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 _resource_apply(self, grad, var, indices=None):
# 准备变量
var_dtype = var.dtype.base_dtype
lr_t = self._decayed_lr(var_dtype)
m = self.get_slot(var, 'm')
v = self.get_slot(var, 'v')
beta_1_t = self._get_hyper('beta_1', var_dtype)
beta_2_t = self._get_hyper('beta_2', var_dtype)
epsilon_t = K.cast(self.epsilon, var_dtype)
local_step = K.cast(self.iterations + 1, var_dtype)
beta_1_t_power = K.pow(beta_1_t, local_step)
beta_2_t_power = K.pow(beta_2_t, local_step)
# 更新公式
if indices is None:
m_t = K.update(m, beta_1_t * m + (1 - beta_1_t) * grad)
v_t = K.update(v, beta_2_t * v + (1 - beta_2_t) * grad**2)
else:
mv_ops = [K.update(m, beta_1_t * m), K.update(v, beta_2_t * v)]
with tf.control_dependencies(mv_ops):
m_t = self._resource_scatter_add(m, indices,
(1 - beta_1_t) * grad)
v_t = self._resource_scatter_add(v, indices,
(1 - beta_2_t) * grad**2)
# 返回算子
with tf.control_dependencies([m_t, v_t]):
if self.bias_correction:
m_t = m_t / (1.0 - beta_1_t_power)
v_t = v_t / (1.0 - beta_2_t_power)
var_t = var - lr_t * m_t / (K.sqrt(v_t) + self.epsilon)
return K.update(var, var_t)
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 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 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:
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 self.attention_scale:
a = a / self.key_size**0.5
a = sequence_masking(a, v_mask, 1, -1)
if a_mask is not None:
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
def sparse_accuracy(self, y_true, y_pred):
"""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')
# 逐标签取最大来粗略评测训练效果
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 get_updates(self, loss, params):
# 更新判据
cond = K.equal(self.iterations % self.grad_accum_steps, 0)
cond = K.cast(cond, K.floatx())
# 获取梯度
grads = self.get_gradients(loss, params)
self.accum_grads = [
K.zeros(K.int_shape(p),
dtype=K.dtype(p),
name='accum_grad_%s' % i) for i, p in enumerate(params)
]
old_update = K.update
def new_update(x, new_x):
new_x = cond * new_x + (1 - cond) * x
return old_update(x, new_x)
K.update = new_update
updates = super(NewOptimizer, self).get_updates(loss, params)
K.update = old_update
# 累积梯度
with tf.control_dependencies(updates):
accum_updates = [
K.update(ag, g + (1 - cond) * ag)
for g, ag in zip(grads, self.accum_grads)
]
return accum_updates
def sparse_loss(self, y_true, y_pred):
"""y_true需要是整数形式（非one hot）
"""
# 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')
# 转为one hot
y_true = K.one_hot(y_true, K.shape(self.trans)[0])
return self.dense_loss(y_true, y_pred)
def compile(self):
# 交叉熵作为loss，并mask掉输入部分的预测
y_true = self.model.input[0][:, 1:] # 目标tokens
y_mask = self.model.input[1][:, 1:] # 目标mask
y_mask = K.cast(y_mask, K.floatx()) # 转为浮点型
y_pred = self.model.output[:, :-1] # 预测tokens，预测与目标错开一位
cross_entropy = K.sparse_categorical_crossentropy(y_true, y_pred)
cross_entropy = K.sum(cross_entropy * y_mask) / K.sum(y_mask)
self.model.add_loss(cross_entropy)
opt = extend_with_gradient_accumulation(Adam)(learning_rate=0.000015,
grad_accum_steps=2)
self.model.compile(optimizer=opt)
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, 0.0),
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 learning_rate(self):
if self._learning_rate is None:
iterations = K.cast(self.iterations + 1, K.floatx())
learning_rate = K.minimum(1.0 / K.sqrt(iterations), 0.01)
if self.multiply_by_parameter_scale:
return learning_rate
else:
return learning_rate * 0.05
else:
if not hasattr(self, '__learning_rate'):
with K.name_scope(self.__class__.__name__):
self.__learning_rate = K.variable(self._learning_rate,
name='learning_rate')
return self.__learning_rate
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
mask = K.all(K.greater(y_pred, -1e6), axis=2, keepdims=True)
mask = K.cast(mask, K.floatx())
# 计算目标分数
y_true, y_pred = y_true * mask, y_pred * mask
target_score = self.target_score(y_true, y_pred)
# 递归计算log Z
init_states = [y_pred[:, 0]]
y_pred = K.concatenate([y_pred, mask], axis=2)
input_length = K.int_shape(y_pred[:, 1:])[1]
log_norm, _, _ = K.rnn(self.log_norm_step,
y_pred[:, 1:],
init_states,
input_length=input_length) # 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):
"""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 _decayed_lr(self, var_dtype):
lr_multiplier = piecewise_linear(self.iterations, self.lr_schedule)
lr_t = super(NewOptimizer, self)._decayed_lr(var_dtype)
return lr_t * K.cast(lr_multiplier, var_dtype)
def reverse_sequence(self, inputs, mask=None):
if mask is None:
return [x[:, ::-1] for x in inputs]
else:
length = K.cast(K.sum(mask, 1), 'int32')
return [tf.reverse_sequence(x, length, seq_axis=1) for x in inputs]
def call(self, inputs, mask=None):
if mask is not None:
mask = K.cast(mask, K.floatx())
return sequence_masking(inputs, mask, 1, 1)
def new_update(x, new_x):
if is_one_of(x, params) and self._do_lazy_optimization(x):
g = self.grads[x]
r = K.any(K.not_equal(g, 0.0), axis=-1, keepdims=True)
new_x = x + (new_x - x) * K.cast(r, K.floatx())
return old_update(x, new_x)
def beta2(self):
if self._beta2 is None:
iterations = K.cast(self.iterations + 1, K.floatx())
return 1.0 - K.pow(iterations, -0.8)
else:
return self._beta2