(inputs, keep_prob, name=name)
if dropout_mask is not None:
return tf.multiply(inputs, dropout_mask, name=name)
# We only check the dims if we are applying per-sequence dropout
check.Ge(inputs.get_shape().ndims, 2, 'inputs must be rank 2 or 3')
check.Le(inputs.get_shape().ndims, 3, 'inputs must be rank 2 or 3')
flat = (inputs.get_shape().ndims == 2)
check.NotNone(stride, 'per-sequence dropout requires stride')
dim = inputs.get_shape().as_list()[-1]
check.NotNone(dim, 'inputs must have static activation dimension, but have '
'static shape %s' % inputs.get_shape().as_list())
# If needed, restore the batch dimension to separate the sequences.
inputs_sxnxd = tf.reshape(inputs, [stride, -1, dim]) if flat else inputs
# Replace |num_steps| with 1 in |noise_shape|, so the dropout mask broadcasts
# to all steps for a particular sequence.
noise_shape = [stride, 1, dim]
masked_sxnxd = tf.nn.dropout(inputs_sxnxd, keep_prob, noise_shape, name=name)
# If needed, flatten out the batch dimension in the return value.
return tf.reshape(masked_sxnxd, [-1, dim]) if flat else masked_sxnxd
