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 _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 get_updates(self, loss, params):
grads = self.get_gradients(loss, params)
self.updates = [K.update_add(self.iterations, 1)]
self.weights = [self.iterations]
lr = self.learning_rate
for i, (p, g) in enumerate(zip(params, grads)):
g2 = K.square(g) + self.epsilon1
shape, dtype = K.int_shape(p), K.dtype(p)
factored_shape = self.factored_shape(shape)
if factored_shape is None:
# 定义参数
v = K.zeros(shape, dtype=dtype, name='v_' + str(i))
self.weights.append(v)
# 定义更新
v_t = self.beta2 * v + (1.0 - self.beta2) * g2
self.updates.append(K.update(v, v_t))
else:
# 定义参数
shape1, axis1, shape2, axis2 = factored_shape
vr = K.zeros(shape1, dtype=dtype, name='vr_' + str(i))
vc = K.zeros(shape2, dtype=dtype, name='vc_' + str(i))
self.weights.extend([vr, vc])
# 定义更新
vr_t = self.beta2 * vr + K.mean(g2, axis=axis1, keepdims=True)
vc_t = self.beta2 * vc + K.mean(g2, axis=axis2, keepdims=True)
self.updates.extend([K.update(vr, vr_t), K.update(vc, vc_t)])
# 合成矩阵
v_t = vr_t * vc_t / K.mean(vr_t, axis=axis2, keepdims=True)
# 增量主体
u = g / K.sqrt(v_t)
# 增量裁剪
if self.clipping_threshold is not None:
u_rms = K.mean(K.sum(K.square(u)))
d = self.clipping_threshold
u = u / K.maximum(1.0, u_rms / d)
# 增量滑动
if self.beta1 > 0.0:
# 定义参数
m = K.zeros(shape, dtype=dtype, name='m_' + str(i))
self.weights.append(m)
# 定义更新
m_t = self.beta1 * m + (1.0 - self.beta1) * u
self.updates.append(K.update(m, m_t))
u = m_t
# 增量调整
if self.multiply_by_parameter_scale:
u = u * K.maximum(K.mean(K.sum(K.square(p))), self.epsilon2)
# 更新参数
self.updates.append(K.update(p, p - lr * u))
return self.updates
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 _resource_apply(self, grad, var, indices=None):
lr = self.learning_rate
g2 = K.square(grad) + self.epsilon1
shape = K.int_shape(var)
factored_shape = self.factored_shape(shape)
if factored_shape is None:
v = self.get_slot(var, 'v')
# 定义更新
v_t = self.beta2 * v + (1.0 - self.beta2) * g2
v_t = K.update(v, v_t)
else:
shape1, axis1, shape2, axis2 = factored_shape
vr = self.get_slot(var, 'vr')
vc = self.get_slot(var, 'vc')
# 定义更新
vr_t = self.beta2 * vr + K.mean(g2, axis=axis1, keepdims=True)
vc_t = self.beta2 * vc + K.mean(g2, axis=axis2, keepdims=True)
vr_t, vc_t = K.update(vr, vr_t), K.update(vc, vc_t)
# 合成矩阵
v_t = vr_t * vc_t / K.mean(vr_t, axis=axis2, keepdims=True)
# 增量主体
u = grad / K.sqrt(v_t)
# 增量裁剪
if self.clipping_threshold is not None:
u_rms = K.mean(K.sum(K.square(u)))
d = self.clipping_threshold
u = u / K.maximum(1.0, u_rms / d)
# 增量滑动
if self.beta1 > 0.0:
m = self.get_slot(var, 'm')
# 定义更新
m_t = self.beta1 * m + (1.0 - self.beta1) * u
u = K.update(m, m_t)
# 增量调整
if self.multiply_by_parameter_scale:
u = u * K.maximum(K.mean(K.sum(K.square(var))), self.epsilon2)
# 更新参数
return K.update(var, var - lr * u)