tensorflow作為一個基於圖結構的深度學習框架，內部通過session實現圖和計算核心的互動，那麼這個圖是什麼樣的結構，session的工作原理又是什麼樣的呢？我們通過幾段程式碼來深入理解一下

tensorflow中的基本數學運算用法

import tensorflow as tf

sess = tf.Session()

a = tf.placeholder("float")
b = tf.placeholder("float")
c = tf.constant(6.0)
d = tf.mul(a, b)
y = tf.mul(d, c)
print sess.run(y, feed_dict={a: 3, b: 3})

A
 
 = [[1.1,2.3],[3.4,4.1]]
Y = tf.matrix_inverse(A)
print sess.run(Y)
sess.close()

主要數字運算還包括：

tf.add
tf.sub
tf.mul
tf.div
tf.mod
tf.abs
tf.neg
tf.sign
tf.inv
tf.square
tf.round
tf.sqrt
tf.pow
tf.exp
tf.log
tf.maximum
tf.minimum
tf.cos
tf.sin

主要矩陣運算還包括：

tf.diag生成對角陣
tf.transpose
tf.matmul
tf.matrix_determinant計算行列式的值
tf.matrix_inverse計算矩陣的逆
 

插播小甜點：tensorboard使用

tensorflow因為程式碼執行過程是先構建圖，然後在執行，所以對中間過程的除錯不太方便，所以提供了一個tensorboard工具來便於除錯，用法如下：

在訓練時會提示寫入事件檔案到哪個目錄(比如：/tmp/tflearn_logs/11U8M4/)

執行如下命令並開啟http://192.168.1.101:6006就能看到tensorboard的介面

tensorboard --logdir=/tmp/tflearn_logs/11U8M4/

什麼是Graph和Session

為了步入正題，我們通過一段程式碼來展示Graph和Session的使用

import tensorflow as
 
 tf

with tf.Graph().as_default() as g:
    with g.name_scope("myscope") as scope: # 有了這個scope，下面的op的name都是類似myscope/Placeholder這樣的字首
        sess = tf.Session(target='', graph = g, config=None) # target表示要連線的tf執行引擎
        print "graph version:", g.version # 0
        a = tf.placeholder("float")
        print a.op # 輸出整個operation資訊，跟下面g.get_operations返回結果一樣
        print "graph version:", g.version # 1
        b = tf.placeholder("float")
        print "graph version:", g.version # 2
        c = tf.placeholder("float")
        print "graph version:", g.version # 3
        y1 = tf.mul(a, b) # 也可以寫成a * b
        print "graph version:", g.version # 4
        y2 = tf.mul(y1, c) # 也可以寫成y1 * c
        print "graph version:", g.version # 5
        operations = g.get_operations()
        for (i, op) in enumerate(operations):
            print "============ operation", i+1, "==========="
            print op # 一個結構，包括：name、op、attr、input等,不同op不一樣
        assert y1.graph is g
        assert sess.graph is g
        print "================ graph object address ================"
        print sess.graph
        print "================ graph define ================"
        print sess.graph_def
        print "================ sess str ================"
        print sess.sess_str
        print sess.run(y1, feed_dict={a: 3, b: 3}) # 9.0 feed_dictgraph中的元素和值的對映
        print sess.run(fetches=[b,y1], feed_dict={a: 3, b: 3}, options=None, run_metadata=None) # 傳入的feches和返回值的shape相同
        print sess.run({'ret_name':y1}, feed_dict={a: 3, b: 3}) # {'ret_name': 9.0} 傳入的feches和返回值的shape相同

        assert tf.get_default_session() is not sess
        with sess.as_default(): # 把sess作為預設的session，那麼tf.get_default_session就是sess, 否則不是
            assert tf.get_default_session() is sess

        h = sess.partial_run_setup([y1, y2], [a, b, c]) # 分階段執行，引數指明瞭feches和feed_dict列表
        res = sess.partial_run(h, y1, feed_dict={a: 3, b: 4}) # 12 執行第一階段
        res = sess.partial_run(h, y2, feed_dict={c: res}) # 144.0 執行第二階段，其中使用了第一階段的執行結果
        print "partial_run res:", res
        sess.close()

輸出如下：

graph version: 0
name: "myscope/Placeholder"
op: "Placeholder"
attr {
  key: "dtype"
  value {
    type: DT_FLOAT
  }
}
attr {
  key: "shape"
  value {
    shape {
    }
  }
}

graph version: 1
graph version: 2
graph version: 3
graph version: 4
graph version: 5
============ operation 1 ===========
name: "myscope/Placeholder"
op: "Placeholder"
attr {
  key: "dtype"
  value {
    type: DT_FLOAT
  }
}
attr {
  key: "shape"
  value {
    shape {
    }
  }
}

============ operation 2 ===========
name: "myscope/Placeholder_1"
op: "Placeholder"
attr {
  key: "dtype"
  value {
    type: DT_FLOAT
  }
}
attr {
  key: "shape"
  value {
    shape {
    }
  }
}

============ operation 3 ===========
name: "myscope/Placeholder_2"
op: "Placeholder"
attr {
  key: "dtype"
  value {
    type: DT_FLOAT
  }
}
attr {
  key: "shape"
  value {
    shape {
    }
  }
}

============ operation 4 ===========
name: "myscope/Mul"
op: "Mul"
input: "myscope/Placeholder"
input: "myscope/Placeholder_1"
attr {
  key: "T"
  value {
    type: DT_FLOAT
  }
}

============ operation 5 ===========
name: "myscope/Mul_1"
op: "Mul"
input: "myscope/Mul"
input: "myscope/Placeholder_2"
attr {
  key: "T"
  value {
    type: DT_FLOAT
  }
}

================ graph object address ================
<tensorflow.python.framework.ops.Graph object at 0x1138702d0>
================ graph define ================
node {
  name: "myscope/Placeholder"
  op: "Placeholder"
  attr {
    key: "dtype"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "shape"
    value {
      shape {
      }
    }
  }
}
node {
  name: "myscope/Placeholder_1"
  op: "Placeholder"
  attr {
    key: "dtype"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "shape"
    value {
      shape {
      }
    }
  }
}
node {
  name: "myscope/Placeholder_2"
  op: "Placeholder"
  attr {
    key: "dtype"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "shape"
    value {
      shape {
      }
    }
  }
}
node {
  name: "myscope/Mul"
  op: "Mul"
  input: "myscope/Placeholder"
  input: "myscope/Placeholder_1"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
}
node {
  name: "myscope/Mul_1"
  op: "Mul"
  input: "myscope/Mul"
  input: "myscope/Placeholder_2"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
}
versions {
  producer: 15
}

================ sess str ================

9.0
[array(3.0, dtype=float32), 9.0]
{'ret_name': 9.0}
partial_run res: 144.0

tensorflow的Session是如何工作的

Session是Graph和執行者之間的媒介，Session.run()實際上將graph、fetches、feed_dict序列化到位元組陣列中，並呼叫tf_session.TF_Run（參見/usr/local/lib/python2.7/site-packages/tensorflow/python/client/session.py）

而這裡的tf_session.TF_Run實際上呼叫了動態連結庫_pywrap_tensorflow.so中實現的_pywrap_tensorflow.TF_Run介面(參見/usr/local/lib/python2.7/site-packages/tensorflow/python/pywrap_tensorflow.py)，這個動態連結庫是tensorflow提供的諸多語言介面中python語言的介面

事實上這裡的_pywrap_tensorflow.so和pywrap_tensorflow.py是通過SWIG工具自動生成，大家都知道tensorflow核心語言是c語言，這裡是通過SWIG生成了各種指令碼語言的介面