Python state_ops.assign_sub函数代码示例

  def _apply_sparse(self, grad, var):
    beta1_power = math_ops.cast(self._beta1_power, var.dtype.base_dtype)
    beta2_power = math_ops.cast(self._beta2_power, var.dtype.base_dtype)
    lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
    beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
    beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
    epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)
    lr = (lr_t * math_ops.sqrt(1 - beta2_power) / (1 - beta1_power))
    # m_t = beta1 * m + (1 - beta1) * g_t
    m = self.get_slot(var, "m")
    m_scaled_g_values = grad.values * (1 - beta1_t)
    m_t = state_ops.assign(m, m * beta1_t,
                           use_locking=self._use_locking)
    m_t = state_ops.scatter_add(m_t, grad.indices, m_scaled_g_values,
                               use_locking=self._use_locking)



    # u_t = max(beta_2 * u_{t-1}, L1(g_t))
    # theta_t = theta_{t-1} - alpha/(1-beta_1).m_t/u_t

    v = self.get_slot(var, "v")
    g_abs_values = tensorflow.abs(g_t)
    v_t = state_ops.assign(v, v * beta_2, use_locking = self._use_locking)
    v_t = state_ops.assign_max(v_t, grad.indices, g_abs_values,
                             use_locking=self._use_locking)
    var_update = state_ops.assign_sub(var,
                                      lr*m_t/(v_t*(1 - beta_1)),
                                      use_locking=self._use_locking)

    return control_flow_ops.group(*[var_update, m_t, v_t])

  def testReadWrite(self):
    """Tests initialization, reading, and writing a resource variable."""
    for dtype in self.numeric_types:
      with self.test_session() as session:
        with self.test_scope():
          with variable_scope.variable_scope("ascope", use_resource=True):
            x = variable_scope.get_variable(
                "x",
                shape=[],
                dtype=dtype,
                initializer=init_ops.constant_initializer(2))
            a = x.read_value()
            with ops.control_dependencies([a]):
              b = state_ops.assign(x, dtype(47))
            with ops.control_dependencies([b]):
              c = x.read_value()
            with ops.control_dependencies([c]):
              d = state_ops.assign_add(x, np.array(6 + 2j).astype(dtype))
            with ops.control_dependencies([d]):
              e = state_ops.assign_sub(x, dtype(3))
            with ops.control_dependencies([e]):
              f = x.read_value()

        session.run(variables.global_variables_initializer())
        v1, v2, v3 = session.run([a, c, f])
        self.assertAllClose(dtype(2), v1)
        self.assertAllClose(dtype(47), v2)
        self.assertAllClose(np.array(50 + 2j).astype(dtype), v3)

  def _resource_apply_sparse(self, grad, var, indices):
    var_dtype = var.dtype.base_dtype
    lr_t = self._decayed_lr(var_dtype)
    beta_1_t = self._get_hyper('beta_1', var_dtype)
    beta_2_t = self._get_hyper('beta_2', var_dtype)
    local_step = math_ops.cast(self.iterations + 1, var_dtype)
    beta_1_power = math_ops.pow(beta_1_t, local_step)
    beta_2_power = math_ops.pow(beta_2_t, local_step)
    epsilon_t = self._get_hyper('epsilon', var_dtype)
    lr = (lr_t * math_ops.sqrt(1 - beta_2_power) / (1 - beta_1_power))

    # m_t = beta1 * m + (1 - beta1) * g_t
    m = self.get_slot(var, 'm')
    m_scaled_g_values = grad * (1 - beta_1_t)
    m_t = state_ops.assign(m, m * beta_1_t, use_locking=self._use_locking)
    with ops.control_dependencies([m_t]):
      m_t = self._resource_scatter_add(m, indices, m_scaled_g_values)

    # v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
    v = self.get_slot(var, 'v')
    v_scaled_g_values = (grad * grad) * (1 - beta_2_t)
    v_t = state_ops.assign(v, v * beta_2_t, use_locking=self._use_locking)
    with ops.control_dependencies([v_t]):
      v_t = self._resource_scatter_add(v, indices, v_scaled_g_values)

    v_sqrt = math_ops.sqrt(v_t)
    var_update = state_ops.assign_sub(
        var, lr * m_t / (v_sqrt + epsilon_t), use_locking=self._use_locking)
    return control_flow_ops.group(*[var_update, m_t, v_t])

 def _apply_sparse_shared(self, grad, var, indices, scatter_add):
   beta1_power = math_ops.cast(self._beta1_power, var.dtype.base_dtype)
   beta2_power = math_ops.cast(self._beta2_power, var.dtype.base_dtype)
   lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
   beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
   beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
   epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)
   lr = (lr_t * math_ops.sqrt(1 - beta2_power) / (1 - beta1_power))
   # m_t = beta1 * m + (1 - beta1) * g_t
   m = self.get_slot(var, "m")
   m_scaled_g_values = grad * (1 - beta1_t)
   m_t = state_ops.assign(m, m * beta1_t,
                          use_locking=self._use_locking)
   with ops.control_dependencies([m_t]):
     m_t = scatter_add(m, indices, m_scaled_g_values)
   # v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
   v = self.get_slot(var, "v")
   v_scaled_g_values = (grad * grad) * (1 - beta2_t)
   v_t = state_ops.assign(v, v * beta2_t, use_locking=self._use_locking)
   with ops.control_dependencies([v_t]):
     v_t = scatter_add(v, indices, v_scaled_g_values)
   v_sqrt = math_ops.sqrt(v_t)
   var_update = state_ops.assign_sub(var,
                                     lr * m_t / (v_sqrt + epsilon_t),
                                     use_locking=self._use_locking)
   return control_flow_ops.group(*[var_update, m_t, v_t])

 def _Update_global_variables():
   local_vars = [v for g, v in grads_and_vars if g is not None]
   global_center_vars = [self._global_map[var] for var in local_vars]
   local_center_vars = [self._local_map[var] for var in local_vars]
   local_center_vars_update = []
   for lvar, var in zip(local_center_vars, global_center_vars):
     local_center_vars_update.append(lvar.assign(var))
   update_ops = []
   differences = []
   with ops.control_dependencies(local_center_vars_update):
     for v, lv in zip(local_vars, local_center_vars):
       with ops.device(v.device):
         differences.append(math_ops.subtract(v, lv))
     for lvar, diff in zip(local_vars, differences):
       with ops.device(lvar.device):
         update_ops.append(
             state_ops.assign_sub(lvar,
                                  math_ops.multiply(self._moving_rate,
                                                    diff)))
     for var, diff in zip(global_center_vars, differences):
       with ops.device(var.device):
         update_ops.append(
             state_ops.assign_add(var,
                                  math_ops.multiply(self._moving_rate,
                                                    diff)))
     if global_step:
       with ops.colocate_with(global_step):
         update_ops.append(state_ops.assign_add(global_step, 1))
   variable_update = control_flow_ops.group(*(update_ops))
   return variable_update

def assign_moving_average(variable, value, decay, name=None):
  """Compute the moving average of a variable.

  The moving average of 'variable' updated with 'value' is:
    variable * decay + value * (1 - decay)

  The returned Operation sets 'variable' to the newly computed moving average.

  The new value of 'variable' can be set with the 'AssignSub' op as:
     variable -= (1 - decay) * (variable - value)

  Args:
    variable: A Variable.
    value: A tensor with the same shape as 'variable'
    decay: A float Tensor or float value.  The moving average decay.
    name: Optional name of the returned operation.

  Returns:
    An Operation that updates 'variable' with the newly computed
    moving average.
  """
  with ops.op_scope([variable, value, decay], name, "AssignMovingAvg") as scope:
    with ops.device(variable.device):
      decay = ops.convert_to_tensor(1.0 - decay, name="decay")
      if decay.dtype != variable.dtype.base_dtype:
        decay = math_ops.cast(decay, variable.dtype.base_dtype)
      return state_ops.assign_sub(variable, (variable - value) * decay,
                                  name=scope)

def assign_moving_average(variable, value, decay, zero_debias=True, name=None):
  """Compute the moving average of a variable.

  The moving average of 'variable' updated with 'value' is:
    variable * decay + value * (1 - decay)

  The returned Operation sets 'variable' to the newly computed moving average.

  The new value of 'variable' can be set with the 'AssignSub' op as:
     variable -= (1 - decay) * (variable - value)

  Since variables that are initialized to a `0` value will be `0` biased,
  `zero_debias` optionally enables scaling by the mathematically correct
  debiasing factor of
    1 - decay ** num_updates
  See `ADAM: A Method for Stochastic Optimization` Section 3 for more details
  (https://arxiv.org/abs/1412.6980).

  The names of the debias shadow variables, by default, include both the scope
  they were created in and the scope of the variables they debias. They are also
  given a uniqifying-suffix.

  E.g.:

  ```
    with tf.variable_scope('scope1'):
      with tf.variable_scope('scope2'):
        var = tf.get_variable('foo')
        tf.assign_moving_average(var, 0.0, 1.0)
        tf.assign_moving_average(var, 0.0, 0.9)

    # var.name: 'scope1/scope2/foo'
    # shadow var names: 'scope1/scope2/scope1/scope2/foo/biased'
    #                   'scope1/scope2/scope1/scope2/foo/biased_1'
  ```

  Args:
    variable: A Variable.
    value: A tensor with the same shape as 'variable'.
    decay: A float Tensor or float value.  The moving average decay.
    zero_debias: A python bool. If true, assume the variable is 0-initialized
      and unbias it, as in https://arxiv.org/abs/1412.6980. See docstring in
      `_zero_debias` for more details.
    name: Optional name of the returned operation.

  Returns:
    A reference to the input 'variable' tensor with the newly computed
    moving average.
  """
  with ops.name_scope(name, "AssignMovingAvg",
                      [variable, value, decay]) as scope:
    with ops.colocate_with(variable):
      decay = ops.convert_to_tensor(1.0 - decay, name="decay")
      if decay.dtype != variable.dtype.base_dtype:
        decay = math_ops.cast(decay, variable.dtype.base_dtype)
      if zero_debias:
        update_delta = _zero_debias(variable, value, decay)
      else:
        update_delta = (variable - value) * decay
      return state_ops.assign_sub(variable, update_delta, name=scope)

    def _apply_dense(self, grad, var):
        beta1_power = math_ops.cast(self._beta1_power, var.dtype.base_dtype)
        beta2_power = math_ops.cast(self._beta2_power, var.dtype.base_dtype)
        lr_t = math_ops.cast(self._lr_t, var.dtype.base_dtype)
        beta1_t = math_ops.cast(self._beta1_t, var.dtype.base_dtype)
        beta2_t = math_ops.cast(self._beta2_t, var.dtype.base_dtype)
        epsilon_t = math_ops.cast(self._epsilon_t, var.dtype.base_dtype)

        lr = (lr_t * math_ops.sqrt(1 - beta2_power) / (1 - beta1_power))

        # m_t = beta1 * m + (1 - beta1) * g_t
        m = self.get_slot(var, "m")
        m_scaled_g_values = grad * (1 - beta1_t)
        m_t = state_ops.assign(m, beta1_t * m + m_scaled_g_values, use_locking=self._use_locking)

        # v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
        v = self.get_slot(var, "v")
        v_scaled_g_values = (grad * grad) * (1 - beta2_t)
        v_t = state_ops.assign(v, beta2_t * v + v_scaled_g_values, use_locking=self._use_locking)

        # amsgrad
        vhat = self.get_slot(var, "vhat")
        vhat_t = state_ops.assign(vhat, math_ops.maximum(v_t, vhat))
        v_sqrt = math_ops.sqrt(vhat_t)

        var_update = state_ops.assign_sub(var, lr * m_t / (v_sqrt + epsilon_t), use_locking=self._use_locking)
        return control_flow_ops.group(*[var_update, m_t, v_t, vhat_t])

 def testVarChange(self):
   with imperative_mode.ImperativeMode(self._target) as mode:
     x = variables.Variable(constant_op.constant(1.0))
     for i in range(10):
       with mode.new_step() as step:
         step.run(state_ops.assign_sub(x, 0.1))
         self.assertAllClose(array_ops.identity(x).value, 1.0 - (i + 1) * 0.1)

 def _assign_moving_average(self, variable, value, momentum):
   with ops.name_scope(None, 'AssignMovingAvg',
                       [variable, value, momentum]) as scope:
     decay = ops.convert_to_tensor(1.0 - momentum, name='decay')
     if decay.dtype != variable.dtype.base_dtype:
       decay = math_ops.cast(decay, variable.dtype.base_dtype)
     update_delta = (variable - value) * decay
     return state_ops.assign_sub(variable, update_delta, name=scope)

def assign_moving_average(variable, value, decay, name=None):
    with ops.op_scope([variable, value, decay], name, "AssignMovingAvg") as name:
        with ops.device(variable.device):
            decay = ops.convert_to_tensor(1.0 - decay, name="decay")
            if decay.dtype != variable.dtype.base_dtype:
                decay = math_ops.cast(decay, variable.dtype.base_dtype)
            return state_ops.assign_sub(variable, (variable - value) * decay,
                                        name=name)

 def _initAssignSubFetch(self, x, y, use_gpu=False):
   """Initialize a param to init, and compute param -= y."""
   with self.test_session(use_gpu=use_gpu):
     p = variables.Variable(x)
     sub = state_ops.assign_sub(p, y)
     p.initializer.run()
     new_value = sub.eval()
     return p.eval(), new_value

 def update_fn(v, value, decay=decay):
   decay = ops.convert_to_tensor(1.0 - decay, name="decay")
   if decay.dtype != v.dtype.base_dtype:
     decay = math_ops.cast(decay, v.dtype.base_dtype)
   if zero_debias:
     update_delta = _zero_debias(v, value, decay)
   else:
     update_delta = (v - value) * decay
   return state_ops.assign_sub(v, update_delta, name=scope)

  def update_fn(v, value, biased_var, local_step):
    update_biased = state_ops.assign_sub(biased_var,
                                         (biased_var - value) * decay)
    update_local_step = local_step.assign_add(1)

    # This function gets `1 - decay`, so use `1.0 - decay` in the exponent.
    bias_factor = 1 - math_ops.pow(1.0 - decay, update_local_step)
    return state_ops.assign(
        v, update_biased / bias_factor, name=ops.get_name_scope() + "/")

    def _apply_dense(self, grad, var):
        lr = self._lr_t * math_ops.sqrt(1 - self._beta2_power) / (1 - self._beta1_power)
        # m_t = beta1 * m + (1 - beta1) * g_t
        m = self.get_slot(var, "m")
        m_scaled_g_values = grad * (1 - self._beta1_t)
        m_t = m * self._beta1_t
        m_t = m_t + m_scaled_g_values
        # v_t = beta2 * v + (1 - beta2) * (g_t * g_t)
        v = self.get_slot(var, "v")
        v_scaled_g_values = tf.pow(grad, 2) * (1 - self._beta2_t)
        v_t = v * self._beta2_t
        v_t = v_t + v_scaled_g_values
        v_sqrt = tf.pow(v_t, self._pow_t)
        var_update = state_ops.assign_sub(var, lr * m_t / (v_sqrt + self._epsilon_t), use_locking=self._use_locking)
        # regularization
        var_update = state_ops.assign_sub(var_update, self._dense_regularization * var, use_locking=self._use_locking)

        return control_flow_ops.group(*[var_update, m_t, v_t])