def _resource_apply_op(self, grad, var, indices=None):
# 更新判据
cond = K.equal(self.iterations % self.grad_accum_steps, 0)
# 获取梯度
ag = self.get_slot(var, 'ag')
old_update = K.update
def new_update(x, new_x):
new_x = K.switch(cond, new_x, x)
return old_update(x, new_x)
K.update = new_update
ag_t = ag / self.grad_accum_steps
op = super(new_optimizer, self)._resource_apply_op(ag_t, var)
K.update = old_update
# 累积梯度
with tf.control_dependencies([op]):
ag_t = K.switch(cond, K.zeros_like(ag), ag)
with tf.control_dependencies([K.update(ag, ag_t)]):
if indices is None:
ag_t = K.update(ag, ag + grad)
else:
ag_t = self._resource_scatter_add(ag, indices, grad)
return ag_t
def _resource_apply_op(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. - beta_1_t_power)
v_t = v_t / (1. - beta_2_t_power)
var_t = var - lr_t * m_t / (K.sqrt(v_t) + self.epsilon)
return K.update(var, var_t)
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 build(self, input_shape):
output_dim = input_shape[-1]
if not isinstance(output_dim, int):
output_dim = output_dim.value
if self.hidden_dim is None:
self.trans = self.add_weight(name='trans',
shape=(output_dim, output_dim),
initializer='glorot_uniform',
trainable=True)
if self.lr_multiplier != 1:
K.set_value(self.trans,
K.eval(self.trans) / self.lr_multiplier)
self.trans = self.lr_multiplier * self.trans
else:
self.l_trans = self.add_weight(name='l_trans',
shape=(output_dim, self.hidden_dim),
initializer='glorot_uniform',
trainable=True)
self.r_trans = self.add_weight(name='r_trans',
shape=(output_dim, self.hidden_dim),
initializer='glorot_uniform',
trainable=True)
if self.lr_multiplier != 1:
K.set_value(self.l_trans,
K.eval(self.l_trans) / self.lr_multiplier)
self.l_trans = self.lr_multiplier * self.l_trans
K.set_value(self.r_trans,
K.eval(self.r_trans) / self.lr_multiplier)
self.r_trans = self.lr_multiplier * self.r_trans
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 new_update(x, new_x):
if is_one_of(x, params) and self._do_layer_adaptation(x):
dx = new_x - x
lr_t = K.clip(self.learning_rate, K.epsilon(), 1e10)
x_norm = tf.norm(x)
g_norm = tf.norm(dx / lr_t)
ratio = K.switch(
x_norm > 0.,
K.switch(g_norm > K.epsilon(), x_norm / g_norm, 1.),
1.)
new_x = x + dx * ratio
return old_update(x, new_x)
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.),
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 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 new_update(x, new_x):
if x is var and self._do_layer_adaptation(x):
dx = new_x - x
lr_t = self._decayed_lr(x.dtype.base_dtype)
lr_t = K.clip(lr_t, K.epsilon(), 1e10)
x_norm = tf.norm(x)
g_norm = tf.norm(dx / lr_t)
ratio = K.switch(
x_norm > 0.,
K.switch(g_norm > K.epsilon(), x_norm / g_norm, 1.),
1.)
new_x = x + dx * ratio
return old_update(x, new_x)
def masked_cross_entropy(y_true, y_pred):
"""交叉熵作为loss，并mask掉padding部分的预测
"""
y_true = K.reshape(y_true, [K.shape(y_true)[0], -1])
y_mask = K.cast(K.not_equal(y_true, 0), K.floatx())
cross_entropy = K.sparse_categorical_crossentropy(y_true, y_pred)
cross_entropy = K.sum(cross_entropy * y_mask) / K.sum(y_mask)
return cross_entropy
def __init__(self,
center=True,
scale=True,
conditional=False,
hidden_units=None,
hidden_activation='linear',
hidden_initializer='glorot_uniform',
**kwargs):
super(LayerNormalization, self).__init__(**kwargs)
self.center = center
self.scale = scale
self.conditional = conditional
self.hidden_units = hidden_units
self.hidden_activation = activations.get(hidden_activation)
self.hidden_initializer = initializers.get(hidden_initializer)
self.epsilon = K.epsilon() * K.epsilon()
def _resource_apply_op(self, grad, var, indices=None):
op = super(new_optimizer,
self)._resource_apply_op(grad, var, indices)
k, alpha = self.steps_per_slow_update, self.slow_step_size
cond = K.equal(self.iterations % k, 0)
slow_var = self.get_slot(var, 'slow_var')
slow_var_t = slow_var + alpha * (var - slow_var)
with tf.control_dependencies([op]):
slow_update = K.update(slow_var,
K.switch(cond, slow_var_t, slow_var))
with tf.control_dependencies([slow_update]):
copy_update = K.update(var, K.switch(cond, slow_var, var))
return copy_update
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 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(new_optimizer, 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 call(self, inputs):
if not hasattr(self, 'kernel'):
embedding_layer = search_layer(inputs, self.embedding_name)
if embedding_layer is None:
raise Exception('Embedding layer not found')
self.kernel = K.transpose(embedding_layer.embeddings)
self.units = K.int_shape(self.kernel)[1] # <<<
if not hasattr(self, 'bias') and self.use_bias: # <<<
self.bias = self.add_weight(name='bias',
shape=(self.units, ),
initializer='zeros')
outputs = K.dot(inputs, self.kernel)
if self.use_bias:
outputs = K.bias_add(outputs, self.bias)
outputs = self.activation(outputs)
return outputs
def train_function(inputs): # 重新定义训练函数
grads = embedding_gradients(inputs)[0] # Embedding梯度
delta = epsilon * grads / (np.sqrt((grads**2).sum()) + 1e-8) # 计算扰动
K.set_value(embeddings, K.eval(embeddings) + delta) # 注入扰动
outputs = old_train_function(inputs) # 梯度下降
K.set_value(embeddings, K.eval(embeddings) - delta) # 删除扰动
return outputs
def adversarial_training(model, embedding_name, epsilon=1):
"""给模型添加对抗训练
其中model是需要添加对抗训练的keras模型，embedding_name
则是model里边Embedding层的名字。要在模型compile之后使用。
"""
if model.train_function is None: # 如果还没有训练函数
model._make_train_function() # 手动make
old_train_function = model.train_function # 备份旧的训练函数
# 查找Embedding层
for output in model.outputs:
embedding_layer = search_layer(output, embedding_name)
if embedding_layer is not None:
break
if embedding_layer is None:
raise Exception('Embedding layer not found')
# 求Embedding梯度
embeddings = embedding_layer.embeddings # Embedding矩阵
gradients = K.gradients(model.total_loss, [embeddings]) # Embedding梯度
gradients = K.zeros_like(embeddings) + gradients[0] # 转为dense tensor
# 封装为函数
inputs = (model._feed_inputs + model._feed_targets +
model._feed_sample_weights) # 所有输入层
embedding_gradients = K.function(
inputs=inputs,
outputs=[gradients],
name='embedding_gradients',
) # 封装为函数
def train_function(inputs): # 重新定义训练函数
grads = embedding_gradients(inputs)[0] # Embedding梯度
delta = epsilon * grads / (np.sqrt((grads**2).sum()) + 1e-8) # 计算扰动
K.set_value(embeddings, K.eval(embeddings) + delta) # 注入扰动
outputs = old_train_function(inputs) # 梯度下降
K.set_value(embeddings, K.eval(embeddings) - delta) # 删除扰动
return outputs
model.train_function = train_function # 覆盖原训练函数
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 __init__(self,
learning_rate=0.001,
beta_1=0.9,
beta_2=0.999,
epsilon=1e-6,
bias_correction=True,
name='Adam',
**kwargs):
kwargs['name'] = name
super(Adam, self).__init__(**kwargs)
self._set_hyper('learning_rate', learning_rate)
self._set_hyper('beta_1', beta_1)
self._set_hyper('beta_2', beta_2)
self.epsilon = epsilon or K.epislon()
self.bias_correction = bias_correction
def set_model(self, model):
"""绑定模型，并初始化参数
"""
super(ExponentialMovingAverage, self).set_model(model)
self.ema_weights = [K.zeros(K.shape(w)) for w in model.weights]
self.old_weights = K.batch_get_value(model.weights)
K.batch_set_value(zip(self.ema_weights, self.old_weights))
self.updates = []
for w1, w2 in zip(self.ema_weights, model.weights):
op = K.moving_average_update(w1, w2, self.momentum)
self.updates.append(op)
def get_updates(self, loss, params):
updates = super(new_optimizer, self).get_updates(loss, params)
k, alpha = self.steps_per_slow_update, self.slow_step_size
cond = K.equal(self.iterations % k, 0)
slow_vars = [
K.zeros(K.int_shape(p),
dtype=K.dtype(p),
name='slow_var_%s' % i) for i, p in enumerate(params)
]
with tf.control_dependencies(updates):
slow_updates = [
K.update(q, K.switch(cond, q + alpha * (p - q), q))
for p, q in zip(params, slow_vars)
]
with tf.control_dependencies(slow_updates):
copy_updates = [
K.update(p, K.switch(cond, q, p))
for p, q in zip(params, slow_vars)
]
return copy_updates
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
def dense_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one hot形式
"""
y_true = K.argmax(y_true, 2)
return self.sparse_accuracy(y_true, y_pred)
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
if self.input_mask is None:
mask = None
else:
mask = K.cast(self.input_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())
if mask is None:
return K.mean(isequal)
else:
return K.sum(isequal * mask) / K.sum(mask)
def dense_loss(self, y_true, y_pred):
"""y_true需要是one hot形式
"""
y_true = K.argmax(y_true, 2)
return self.sparse_loss(y_true, y_pred)
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
if self.input_mask is None:
mask = None
else:
mask = K.cast(self.input_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)
if mask is None:
return K.mean(loss)
else:
return K.sum(loss * mask) / K.sum(mask)
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 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
if self.input_mask is None:
mask = None
else:
mask = K.cast(self.input_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())
if mask is None:
return K.mean(isequal)
else:
return K.sum(isequal * mask) / K.sum(mask)
