Python ops.OpStats方法代码示例

# 需要导入模块: from tensorflow.python.framework import ops [as 别名]
# 或者: from tensorflow.python.framework.ops import OpStats [as 别名]
def _pool_flops(graph, node):
  """Common code which compute flops for pooling operations."""
  # compute flops for average and max pooling
  _verify_conv_data_format(node)
  #
  # Pooling declaration:
  #   Inputs:
  #     - value
  #   Outputs:
  #     - output
  #   Attributes:
  #     - ksize
  #     - strides
  #     - padding
  #     - data_format
  #
  # Pooling implenetation:
  out_shape = graph_util.tensor_shape_from_node_def_name(graph, node.name)
  out_shape.assert_is_fully_defined()
  kernel_shape = list(node.attr["ksize"].list.i)
  kernel_area = _list_product(kernel_shape)
  return ops.OpStats("flops", kernel_area * out_shape.num_elements()) 

# 需要导入模块: from tensorflow.python.framework import ops [as 别名]
# 或者: from tensorflow.python.framework.ops import OpStats [as 别名]
def _max_pool_grad_flops(graph, node):
  """Compute flops for MaxPoolGrad operation."""
  _verify_conv_data_format(node)
  #
  # MaxPoolGrad declaration:
  #   Inputs:
  #     - orig_input  -- original input tensor (of max_pool)
  #     - orig_output  -- original output tensor (of max_pool)
  #     - grad --  gradient with respect to output of max_pool
  #   Outputs:
  #     - output -- gradient with respect to input of max_pool
  #   Attributes:
  #     - ksize
  #     - strides
  #     - padding
  #     - data_format
  # It computes MaxPool first, then one flop per each element of original output
  #
  kernel_shape = list(node.attr["ksize"].list.i)
  kernel_area = _list_product(kernel_shape)
  orig_out_shape = graph_util.tensor_shape_from_node_def_name(graph,
                                                              node.input[1])
  max_pool_ops = kernel_area * orig_out_shape.num_elements()
  return ops.OpStats("flops", max_pool_ops + orig_out_shape.num_elements()) 

# 需要导入模块: from tensorflow.python.framework import ops [as 别名]
# 或者: from tensorflow.python.framework.ops import OpStats [as 别名]
def _conv_2d_backprop_filter_flops(graph, node):
  """Compute flops for Conv2DBackpropFilter operation."""
  # Formula same as for Conv2DBackpropInput:
  #  batch_size * image_x_dim * image_y_dim * kernel_x_dim * kernel_y_dim
  #  * input_depth * output_depth * 2 / (image_x_stride * image_x_stride)
  #
  _verify_conv_data_format(node)
  # image_shape = [batch_size, image_y_dim, image_x_dim, input_depth]
  image_shape = graph_util.tensor_shape_from_node_def_name(graph, node.input[0])
  image_shape.assert_is_fully_defined()
  # kernel_shape = [kernel_y_dim, kernel_x_dim, input_depth, output_depth]
  kernel_shape = graph_util.tensor_shape_from_node_def_name(graph, node.name)
  kernel_shape.assert_is_fully_defined()
  # strides
  strides_shape = list(node.attr["strides"].list.i)
  strides_product = strides_shape[1] * strides_shape[2]
  return ops.OpStats("flops",
                     (2 * image_shape.num_elements()
                      * kernel_shape.num_elements()
                      / (image_shape[-1].value * strides_product)))

################################################################################
# Other ops
################################################################################ 

# 需要导入模块: from tensorflow.python.framework import ops [as 别名]
# 或者: from tensorflow.python.framework.ops import OpStats [as 别名]
def _calc_conv_flops(graph, node):
  """Calculates the compute resources needed for Conv2D."""
  input_shape = graph_util.tensor_shape_from_node_def_name(graph, node.input[0])
  input_shape.assert_is_fully_defined()
  filter_shape = graph_util.tensor_shape_from_node_def_name(graph,
                                                            node.input[1])
  filter_shape.assert_is_fully_defined()
  output_shape = graph_util.tensor_shape_from_node_def_name(graph, node.name)
  output_shape.assert_is_fully_defined()
  filter_height = int(filter_shape[0])
  filter_width = int(filter_shape[1])
  filter_in_depth = int(filter_shape[2])
  output_count = np.prod(output_shape.as_list())
  return ops.OpStats("flops", (output_count * filter_in_depth * filter_height *
                               filter_width * 2)) 

# 需要导入模块: from tensorflow.python.framework import ops [as 别名]
# 或者: from tensorflow.python.framework.ops import OpStats [as 别名]
def _calc_depthwise_conv_flops(graph, node):
  """Calculates the compute resources needed for DepthwiseConv2dNative."""
  input_shape = graph_util.tensor_shape_from_node_def_name(graph, node.input[0])
  input_shape.assert_is_fully_defined()
  filter_shape = graph_util.tensor_shape_from_node_def_name(graph,
                                                            node.input[1])
  filter_shape.assert_is_fully_defined()
  output_shape = graph_util.tensor_shape_from_node_def_name(graph, node.name)
  output_shape.assert_is_fully_defined()
  filter_height = int(filter_shape[0])
  filter_width = int(filter_shape[1])
  output_count = np.prod(output_shape.as_list())
  return ops.OpStats("flops", (output_count * filter_height * filter_width * 2)) 

# 需要导入模块: from tensorflow.python.framework import ops [as 别名]
# 或者: from tensorflow.python.framework.ops import OpStats [as 别名]
def _calc_bias_add_flops(graph, node):
  """Calculates the computing needed for BiasAdd."""
  input_shape = graph_util.tensor_shape_from_node_def_name(graph, node.input[0])
  input_shape.assert_is_fully_defined()
  input_count = np.prod(input_shape.as_list())
  return ops.OpStats("flops", input_count) 

# 需要导入模块: from tensorflow.python.framework import ops [as 别名]
# 或者: from tensorflow.python.framework.ops import OpStats [as 别名]
def _calc_dilation2d_flops(graph, node):
  """Calculates the compute resources needed for Dilation2D."""
  input_shape = graph_util.tensor_shape_from_node_def_name(graph, node.input[0])
  input_shape.assert_is_fully_defined()
  filter_shape = graph_util.tensor_shape_from_node_def_name(graph,
                                                            node.input[1])
  filter_shape.assert_is_fully_defined()
  output_shape = graph_util.tensor_shape_from_node_def_name(graph, node.name)
  output_shape.assert_is_fully_defined()
  filter_height = int(filter_shape[0])
  filter_width = int(filter_shape[1])
  output_count = np.prod(output_shape.as_list())
  return ops.OpStats("flops", (output_count * filter_height * filter_width * 2)) 

# 需要导入模块: from tensorflow.python.framework import ops [as 别名]
# 或者: from tensorflow.python.framework.ops import OpStats [as 别名]
def _calc_mat_mul_flops(graph, node):
  """Calculates the compute resources needed for MatMul."""
  transpose_a = node.attr["transpose_a"].b
  a_shape = graph_util.tensor_shape_from_node_def_name(graph, node.input[0])
  a_shape.assert_is_fully_defined()
  if transpose_a:
    k = int(a_shape[0])
  else:
    k = int(a_shape[1])
  output_shape = graph_util.tensor_shape_from_node_def_name(graph, node.name)
  output_shape.assert_is_fully_defined()
  output_count = np.prod(output_shape.as_list())
  return ops.OpStats("flops", (k * output_count * 2)) 

# 需要导入模块: from tensorflow.python.framework import ops [as 别名]
# 或者: from tensorflow.python.framework.ops import OpStats [as 别名]
def testAccumulateStatistics(self):
    flops_total = ops.OpStats("flops")
    self.assertEqual(None, flops_total.value)
    second_flops = ops.OpStats("flops", 3)
    flops_total += second_flops
    self.assertEqual(3, flops_total.value) 

# 需要导入模块: from tensorflow.python.framework import ops [as 别名]
# 或者: from tensorflow.python.framework.ops import OpStats [as 别名]
def calculate_graph_metrics(graph_def, statistic_types, input_layer,
                            input_shape_override, batch_size):
  """Looks at the performance statistics of all nodes in the graph."""
  _ = tf.import_graph_def(graph_def, name="")
  total_stats = {}
  node_stats = {}
  for statistic_type in statistic_types:
    total_stats[statistic_type] = ops.OpStats(statistic_type)
    node_stats[statistic_type] = {}
  # Make sure we get pretty-printed numbers with separators.
  locale.setlocale(locale.LC_ALL, "")
  with tf.Session() as sess:
    input_tensor = sess.graph.get_tensor_by_name(input_layer)
    input_shape_tensor = input_tensor.get_shape()
    if input_shape_tensor:
      input_shape = input_shape_tensor.as_list()
    else:
      input_shape = None
    if input_shape_override:
      input_shape = input_shape_override
    if input_shape is None:
      raise ValueError("""No input shape was provided on the command line,"""
                       """ and the input op itself had no default shape, so"""
                       """ shape inference couldn't be performed. This is"""
                       """ required for metrics calculations.""")
    input_shape[0] = batch_size
    input_tensor.set_shape(input_shape)
    for node in graph_def.node:
      # Ensure that the updated input shape has been fully-propagated before we
      # ask for the statistics, since they may depend on the output size.
      op = sess.graph.get_operation_by_name(node.name)
      ops.set_shapes_for_outputs(op)
      for statistic_type in statistic_types:
        current_stats = ops.get_stats_for_node_def(sess.graph, node,
                                                   statistic_type)
        node_stats[statistic_type][node.name] = current_stats
        total_stats[statistic_type] += current_stats
  return total_stats, node_stats 

# 需要导入模块: from tensorflow.python.framework import ops [as 别名]
# 或者: from tensorflow.python.framework.ops import OpStats [as 别名]
def _unary_op_flops(graph, node, ops_per_element=1):
  """Common code which compute flops for unary operations."""
  in_shape = graph_util.tensor_shape_from_node_def_name(graph, node.input[0])
  in_shape.assert_is_fully_defined()
  return ops.OpStats("flops", in_shape.num_elements() * ops_per_element) 

# 需要导入模块: from tensorflow.python.framework import ops [as 别名]
# 或者: from tensorflow.python.framework.ops import OpStats [as 别名]
def _l2_loss_flops(graph, node):
  """Compute flops for L2Loss operation."""
  in_shape = graph_util.tensor_shape_from_node_def_name(graph, node.input[0])
  in_shape.assert_is_fully_defined()
  # Tensorflow uses inefficient implementation, with (3*N-1) flops:
  # Optimal implementation is 2*N flops
  return ops.OpStats("flops", in_shape.num_elements() * 3 - 1) 

# 需要导入模块: from tensorflow.python.framework import ops [as 别名]
# 或者: from tensorflow.python.framework.ops import OpStats [as 别名]
def _binary_per_element_op_flops(graph, node, ops_per_element=1):
  """Common code which compute flops for binary operations."""
  out_shape = graph_util.tensor_shape_from_node_def_name(graph, node.name)
  out_shape.assert_is_fully_defined()
  return ops.OpStats("flops", out_shape.num_elements() * ops_per_element) 

# 需要导入模块: from tensorflow.python.framework import ops [as 别名]
# 或者: from tensorflow.python.framework.ops import OpStats [as 别名]
def _reduction_op_flops(graph, node, reduce_flops=1, finalize_flops=0):
  """Common code which compute flops for reduction operations."""
  in_shape = graph_util.tensor_shape_from_node_def_name(graph, node.input[0])
  in_shape.assert_is_fully_defined()
  out_shape = graph_util.tensor_shape_from_node_def_name(graph, node.name)
  out_shape.assert_is_fully_defined()
  num_flops = (in_shape.num_elements() * reduce_flops
               + out_shape.num_elements() * (finalize_flops - reduce_flops))
  return ops.OpStats("flops", num_flops) 

# 需要导入模块: from tensorflow.python.framework import ops [as 别名]
# 或者: from tensorflow.python.framework.ops import OpStats [as 别名]
def _conv_2d_backprop_input_flops(graph, node):
  """Compute flops for Conv2DBackpropInput operation."""
  # Formula:
  #  batch_size * image_x_dim * image_y_dim * kernel_x_dim * kernel_y_dim
  #  * input_depth * output_depth * 2 / (image_x_stride * image_x_stride)
  #
  # Where:
  # image_x_dim, image_y_dim and input_depth --- size of input to source (no
  #   backprop) convolution, in other words they are sizes of backprop output.
  # output_depth --- number of filters in the original convolution, thus
  #   depth of backprop input.
  # kernel_x_dim and kernel_y_dim --- sizes of filter in spatial dimension
  # image_x_stride and image_x_stride --- strides of the convolution
  #
  _verify_conv_data_format(node)
  # out_shape = [batch_size, image_y_dim, image_x_dim, input_depth]
  out_shape = graph_util.tensor_shape_from_node_def_name(graph, node.name)
  out_shape.assert_is_fully_defined()
  # kernel_shape = [kernel_y_dim, kernel_x_dim, input_depth, output_depth]
  kernel_shape = graph_util.tensor_shape_from_node_def_name(graph,
                                                            node.input[1])
  kernel_shape.assert_is_fully_defined()
  # strides
  strides_shape = list(node.attr["strides"].list.i)
  strides_product = strides_shape[1] * strides_shape[2]
  return ops.OpStats("flops",
                     (2 * out_shape.num_elements()
                      * kernel_shape.num_elements()
                      / (out_shape[-1].value * strides_product)))