def bidirectional_rnn(incoming, rnncell_fw, rnncell_bw, return_seq=False,
return_states=False, initial_state_fw=None,
initial_state_bw=None, dynamic=False, scope=None,
name="BiRNN"):
""" Bidirectional RNN.
Build a bidirectional recurrent neural network, it requires 2 RNN Cells
to process sequence in forward and backward order. Any RNN Cell can be
used i.e. SimpleRNN, LSTM, GRU... with its own parameters. But the two
cells number of units must match.
Input:
3-D Tensor Layer [samples, timesteps, input dim].
Output:
if `return_seq`: 3-D Tensor [samples, timesteps, output dim].
else: 2-D Tensor Layer [samples, output dim].
Arguments:
incoming: `Tensor`. The incoming Tensor.
rnncell_fw: `RNNCell`. The RNN Cell to use for foward computation.
rnncell_bw: `RNNCell`. The RNN Cell to use for backward computation.
return_seq: `bool`. If True, returns the full sequence instead of
last sequence output only.
return_states: `bool`. If True, returns a tuple with output and
states: (output, states).
initial_state_fw: `Tensor`. An initial state for the forward RNN.
This must be a tensor of appropriate type and shape [batch_size
x cell.state_size].
initial_state_bw: `Tensor`. An initial state for the backward RNN.
This must be a tensor of appropriate type and shape [batch_size
x cell.state_size].
dynamic: `bool`. If True, dynamic computation is performed. It will not
compute RNN steps above the sequence length. Note that because TF
requires to feed sequences of same length, 0 is used as a mask.
So a sequence padded with 0 at the end must be provided. When
computation is performed, it will stop when it meets a step with
a value of 0.
scope: `str`. Define this layer scope (optional). A scope can be
used to share varibales between layers. Note that scope will
override name.
name: `str`. A name for this layer (optional).
"""
assert (rnncell_fw._num_units == rnncell_bw._num_units), \
"RNN Cells number of units must match!"
sequence_length = None
if dynamic:
sequence_length = retrieve_seq_length_op(
incoming if isinstance(incoming, tf.Tensor) else tf.pack(incoming))
input_shape = utils.get_incoming_shape(incoming)
with tf.variable_op_scope([incoming], scope, name) as scope:
name = scope.name
# TODO: DropoutWrapper
inference = incoming
# If a tensor given, convert it to a per timestep list
if type(inference) not in [list, np.array]:
ndim = len(input_shape)
assert ndim >= 3, "Input dim should be at least 3."
axes = [1, 0] + list(range(2, ndim))
inference = tf.transpose(inference, (axes))
inference = tf.unpack(inference)
outputs, states_fw, states_bw = _brnn(
rnncell_fw, rnncell_bw, inference,
initial_state_fw=initial_state_fw,
initial_state_bw=initial_state_bw,
sequence_length=sequence_length,
dtype=tf.float32)
c = tf.GraphKeys.LAYER_VARIABLES + '/' + scope.name
for v in [rnncell_fw.W, rnncell_fw.b, rnncell_bw.W, rnncell_bw.b]:
if hasattr(v, "__len__"):
for var in v: tf.add_to_collection(c, var)
else:
tf.add_to_collection(c, v)
# Track activations.
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, outputs[-1])
if dynamic:
outputs = tf.transpose(tf.pack(outputs), [1, 0, 2])
o = advanced_indexing_op(outputs, sequence_length)
else:
o = outputs if return_seq else outputs[-1]
sfw = states_fw
sbw = states_fw
return (o, sfw, sbw) if return_states else o
def bidirectional_rnn(incoming,
rnncell_fw,
rnncell_bw,
return_seq=False,
return_states=False,
initial_state_fw=None,
initial_state_bw=None,
dynamic=False,
scope=None,
name="BiRNN"):
""" Bidirectional RNN.
Build a bidirectional recurrent neural network, it requires 2 RNN Cells
to process sequence in forward and backward order. Any RNN Cell can be
used i.e. SimpleRNN, LSTM, GRU... with its own parameters. But the two
cells number of units must match.
Input:
3-D Tensor Layer [samples, timesteps, input dim].
Output:
if `return_seq`: 3-D Tensor [samples, timesteps, output dim].
else: 2-D Tensor Layer [samples, output dim].
Arguments:
incoming: `Tensor`. The incoming Tensor.
rnncell_fw: `RNNCell`. The RNN Cell to use for foward computation.
rnncell_bw: `RNNCell`. The RNN Cell to use for backward computation.
return_seq: `bool`. If True, returns the full sequence instead of
last sequence output only.
return_states: `bool`. If True, returns a tuple with output and
states: (output, states).
initial_state_fw: `Tensor`. An initial state for the forward RNN.
This must be a tensor of appropriate type and shape [batch_size
x cell.state_size].
initial_state_bw: `Tensor`. An initial state for the backward RNN.
This must be a tensor of appropriate type and shape [batch_size
x cell.state_size].
dynamic: `bool`. If True, dynamic computation is performed. It will not
compute RNN steps above the sequence length. Note that because TF
requires to feed sequences of same length, 0 is used as a mask.
So a sequence padded with 0 at the end must be provided. When
computation is performed, it will stop when it meets a step with
a value of 0.
scope: `str`. Define this layer scope (optional). A scope can be
used to share variables between layers. Note that scope will
override name.
name: `str`. A name for this layer (optional).
"""
assert (rnncell_fw._num_units == rnncell_bw._num_units), \
"RNN Cells number of units must match!"
sequence_length = None
if dynamic:
sequence_length = retrieve_seq_length_op(
incoming if isinstance(incoming, tf.Tensor) else tf.pack(incoming))
input_shape = utils.get_incoming_shape(incoming)
with tf.variable_scope(scope, name, values=[incoming]) as scope:
name = scope.name
# TODO: DropoutWrapper
inference = incoming
# If a tensor given, convert it to a per timestep list
if type(inference) not in [list, np.array]:
ndim = len(input_shape)
assert ndim >= 3, "Input dim should be at least 3."
axes = [1, 0] + list(range(2, ndim))
inference = tf.transpose(inference, (axes))
inference = tf.unpack(inference)
outputs, states_fw, states_bw = _brnn(
rnncell_fw,
rnncell_bw,
inference,
initial_state_fw=initial_state_fw,
initial_state_bw=initial_state_bw,
sequence_length=sequence_length,
dtype=tf.float32)
c = tf.GraphKeys.LAYER_VARIABLES + '/' + scope.name
for v in [rnncell_fw.W, rnncell_fw.b, rnncell_bw.W, rnncell_bw.b]:
if hasattr(v, "__len__"):
for var in v:
tf.add_to_collection(c, var)
else:
tf.add_to_collection(c, v)
# Track activations.
tf.add_to_collection(tf.GraphKeys.ACTIVATIONS, outputs[-1])
if dynamic:
if return_seq:
o = outputs
else:
outputs = tf.transpose(tf.pack(outputs), [1, 0, 2])
o = advanced_indexing_op(outputs, sequence_length)
else:
o = outputs if return_seq else outputs[-1]
sfw = states_fw
sbw = states_fw
# Track output tensor.
tf.add_to_collection(tf.GraphKeys.LAYER_TENSOR + '/' + name, o)
return (o, sfw, sbw) if return_states else o