def compute_position_ids(self, inputs):
"""T5的相对位置分桶（直接翻译自官方T5源码）
"""
q, v = inputs
# 计算位置差
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
# 后处理操作
num_buckets, max_distance = self.input_dim, self.max_distance
ret = 0
n = -pos_ids
if self.bidirectional:
num_buckets //= 2
ret += K.cast(K.less(n, 0), 'int32') * num_buckets
n = K.abs(n)
else:
n = K.maximum(n, 0)
# now n is in the range [0, inf)
max_exact = num_buckets // 2
is_small = K.less(n, max_exact)
val_if_large = max_exact + K.cast(
K.log(K.cast(n, K.floatx()) / max_exact) /
np.log(max_distance / max_exact) * (num_buckets - max_exact),
'int32',
)
val_if_large = K.minimum(val_if_large, num_buckets - 1)
ret += K.switch(is_small, n, val_if_large)
return ret
def log_norm_step(self, inputs, states):
"""递归计算归一化因子
要点：1、递归计算；2、用logsumexp避免溢出。
"""
inputs, mask = inputs[:, :-1], inputs[:, -1:]
states = K.expand_dims(states[0], 2) # (batch_size, output_dim, 1)
trans = K.expand_dims(self.trans, 0) # (1, output_dim, output_dim)
outputs = tf.reduce_logsumexp(states + trans,
1) # (batch_size, output_dim)
outputs = outputs + inputs
outputs = mask * outputs + (1 - mask) * states[:, :, 0]
return outputs, [outputs]
def compute_position_ids(self, inputs):
q, v = inputs
# 计算位置差
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
# 后处理操作
max_position = (self.input_dim - 1) // 2
pos_ids = K.clip(pos_ids, -max_position, max_position)
pos_ids = pos_ids + max_position
return pos_ids
def call(self, inputs):
"""如果是条件Layer Norm，则默认以list为输入，第二个是condition
"""
if self.conditional:
inputs, cond = inputs
if self.hidden_units is not None:
cond = self.hidden_dense(cond)
for _ in range(K.ndim(inputs) - K.ndim(cond)):
cond = K.expand_dims(cond, 1)
if self.center:
beta = self.beta_dense(cond) + self.beta
if self.scale:
gamma = self.gamma_dense(cond) + self.gamma
else:
if self.center:
beta = self.beta
if self.scale:
gamma = self.gamma
outputs = inputs
if self.center:
mean = K.mean(outputs, axis=-1, keepdims=True)
outputs = outputs - mean
if self.scale:
variance = K.mean(K.square(outputs), axis=-1, keepdims=True)
std = K.sqrt(variance + self.epsilon)
outputs = outputs / std
outputs = outputs * gamma
if self.center:
outputs = outputs + beta
return outputs
def call(self, inputs):
input_shape = K.shape(inputs)
batch_size, seq_len = input_shape[0], input_shape[1]
pos_embeddings = self.embeddings[:seq_len]
pos_embeddings = K.expand_dims(pos_embeddings, 0)
if self.merge_mode == 'add':
return inputs + pos_embeddings
else:
pos_embeddings = K.tile(pos_embeddings, [batch_size, 1, 1])
return K.concatenate([inputs, pos_embeddings])
def dense_loss(self, y_true, y_pred):
"""y_true需要是one hot形式
"""
# 导出mask并转换数据类型
if self.input_mask is None:
mask = None
else:
mask = K.cast(self.input_mask, K.floatx())
# 计算目标分数
target_score = self.target_score(y_true, y_pred, mask)
# 递归计算log Z
init_states = [y_pred[:, 0]]
if mask is None:
mask = K.ones_like(y_pred[:, :, :1])
else:
mask = K.expand_dims(mask, 2)
y_pred = K.concatenate([y_pred, mask])
log_norm, _, _ = K.rnn(self.log_norm_step, y_pred[:, 1:],
init_states) # 最后一步的log Z向量
log_norm = tf.reduce_logsumexp(log_norm, 1) # logsumexp得标量
# 计算损失 -log p
return log_norm - target_score
def call(self, inputs, mask=None, a_mask=None, p_bias=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。
p_bias: 在attention里的位置偏置。
一般用来指定相对位置编码的种类。
"""
q, k, v = inputs[:3]
q_mask, v_mask, n = None, None, 3
if mask is not None:
if mask[0] is not None:
q_mask = K.cast(mask[0], K.floatx())
if mask[2] is not None:
v_mask = K.cast(mask[2], K.floatx())
if a_mask:
if len(inputs) == 3:
a_mask = 'history_only'
else:
a_mask = inputs[n]
n += 1
# 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)
# 处理位置编码
if p_bias == 'typical_relative':
pos_embeddings = inputs[n]
a = a + tf.einsum('bjhd,jkd->bhjk', qw, pos_embeddings)
elif p_bias == 't5_relative':
pos_embeddings = K.permute_dimensions(inputs[n], (2, 0, 1))
a = a + K.expand_dims(pos_embeddings, 0)
# Attention（续）
if p_bias != 't5_relative': # T5不用缩放
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 p_bias == 'typical_relative':
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)
# 返回结果
o = sequence_masking(o, q_mask, 0)
return o
