def __call__(self, query, state, softmaxed=True):
"""Score the query based on the keys and values.
Args:
query: Tensor of dtype matching `self.values` and shape
`[batch_size, query_depth]`.
state: Tensor of dtype matching `self.values` and shape
`[batch_size, alignments_size]`
(`alignments_size` is memory's `max_time`).
Returns:
alignments: Tensor of dtype matching `self.values` and shape
`[batch_size, alignments_size]` (`alignments_size` is memory's
`max_time`).
"""
with variable_scope.variable_scope(None, "bahdanau_attention",
[query]):
processed_query = self.query_layer(
query) if self.query_layer else query
score = _bahdanau_score(processed_query, self._keys,
self._normalize)
if softmaxed:
alignments = self._probability_fn(score, state)
next_state = alignments
return alignments, next_state
score = masked_mean_pooling(score, self._memory_sequence_length,
int(self._values.shape[1]))
return score
def __call__(self, query, state=None):
with vs.variable_scope(self._attention_scope, reuse=True):
processed_query = (self.query_layer(query)
if self.query_layer else query)
score = _bahdanau_score(processed_query=processed_query,
keys=self._keys,
normalize=self._normalize)
alignments = self._probability_fn(score, state)
next_state = alignments
return alignments, next_state
def __call__(self, query, state):
processed_query = self.query_layer(
query) if self.query_layer else query
with tf.variable_scope("bahdanau_attention", reuse=tf.AUTO_REUSE):
score = _bahdanau_score(processed_query, self._keys,
self._normalize)
alignments = self._probability_fn(score, state)
next_state = alignments
return alignments, next_state
def __call__(self, query, state):
"""Score the query based on the keys and values.
Args:
query: Tensor of dtype matching `self.values` and shape
`[batch_size, query_depth]`.
state: Tensor of dtype matching `self.values` and shape
`[batch_size, alignments_size]`
(`alignments_size` is memory's `max_time`).
Returns:
alignments: Tensor of dtype matching `self.values` and shape
`[batch_size, alignments_size]` (`alignments_size` is memory's
`max_time`).
"""
with tf.variable_scope(None, "bahdanau_monotonic_hccho_attention",
[query]):
processed_query = self.query_layer(
query) if self.query_layer else query
# processed_query: (N,num_units) ==> self._keys와 더하기 위해서 _bahdanau_score 내부에서 (N,1,num_units) 으로 변환.
# self._keys: (N, encoder_dim, num_units)
score = _bahdanau_score(processed_query, self._keys,
self._normalize) # keys 가 memory임
score_bias = tf.get_variable("attention_score_bias",
dtype=processed_query.dtype,
initializer=self._score_bias_init)
#alignments_bias = tf.get_variable("alignments_bias", shape = state.get_shape()[-1],dtype=processed_query.dtype, initializer=tf.zeros_initializer()) # hccho
alignments_bias = tf.get_variable(
"alignments_bias",
shape=(1),
dtype=processed_query.dtype,
initializer=tf.zeros_initializer()) # hccho
score += score_bias
alignments = self._probability_fn(
score, state) #BahdanauAttention에서 _probability_fn = softmax
next_state = alignments # 다음 alignment 계산에 사용할 state 값
# hccho. alignment가 attention 계산에 직접 사용된다.
alignments = tf.nn.relu(alignments + alignments_bias)
alignments = alignments / (tf.reduce_sum(
alignments, axis=-1, keepdims=True) + 1.0e-12) # hccho 수정
return alignments, next_state
def __call__(self, query, state):
"""Score the query based on the keys and values.
Args:
query: Tensor of dtype matching `self.values` and shape
`[batch_size, query_depth]`.
state: Tensor of dtype matching `self.values` and shape
`[batch_size, alignments_size]`
(`alignments_size` is memory's `max_time`).
Returns:
alignments: Tensor of dtype matching `self.values` and shape
`[batch_size, alignments_size]` (`alignments_size` is memory's
`max_time`).
"""
with variable_scope.variable_scope(None, "bahdanau_monotonic_attention", [query]):
processed_query = self.query_layer(query) if self.query_layer else query
score = _bahdanau_score(processed_query, self._keys, self._normalize)
score_bias = variable_scope.get_variable(
"attention_score_bias", dtype=processed_query.dtype,
initializer=self._score_bias_init)
score += score_bias
alignments = self._probability_fn(score, state)
next_state = alignments
return alignments, next_state