Python v1.assign_add方法代码示例

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import assign_add [as 别名]
def weight_noise(noise_rate, learning_rate, var_list):
  """Apply weight noise to vars in var_list."""
  if not noise_rate:
    return [tf.no_op()]

  tf.logging.info("Applying weight noise scaled by learning rate, "
                  "noise_rate: %0.5f", noise_rate)

  noise_ops = []

  for v in var_list:
    with tf.device(v.device):  # pylint: disable=protected-access
      scale = noise_rate * learning_rate * 0.001
      if common_layers.should_generate_summaries():
        tf.summary.scalar("weight_noise_scale", scale)
      noise = tf.truncated_normal(v.shape) * scale
      noise_op = v.assign_add(noise)
      noise_ops.append(noise_op)

  return noise_ops 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import assign_add [as 别名]
def _apply_cond(self, apply_fn, grad, var, *args, **kwargs):
    """Apply conditionally if counter is zero."""
    grad_acc = self.get_slot(var, "grad_acc")

    def apply_adam(grad_acc, apply_fn, grad, var, *args, **kwargs):
      total_grad = (grad_acc + grad) / tf.cast(self._n_t, grad.dtype)
      adam_op = apply_fn(total_grad, var, *args, **kwargs)
      with tf.control_dependencies([adam_op]):
        grad_acc_to_zero_op = grad_acc.assign(
            tf.zeros_like(grad_acc), use_locking=self._use_locking)
      return tf.group(adam_op, grad_acc_to_zero_op)

    def accumulate_gradient(grad_acc, grad):
      assign_op = tf.assign_add(grad_acc, grad, use_locking=self._use_locking)
      return tf.group(assign_op)  # Strip return value

    return tf.cond(
        tf.equal(self._get_iter_variable(), 0),
        lambda: apply_adam(grad_acc, apply_fn, grad, var, *args, **kwargs),
        lambda: accumulate_gradient(grad_acc, grad)) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import assign_add [as 别名]
def _apply_cond(self, apply_fn, grad, var, *args, **kwargs):
    """Apply conditionally if counter is zero."""
    grad_acc = self.get_slot(var, "grad_acc")

    def apply_adam(grad_acc, apply_fn, grad, var, *args, **kwargs):
      total_grad = (grad_acc + grad) / tf.cast(self._n_t, grad.dtype)
      adam_op = apply_fn(total_grad, var, *args, **kwargs)
      with tf.control_dependencies([adam_op]):
        grad_acc_to_zero_op = grad_acc.assign(tf.zeros_like(grad_acc),
                                              use_locking=self._use_locking)
      return tf.group(adam_op, grad_acc_to_zero_op)

    def accumulate_gradient(grad_acc, grad):
      assign_op = tf.assign_add(grad_acc, grad, use_locking=self._use_locking)
      return tf.group(assign_op)  # Strip return value

    return tf.cond(
        tf.equal(self._get_iter_variable(), 0),
        lambda: apply_adam(grad_acc, apply_fn, grad, var, *args, **kwargs),
        lambda: accumulate_gradient(grad_acc, grad)) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import assign_add [as 别名]
def _build_select_slate_op(self):
    p_no_click = self._prob_no_click_ph
    p = self._doc_affinity_scores_ph
    q = self._net_outputs.q_values[0]
    with tf.name_scope('select_slate'):
      self._output_slate = self._select_slate_fn(self._slate_size, p_no_click,
                                                 p, q)

    self._output_slate = tf.Print(
        self._output_slate, [tf.constant('cp 1'), self._output_slate, p, q],
        summarize=10000)
    self._output_slate = tf.reshape(self._output_slate, (self._slate_size,))

    self._action_counts = tf.get_variable(
        'action_counts',
        shape=[self._num_candidates],
        initializer=tf.zeros_initializer())
    output_slate = tf.reshape(self._output_slate, [-1])
    output_one_hot = tf.one_hot(output_slate, self._num_candidates)
    update_ops = []
    for i in range(self._slate_size):
      update_ops.append(tf.assign_add(self._action_counts, output_one_hot[i]))
    self._select_action_update_op = tf.group(*update_ops) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import assign_add [as 别名]
def _make_take_sample(self):
    assignments = []
    n = tf.cast(self._num_samples, tf.float32)
    mu = 1.0 / (1.0 + n)
    for tensor, average in zip(self._tensors, self._averages):
      assignments.append(tf.assign_add(average, (tensor-average)*mu))
    add_to_averages = tf.group(assignments)
    with tf.control_dependencies([add_to_averages]):
      incr_num_samples = tf.assign(self._num_samples, self._num_samples + 1)
    return incr_num_samples 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import assign_add [as 别名]
def _reset_non_empty(self, indices):
    """Reset the batch of environments.

    Args:
      indices: The batch indices of the environments to reset; defaults to all.

    Returns:
      Batch tensor of the new observations.
    """
    reset_video_op = tf.cond(
        self._video_condition,
        lambda: tf.py_func(self._video_reset_writer, [], []),
        tf.no_op)
    with tf.control_dependencies([reset_video_op]):
      inc_op = tf.assign_add(self._episode_counter, 1)
      with tf.control_dependencies([self.history_buffer.reset(indices),
                                    inc_op]):
        initial_frame_dump_op = tf.cond(
            self._video_condition,
            lambda: tf.py_func(self._video_dump_frames,  # pylint: disable=g-long-lambda
                               [self.history_buffer.get_all_elements()], []),
            tf.no_op)
        observ_assign_op = self._observ.assign(
            self.history_buffer.get_all_elements()[:, -1, ...])
        with tf.control_dependencies([observ_assign_op, initial_frame_dump_op]):
          reset_model_op = tf.assign(self._reset_model, tf.constant(1.0))
          with tf.control_dependencies([reset_model_op]):
            return tf.gather(self._observ.read_value(), indices) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import assign_add [as 别名]
def apply_gradients(self, grads_and_vars, global_step=None, name=None):
    if contrib.is_tf2:
      with tf.control_dependencies(
          [tf.assign_add(tf.train.get_or_create_global_step(), 1)]):
        return self._opt.apply_gradients(grads_and_vars, name=name)
    else:
      return self._opt.apply_gradients(
          grads_and_vars, global_step=global_step, name=name) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import assign_add [as 别名]
def testDiet(self):

    params = diet.diet_adam_optimizer_params()

    @diet.fn_with_diet_vars(params)
    def model_fn(x):
      y = tf.layers.dense(x, 10, use_bias=False)
      return y

    @diet.fn_with_diet_vars(params)
    def model_fn2(x):
      y = tf.layers.dense(x, 10, use_bias=False)
      return y

    x = tf.random_uniform((10, 10))
    y = model_fn(x) + 10.
    y = model_fn2(y) + 10.
    grads = tf.gradients(y, [x])
    with tf.control_dependencies(grads):
      incr_step = tf.assign_add(tf.train.get_or_create_global_step(), 1)

    train_op = tf.group(incr_step, *grads)
    with self.test_session() as sess:
      sess.run(tf.global_variables_initializer())
      orig_vals = sess.run(tf.global_variables())
      for _ in range(10):
        sess.run(train_op)
      new_vals = sess.run(tf.global_variables())

      different = []
      for old, new in zip(orig_vals, new_vals):
        try:
          self.assertAllClose(old, new)
        except AssertionError:
          different.append(True)
      self.assertEqual(len(different), len(tf.global_variables())) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import assign_add [as 别名]
def call_fake_controller(push_values, pop_values, write_values, output_values):
  """Mock a RNN controller from a set of expected outputs.

  Args:
    push_values: Expected controller push values.
    pop_values: Expected controller pop values.
    write_values: Expected controller write values.
    output_values: Expected controller output values.

  Returns:
    A callable which behaves like the call method of an NeuralStackCell.
  """
  def call(cell, inputs, prev_read_values, controller_state, batch_size):
    del inputs
    del prev_read_values
    del batch_size
    next_step = tf.constant(0)
    if hasattr(cell, "current_step"):
      next_step = tf.assign_add(cell.current_step, tf.constant(1))
    return neural_stack.NeuralStackControllerInterface(
        push_strengths=tf.slice(tf.constant(push_values),
                                [next_step, 0, 0, 0],
                                [1, -1, -1, -1]),
        pop_strengths=tf.slice(tf.constant(pop_values),
                               [next_step, 0, 0, 0],
                               [1, -1, -1, -1]),
        write_values=tf.slice(tf.constant(write_values),
                              [next_step, 0, 0],
                              [1, -1, -1]),
        outputs=tf.slice(tf.constant(output_values),
                         [next_step, 0, 0],
                         [1, -1, -1]),
        state=controller_state
    )
  return call 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import assign_add [as 别名]
def __init__(self, train_time, time_limit=None):
    super(TrainTimeHook, self).__init__()
    self._train_time = train_time
    self._time_limit = time_limit
    self._increment_amount = tf.placeholder(tf.float32, None)
    self._increment_op = tf.assign_add(train_time, self._increment_amount)
    self._last_run_duration = None 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import assign_add [as 别名]
def metric_sum(values, name=None, **kwargs):
  del kwargs
  with tf.variable_scope(name, "metric_sum", [values]):
    accum = tf.get_variable(
        "accum", shape=[], dtype=tf.float32, trainable=False,
        collections=[tf.GraphKeys.LOCAL_VARIABLES],
        initializer=tf.zeros_initializer())
    update_op = tf.assign_add(accum, tf.reduce_sum(tf.cast(values, tf.float32)))
    return accum, update_op 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import assign_add [as 别名]
def __init__(self,
               update_batchnorm_params=True):
    self.update_batchnorm_params = update_batchnorm_params

    num_samples = datasets.get_count(FLAGS.train_split)
    if FLAGS.num_supervised_examples:
      num_samples = FLAGS.num_supervised_examples
    steps_per_epoch = num_samples // FLAGS.batch_size
    self.steps_per_epoch = steps_per_epoch

    global_step = tf.train.get_or_create_global_step()
    self.global_step_inc = tf.assign_add(global_step, 1)

    # lr_scale_batch_size defines a canonical batch size that is coupled with
    # the initial learning rate. If actual batch size is not the same as
    # canonical than learning rate is linearly scaled. This is very convinient
    # as this allows to vary batch size without recomputing learning rate.
    lr_factor = 1.0
    if FLAGS.lr_scale_batch_size:
      lr_factor = FLAGS.batch_size / float(FLAGS.lr_scale_batch_size)

    # We actually also accept fractional epochs.
    schedule_in_steps = utils.get_schedule_from_config(
        FLAGS.schedule, steps_per_epoch)
    warmup, decays = schedule_in_steps[0], schedule_in_steps[1:-1]

    self.lr = get_lr(
        global_step,
        base_lr=FLAGS.lr * lr_factor,
        decay_steps=decays,
        lr_decay_factor=FLAGS.lr_decay_factor,
        warmup_steps=warmup)

    # TODO(marvinritter): Re-enable summaries with support for TPU training.
    # tf.summary.scalar('learning_rate', self.lr) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import assign_add [as 别名]
def testPeriodicTargetUpdate(self, use_locking, update_period):
    """Tests that the simple success case works as expected.

    This is an integration test. The periodically and update parts are
    unit-tested in the preceding.

    Args:
      use_locking: value for `periodic_target_update`'s `use_locking` argument.
      update_period: how often an update should happen.
    """
    target_variables = [tf.Variable(tf.zeros([1, 2]))]
    source_variables = [tf.Variable(tf.random_normal([1, 2]))]
    increment = tf.ones([1, 2])

    update_source_op = tf.assign_add(source_variables[0], increment)
    updated = target_update_ops.periodic_target_update(
        target_variables,
        source_variables,
        update_period=update_period,
        use_locking=use_locking)

    with self.test_session() as sess:
      sess.run(tf.global_variables_initializer())

      for step in range(3 * update_period):
        sess.run(update_source_op)
        sess.run(updated)
        targets, sources = sess.run([target_variables, source_variables])

        if step % update_period == 0:
          self.assertAllClose(targets, sources)
        else:
          self.assertNotAllClose(targets, sources) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import assign_add [as 别名]
def apply_gradients(self, grads_and_vars, global_step=None, name=None):
    """Applying gradients and tune hyperparams with YellowFin.

    Args:
      grads_and_vars: List of (gradient, variable) pairs as returned by
        compute_gradients().
      global_step: Optional Variable to increment by one after the
        variables have been updated.
      name:  Optional name for the returned operation. Default to the
        name passed to the Optimizer constructor.

    Returns:
        (A group of operations)
        Variable Update with Momentum ops,
        YellowFin ops(Curvature, Variance, Distance) ops,
        SingleStep and lr_mu tuning ops,
        Step increment ops.
    """
    self._grad, self._vars = zip(*[(g, t)
                                   for g, t in grads_and_vars if g is not None])

    # Var update with Momentum.
    with tf.variable_scope("apply_updates"):
      # Gradient Clipping?
      if self._clip_thresh_var is not None:
        self._grad, _ = tf.clip_by_global_norm(
            self._grad, self._clip_thresh_var)

        apply_grad_op = self._momentum_optimizer.apply_gradients(
            zip(self._grad, self._vars),
            global_step=global_step,
            name=name)
      else:
        apply_grad_op = self._momentum_optimizer.apply_gradients(
            zip(self._grad, self._vars),
            global_step=global_step,
            name=name)

    # Begin lr and mu tuning.
    with tf.variable_scope("prepare_yellowFin_variables"):
      # the dependencies ideally only need to be after clip is done,
      # i.e. depends on self._grads. However, the control_dependencies
      # does not support indexed slice for sparse gradients.
      # The alternative dependencies here might be slightly slower due
      # to less parallelization.
      with tf.control_dependencies([apply_grad_op,]):
        prepare_variables_op = self._prepare_variables()

    with tf.variable_scope("yellowfin"):
      with tf.control_dependencies([prepare_variables_op]):
        yellowfin_op = self._yellowfin()

    # Update YellowFin step variable.
    with tf.control_dependencies([yellowfin_op]):
      self._increment_step_op = tf.assign_add(self._step, 1).op

    return tf.group(apply_grad_op,
                    prepare_variables_op,
                    yellowfin_op,
                    self._increment_step_op) 

# 需要导入模块: from tensorflow.compat import v1 [as 别名]
# 或者: from tensorflow.compat.v1 import assign_add [as 别名]
def testEarlyStoppingHook(self):
    global_step = tf.train.create_global_step()
    counter = tf.get_variable("count", initializer=0, dtype=tf.int32)
    tf.summary.scalar("count", counter)
    incr_global_step = tf.assign_add(global_step, 1)
    incr_counter = tf.assign_add(counter, 1)

    # Stop if the global step has not gone up by more than 1 in 20 steps.

    ckpt_dir = self.ckpt_dir("early")
    stop_hook = metrics_hook.EarlyStoppingHook(
        ckpt_dir,
        "count_1",
        num_plateau_steps=20,
        plateau_delta=1.,
        plateau_decrease=False,
        every_n_steps=10)
    with self.sess(stop_hook, ckpt_dir) as sess:
      for _ in range(20):
        sess.run((incr_global_step, incr_counter))

      # Summary files should now have 2 values in them
      self.flush()

      # Run for more steps so that the hook gets triggered and we verify that we
      # don't stop.
      for _ in range(30):
        sess.run((incr_global_step, incr_counter))

      self.flush()

      # Run without incrementing the counter
      for _ in range(40):
        sess.run(incr_global_step)

      # Metrics should be written such that now the counter has gone >20 steps
      # without being incremented.
      self.flush()

      # Check that we ask for stop
      with self.assertRaisesRegexp(RuntimeError, "after should_stop requested"):
        for _ in range(30):
          sess.run(incr_global_step)