def _resource_apply(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(NewOptimizer, self)._resource_apply(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(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 _resource_apply(self, grad, var, indices=None):
op = super(NewOptimizer, self)._resource_apply(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 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 _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)
def get_updates(self, loss, params):
updates = super(NewOptimizer, self).get_updates(loss, params)
self.model_weights = params
self.ema_weights = [K.zeros(K.shape(w)) for w in params]
self.old_weights = K.batch_get_value(params)
K.batch_set_value(zip(self.ema_weights, self.old_weights))
ema_updates, ema_momentum = [], self.ema_momentum
with tf.control_dependencies(updates):
for w1, w2 in zip(self.ema_weights, params):
new_w = ema_momentum * w1 + (1 - ema_momentum) * w2
ema_updates.append(K.update(w1, new_w))
return ema_updates
def get_updates(self, loss, params):
updates = super(NewOptimizer, 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