Python component.MasterState方法代碼示例

# 需要導入模塊: from dragnn.python import component [as 別名]
# 或者: from dragnn.python.component import MasterState [as 別名]
def build_inference(self, handle, use_moving_average=False):
    """Builds an inference pipeline.

    This always uses the whole pipeline.

    Args:
      handle: Handle tensor for the ComputeSession.
      use_moving_average: Whether or not to read from the moving
        average variables instead of the true parameters. Note: it is not
        possible to make gradient updates when this is True.

    Returns:
      handle: Handle after annotation.
    """
    self.read_from_avg = use_moving_average
    network_states = {}

    for comp in self.components:
      network_states[comp.name] = component.NetworkState()
      handle = dragnn_ops.init_component_data(
          handle, beam_size=comp.inference_beam_size, component=comp.name)
      master_state = component.MasterState(handle,
                                           dragnn_ops.batch_size(
                                               handle, component=comp.name))
      with tf.control_dependencies([handle]):
        handle = comp.build_greedy_inference(master_state, network_states)
      handle = dragnn_ops.write_annotations(handle, component=comp.name)

    self.read_from_avg = False
    return handle 

# 需要導入模塊: from dragnn.python import component [as 別名]
# 或者: from dragnn.python.component import MasterState [as 別名]
def setUp(self):
    self.master = MockMaster()
    self.master_state = component.MasterState(
        handle='handle', current_batch_size=2)
    self.network_states = {
        'mock': component.NetworkState(),
        'test': component.NetworkState(),
    } 

# 需要導入模塊: from dragnn.python import component [as 別名]
# 或者: from dragnn.python.component import MasterState [as 別名]
def build_inference(self,
                      handle,
                      use_moving_average=False,
                      build_runtime_graph=False):
    """Builds an inference pipeline.

    This always uses the whole pipeline.

    Args:
      handle: Handle tensor for the ComputeSession.
      use_moving_average: Whether or not to read from the moving
        average variables instead of the true parameters. Note: it is not
        possible to make gradient updates when this is True.
      build_runtime_graph: Whether to build a graph for use by the runtime.

    Returns:
      handle: Handle after annotation.
    """
    self.read_from_avg = use_moving_average
    self.build_runtime_graph = build_runtime_graph
    network_states = {}

    for comp in self.components:
      network_states[comp.name] = component.NetworkState()
      handle = dragnn_ops.init_component_data(
          handle, beam_size=comp.inference_beam_size, component=comp.name)
      if build_runtime_graph:
        batch_size = 1  # runtime uses singleton batches
      else:
        batch_size = dragnn_ops.batch_size(handle, component=comp.name)
      master_state = component.MasterState(handle, batch_size)
      with tf.control_dependencies([handle]):
        handle = comp.build_greedy_inference(master_state, network_states)
      handle = dragnn_ops.write_annotations(handle, component=comp.name)

    self.read_from_avg = False
    self.build_runtime_graph = False
    return handle 

# 需要導入模塊: from dragnn.python import component [as 別名]
# 或者: from dragnn.python.component import MasterState [as 別名]
def setUp(self):
    # Clear the graph and all existing variables.  Otherwise, variables created
    # in different tests may collide with each other.
    tf.reset_default_graph()
    self.master = MockMaster()
    self.master_state = component.MasterState(
        handle=tf.constant(['foo', 'bar']), current_batch_size=2)
    self.network_states = {
        'mock': component.NetworkState(),
        'test': component.NetworkState(),
    }