def basic_loss(self, y_true, y_pred, go_backwards=False):
"""y_true需要是整数形式（非one hot）
"""
# 导出mask并转换数据类型
mask = K.all(K.greater(y_pred, -1e6), axis=2)
mask = K.cast(mask, K.floatx())
# y_true需要重新明确一下shape和dtype
y_true = K.reshape(y_true, K.shape(y_pred)[:-1])
y_true = K.cast(y_true, 'int32')
# 反转相关
if self.hidden_dim is None:
if go_backwards: # 是否反转序列
y_true, y_pred = self.reverse_sequence([y_true, y_pred], mask)
trans = K.transpose(self.trans)
else:
trans = self.trans
histoty = K.gather(trans, y_true)
else:
if go_backwards: # 是否反转序列
y_true, y_pred = self.reverse_sequence([y_true, y_pred], mask)
r_trans, l_trans = self.l_trans, self.r_trans
else:
l_trans, r_trans = self.l_trans, self.r_trans
histoty = K.gather(l_trans, y_true)
histoty = tf.einsum('bnd,kd->bnk', histoty, r_trans)
# 计算loss
histoty = K.concatenate([y_pred[:, :1], histoty[:, :-1]], 1)
y_pred = (y_pred + histoty) / 2
loss = K.sparse_categorical_crossentropy(
y_true, y_pred, from_logits=True
)
return K.sum(loss * mask) / K.sum(mask)
def call(self, inputs):
"""如果custom_position_ids，那么第二个输入为自定义的位置id
"""
input_shape = K.shape(inputs)
batch_size, seq_len = input_shape[0], input_shape[1]
if self.custom_position_ids:
inputs, position_ids = inputs
if K.dtype(position_ids) != 'int32':
position_ids = K.cast(position_ids, 'int32')
else:
position_ids = K.arange(0, seq_len, dtype='int32')[None]
if self.hierarchical:
alpha = 0.4 if self.hierarchical is True else self.hierarchical
embeddings = self.embeddings - alpha * self.embeddings[:1]
embeddings = embeddings / (1 - alpha)
embeddings_x = K.gather(embeddings, position_ids // self.input_dim)
embeddings_y = K.gather(embeddings, position_ids % self.input_dim)
pos_embeddings = alpha * embeddings_x + (1 - alpha) * embeddings_y
else:
if self.custom_position_ids:
pos_embeddings = K.gather(self.embeddings, position_ids)
else:
pos_embeddings = self.embeddings[None, :seq_len]
if self.merge_mode == 'add':
return inputs + pos_embeddings
elif self.merge_mode == 'mul':
return inputs * pos_embeddings
else:
if not self.custom_position_ids:
pos_embeddings = K.tile(pos_embeddings, [batch_size, 1, 1])
return K.concatenate([inputs, pos_embeddings])
def basic_loss(self, y_true, y_pred, go_backwards=False):
"""y_true需要是整数形式（非one hot）
"""
mask = self.output_mask
# y_true需要重新明确一下dtype和shape
y_true = K.cast(y_true, 'int32')
y_true = K.reshape(y_true, [K.shape(y_true)[0], -1])
# 反转相关
if self.hidden_dim is None:
if go_backwards: # 是否反转序列
y_true, y_pred = self.reverse_sequence([y_true, y_pred], mask)
trans = K.transpose(self.trans)
else:
trans = self.trans
histoty = K.gather(trans, y_true)
else:
if go_backwards: # 是否反转序列
y_true, y_pred = self.reverse_sequence([y_true, y_pred], mask)
r_trans, l_trans = self.l_trans, self.r_trans
else:
l_trans, r_trans = self.l_trans, self.r_trans
histoty = K.gather(l_trans, y_true)
histoty = tf.einsum('bnd,kd->bnk', histoty, r_trans)
# 计算loss
histoty = K.concatenate([y_pred[:, :1], histoty[:, :-1]], 1)
y_pred = (y_pred + histoty) / 2
loss = K.sparse_categorical_crossentropy(y_true,
y_pred,
from_logits=True)
if mask is None:
return K.mean(loss)
else:
return K.sum(loss * mask) / K.sum(mask)
def basic_accuracy(self, y_true, y_pred, go_backwards=False):
"""训练过程中显示逐帧准确率的函数，排除了mask的影响
此处y_true需要是整数形式（非one hot）
"""
# 导出mask并转换数据类型
mask = K.all(K.greater(y_pred, -1e6), axis=2)
mask = K.cast(mask, K.floatx())
# y_true需要重新明确一下shape和dtype
y_true = K.reshape(y_true, K.shape(y_pred)[:-1])
y_true = K.cast(y_true, 'int32')
# 反转相关
if self.hidden_dim is None:
if go_backwards: # 是否反转序列
y_true, y_pred = self.reverse_sequence([y_true, y_pred], mask)
trans = K.transpose(self.trans)
else:
trans = self.trans
histoty = K.gather(trans, y_true)
else:
if go_backwards: # 是否反转序列
y_true, y_pred = self.reverse_sequence([y_true, y_pred], mask)
r_trans, l_trans = self.l_trans, self.r_trans
else:
l_trans, r_trans = self.l_trans, self.r_trans
histoty = K.gather(l_trans, y_true)
histoty = tf.einsum('bnd,kd->bnk', histoty, r_trans)
# 计算逐标签accuracy
histoty = K.concatenate([y_pred[:, :1], histoty[:, :-1]], 1)
y_pred = (y_pred + histoty) / 2
y_pred = K.cast(K.argmax(y_pred, 2), 'int32')
isequal = K.cast(K.equal(y_true, y_pred), K.floatx())
return K.sum(isequal * mask) / K.sum(mask)
def new_update(x, new_x):
if x is var and self._do_lazy_optimization(x):
if indices is None:
r = K.any(K.not_equal(grad, 0.),
axis=-1,
keepdims=True)
new_x = x + (new_x - x) * K.cast(r, K.floatx())
return old_update(x, new_x)
else:
return self._resource_scatter_add(
x, indices, K.gather(new_x - x, indices))
return old_update(x, new_x)
def call(self, inputs):
"""如果inputs是一个list，则默认第二个输入是传入的位置id，否则
是默认顺序id，即[0, 1, 2, 3, ...]
"""
if isinstance(inputs, list):
inputs, pos_ids = inputs
pos_embeddings = K.gather(self.embeddings, pos_ids)
else:
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)
pos_embeddings = K.tile(pos_embeddings, [batch_size, 1, 1])
if self.merge_mode == 'add':
return inputs + pos_embeddings
else:
return K.concatenate([inputs, pos_embeddings])
def call(self, inputs):
"""如果custom_position_ids，那么第二个输入为自定义的位置id
"""
if self.custom_position_ids:
inputs, position_ids = inputs
if K.dtype(position_ids) != 'int32':
position_ids = K.cast(position_ids, 'int32')
pos_embeddings = K.gather(self.embeddings, position_ids)
else:
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':
pos_embeddings = K.tile(pos_embeddings, [batch_size, 1, 1])
if self.merge_mode == 'add':
return inputs + pos_embeddings
else:
return K.concatenate([inputs, pos_embeddings])
def basic_accuracy(self, y_true, y_pred, go_backwards=False):
"""训练过程中显示逐帧准确率的函数，排除了mask的影响
此处y_true需要是整数形式（非one hot）
"""
mask = self.output_mask
# y_true需要重新明确一下dtype和shape
y_true = K.cast(y_true, 'int32')
y_true = K.reshape(y_true, [K.shape(y_true)[0], -1])
# 是否反转序列
if go_backwards:
y_true, y_pred = self.reverse_sequence([y_true, y_pred], mask)
trans = K.transpose(self.trans)
else:
trans = self.trans
# 计算逐标签accuracy
histoty = K.gather(trans, y_true)
histoty = K.concatenate([y_pred[:, :1], histoty[:, :-1]], 1)
y_pred = (y_pred + histoty) / 2
y_pred = K.cast(K.argmax(y_pred, 2), 'int32')
isequal = K.cast(K.equal(y_true, y_pred), K.floatx())
if mask is None:
return K.mean(isequal)
else:
return K.sum(isequal * mask) / K.sum(mask)
def call(self, inputs):
pos_ids = self.compute_position_ids(inputs)
return K.gather(self.embeddings, pos_ids)
def call(self, inputs, q_mask=False, v_mask=False, a_mask=False):
"""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。
"""
# 处理mask
inputs = inputs[:]
for i, mask in enumerate([q_mask, v_mask, a_mask]):
if not mask:
inputs.insert(3 + i, None)
q, k, v, q_mask, v_mask = inputs[:5]
if len(inputs) == 5:
a_mask = 'history_only'
elif len(inputs) == 6:
a_mask = inputs[-1]
else:
raise ValueError('wrong inputs for MultiHeadAttention.')
# 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)
o = sequence_masking(o, q_mask, 0)
return o
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
def call(self, inputs):
if K.dtype(inputs) != 'int32':
inputs = K.cast(inputs, 'int32')
outputs = K.gather(self._embeddings, inputs)
outputs = K.dot(outputs, self._project_kernel)
return outputs
