def __init__(self,
cell,
memory,
mask=None,
controller=None,
mapper=None,
input_keep_prob=1.0,
is_train=None):
"""
Early fusion attention cell: uses the (inputs, state) to control the current attention.
:param cell:
:param memory: [N, M, m]
:param mask:
:param controller: (inputs, prev_state, memory) -> memory_logits
"""
self._cell = cell
self._memory = memory
self._mask = mask
self._flat_memory = flatten(memory, 2)
self._flat_mask = flatten(mask, 1)
if controller is None:
controller = AttentionCell.get_linear_controller(True,
is_train=is_train)
self._controller = controller
if mapper is None:
mapper = AttentionCell.get_concat_mapper()
elif mapper == 'sim':
mapper = AttentionCell.get_sim_mapper()
self._mapper = mapper
def dynamic_rnn(cell,
inputs,
sequence_length=None,
initial_state=None,
dtype=None,
parallel_iterations=None,
swap_memory=False,
time_major=False,
scope=None):
assert not time_major # TODO : to be implemented later!
print("dynamic rnn input")
print(inputs.get_shape())
flat_inputs = flatten(inputs, 2) # [-1, J, d]
print("dynamic rnn flatten shape")
print(flat_inputs.get_shape())
flat_len = None if sequence_length is None else tf.cast(
flatten(sequence_length, 0), 'int64')
flat_outputs, final_state = _dynamic_rnn(
cell,
flat_inputs,
sequence_length=flat_len,
initial_state=initial_state,
dtype=dtype,
parallel_iterations=parallel_iterations,
swap_memory=swap_memory,
time_major=time_major,
scope=scope)
print("flat_outputs shape")
print(flat_outputs.get_shape())
outputs = reconstruct(flat_outputs, inputs, 2)
return outputs, final_state
def bidirectional_dynamic_rnn(cell_fw,
cell_bw,
inputs,
sequence_length=None,
initial_state_fw=None,
initial_state_bw=None,
dtype=None,
parallel_iterations=None,
swap_memory=False,
time_major=False,
scope=None):
assert not time_major
flat_inputs = flatten(inputs, 2) # [-1, J, d]
flat_len = None if sequence_length is None else tf.cast(
flatten(sequence_length, 0), 'int64')
(flat_fw_outputs, flat_bw_outputs), final_state = \
_bidirectional_dynamic_rnn(cell_fw, cell_bw, flat_inputs, sequence_length=flat_len,
initial_state_fw=initial_state_fw, initial_state_bw=initial_state_bw,
dtype=dtype, parallel_iterations=parallel_iterations, swap_memory=swap_memory,
time_major=time_major, scope=scope)
fw_outputs = reconstruct(flat_fw_outputs, inputs, 2)
bw_outputs = reconstruct(flat_bw_outputs, inputs, 2)
# FIXME : final state is not reshaped!
return (fw_outputs, bw_outputs), final_state
def bw_dynamic_rnn(cell,
inputs,
sequence_length=None,
initial_state=None,
dtype=None,
parallel_iterations=None,
swap_memory=False,
time_major=False,
scope=None):
assert not time_major # TODO : to be implemented later!
flat_inputs = flatten(inputs, 2) # [-1, J, d]
flat_len = None if sequence_length is None else tf.cast(
flatten(sequence_length, 0), 'int64')
flat_inputs = tf.reverse(flat_inputs, 1) if sequence_length is None \
else tf.reverse_sequence(flat_inputs, sequence_length, 1)
flat_outputs, final_state = _dynamic_rnn(
cell,
flat_inputs,
sequence_length=flat_len,
initial_state=initial_state,
dtype=dtype,
parallel_iterations=parallel_iterations,
swap_memory=swap_memory,
time_major=time_major,
scope=scope)
flat_outputs = tf.reverse(flat_outputs, 1) if sequence_length is None \
else tf.reverse_sequence(flat_outputs, sequence_length, 1)
outputs = reconstruct(flat_outputs, inputs, 2)
return outputs, final_state
def linear(args,
output_size,
bias,
bias_start=0.0,
scope=None,
squeeze=False,
wd=0.0,
input_keep_prob=1.0,
is_train=None):
with tf.variable_scope(scope or "linear"):
if args is None or (nest.is_sequence(args) and not args):
raise ValueError("`args` must be specified")
if not nest.is_sequence(args):
args = [args]
flat_args = [flatten(arg, 1) for arg in args]
# if input_keep_prob < 1.0:
assert is_train is not None
flat_args = [
tf.cond(is_train, lambda: tf.nn.dropout(arg, input_keep_prob),
lambda: arg) for arg in flat_args
]
flat_out = _linear(flat_args, output_size, bias)
out = reconstruct(flat_out, args[0], 1)
if squeeze:
out = tf.squeeze(out, [len(args[0].get_shape().as_list()) - 1])
if wd:
add_wd(wd)
return out
def softmax(logits, mask=None, scope=None):
with tf.name_scope(scope or "Softmax"):
if mask is not None:
logits = exp_mask(logits, mask)
flat_logits = flatten(logits, 1)
flat_out = tf.nn.softmax(flat_logits)
out = reconstruct(flat_out, logits, 1)
return out