了解成本模型张量流

Question

tf.GraphOptions.build_cost_model选项定义为在生成成本模型之前将要运行的步骤数。为什么我们有此选项？换句话说，“步骤数”到底是什么意思？

我有这段代码可以看到矩阵乘法的成本估算：

import tensorflow as tf
import datetime
import numpy as np
A = np.random.rand(100, 10).astype('float32')
B = np.random.rand(10, 200).astype('float32')
with tf.device('/cpu:0'):
    a = tf.placeholder(tf.float32, [100, 10])
    b = tf.placeholder(tf.float32, [10, 200])
mul = tf.matmul(a,b)
metadata = tf.RunMetadata()
run_options = tf.RunOptions(trace_level=tf.RunOptions.FULL_TRACE)
ops = tf.GraphOptions(build_cost_model=2)
with tf.Session(config=tf.ConfigProto(log_device_placement=True, graph_options=ops)) as sess:
    # Run the op.
    for _ in range(2):
        sess.run([mul], feed_dict={a:A, b:B}, options=run_options,run_metadata=metadata)
        print ("Cost graph " + str(metadata.cost_graph)

然后，我将其作为输出：

Cost graph 
Cost graph node {
  name: "_SOURCE"
  compute_cost: 2
}
node {
  name: "_SINK"
  id: 1
  control_input: 0
  control_input: 0
  control_input: 5
}
node {
  name: "_arg_Placeholder_38_0_1"
  device: "/job:localhost/replica:0/task:0/device:CPU:0"
  id: 2
  output_info {
    alias_input_port: -1
    shape {
      dim {
        size: 10
      }
      dim {
        size: 200
      }
    }
    dtype: DT_FLOAT
  }
  control_input: 0
  compute_cost: 1
}
node {
  name: "_arg_Placeholder_37_0_0"
  device: "/job:localhost/replica:0/task:0/device:CPU:0"
  id: 3
  output_info {
    alias_input_port: -1
    shape {
      dim {
        size: 100
      }
      dim {
        size: 10
      }
    }
    dtype: DT_FLOAT
  }
  control_input: 0
  compute_cost: 1
}
node {
  name: "MatMul_11"
  device: "/job:localhost/replica:0/task:0/device:CPU:0"
  id: 4
  input_info {
    preceding_node: 3
  }
  input_info {
    preceding_node: 2
  }
  output_info {
    size: 80000
    alias_input_port: -1
    shape {
      dim {
        size: 100
      }
      dim {
        size: 200
      }
    }
    dtype: DT_FLOAT
  }
  compute_cost: 24
}
node {
  name: "_retval_MatMul_11_0_0"
  device: "/job:localhost/replica:0/task:0/device:CPU:0"
  id: 5
  input_info {
    preceding_node: 4
  }
  compute_cost: 2
}

这里的计算成本如何估算？为什么运行更多次会影响成本？