def call(self, inputs, mask=None):
x = inputs
# Pooling
if self.pool_size > 1:
if mask is not None:
mask = K.cast(mask, K.floatx())
x_in_len = K.shape(x)[1]
x = sequence_masking(x, mask, 0)
x = divisible_temporal_padding(x, self.pool_size)
x = pool1d(x, self.pool_size, self.pool_size, pool_mode='avg')
# 执行FFN
x = self.dense_1(x)
x = self.dense_2(x)
# 恢复长度
if self.pool_size > 1:
x = K.repeat_elements(x, self.pool_size, 1)[:, :x_in_len]
# 返回结果
return x
def call(self, inputs, mask=None):
x = inputs
# Pooling
if self.pool_size > 1:
if mask is not None:
if not hasattr(self, 'mask_layer'):
self.mask_layer = search_layer(x, mask)
mask = self.mask_layer.output_mask
x_in_len = K.shape(x)[1]
x = sequence_masking(x, mask, 0)
x = divisible_temporal_padding(x, self.pool_size)
x = pool1d(x, self.pool_size, self.pool_size, pool_mode='avg')
# 执行FFN
x = self.dense_1(x)
x = self.dense_2(x)
# 恢复长度
if self.pool_size > 1:
x = K.repeat_elements(x, self.pool_size, 1)[:, :x_in_len]
# 返回结果
return x
def call(self, inputs, q_mask=None, v_mask=None, a_mask=None):
"""实现多头注意力
q_mask: 对输入的query序列的mask。
主要是将输出结果的padding部分置0。
v_mask: 对输入的value序列的mask。
主要是防止attention读取到padding信息。
a_mask: 对attention矩阵的mask。
不同的attention mask对应不同的应用。
"""
q, k, v = inputs[:3]
if a_mask:
if len(inputs) == 3:
a_mask = 'history_only'
else:
a_mask = inputs[3]
if q_mask is not None:
if not hasattr(self, 'q_mask_layer'):
self.q_mask_layer = search_layer(q, q_mask)
q_mask = self.q_mask_layer.output_mask
if v_mask is not None:
if not hasattr(self, 'v_mask_layer'):
self.v_mask_layer = search_layer(v, v_mask)
v_mask = self.v_mask_layer.output_mask
# Pooling
if self.pool_size > 1:
is_self_attention = (q is k is v)
q_in_len = K.shape(q)[1]
q = sequence_masking(q, q_mask, 0)
q = divisible_temporal_padding(q, self.pool_size)
q = pool1d(q, self.pool_size, self.pool_size, pool_mode='avg')
if is_self_attention:
k = v = q
else:
k = sequence_masking(k, v_mask, 0)
k = divisible_temporal_padding(k, self.pool_size)
k = pool1d(k, self.pool_size, self.pool_size, pool_mode='avg')
v = sequence_masking(v, v_mask, 0)
v = divisible_temporal_padding(v, self.pool_size)
v = pool1d(v, self.pool_size, self.pool_size, pool_mode='avg')
if v_mask is not None:
v_mask = v_mask[:, ::self.pool_size]
if a_mask is not None and not is_string(a_mask):
a_mask = a_mask[..., ::self.pool_size, ::self.pool_size]
# 线性变换
qw = self.q_dense(q)
kw = self.k_dense(k)
vw = self.v_dense(v)
# 形状变换
qw = K.reshape(qw, (-1, K.shape(q)[1], self.heads, self.key_size))
kw = K.reshape(kw, (-1, K.shape(k)[1], self.heads, self.key_size))
vw = K.reshape(vw, (-1, K.shape(v)[1], self.heads, self.head_size))
# Attention
a = tf.einsum('bjhd,bkhd->bhjk', qw, kw)
# 相对位置编码
if self.max_relative_position is not None:
q_idxs = K.arange(0, K.shape(q)[1], dtype='int32')
q_idxs = K.expand_dims(q_idxs, 1)
v_idxs = K.arange(0, K.shape(v)[1], dtype='int32')
v_idxs = K.expand_dims(v_idxs, 0)
pos_ids = v_idxs - q_idxs
pos_ids = K.clip(pos_ids, -self.max_relative_position,
self.max_relative_position)
pos_ids = pos_ids + self.max_relative_position
pos_embeddings = K.gather(self.relative_embeddings, pos_ids)
a = a + tf.einsum('bjhd,jkd->bhjk', qw, pos_embeddings)
# Attention（续）
a = a / self.key_size**0.5
a = sequence_masking(a, v_mask, 1, -1)
if a_mask is not None:
if is_string(a_mask):
ones = K.ones_like(a[:1, :1])
a_mask = (ones - tf.linalg.band_part(ones, -1, 0)) * 1e12
a = a - a_mask
else:
a = a - (1 - a_mask) * 1e12
a = K.softmax(a)
# 完成输出
o = tf.einsum('bhjk,bkhd->bjhd', a, vw)
if self.max_relative_position is not None:
o = o + tf.einsum('bhjk,jkd->bjhd', a, pos_embeddings)
o = K.reshape(o, (-1, K.shape(o)[1], self.out_dim))
o = self.o_dense(o)
# 恢复长度
if self.pool_size > 1:
o = K.repeat_elements(o, self.pool_size, 1)[:, :q_in_len]
# 返回结果
o = sequence_masking(o, q_mask, 0)
return o
