笔记 - tensorflow：sess.run机制

实验结果

import tensorflow as tf

def read_data():
    print("read data ...")
    return tf.constant(value=[1.0, 2.0, 3.0], dtype=tf.float32)

X = read_data()
X_train = tf.placeholder(dtype=tf.float32)
with tf.Session() as sess:
    for epoch in range(3):
        for batch in range(3):
            x = sess.run(X)
            print(sess.run(X_train, feed_dict={X_train: x}))

• 运行结果

read data ...
[1. 2. 3.]
[1. 2. 3.]
[1. 2. 3.]
[1. 2. 3.]
[1. 2. 3.]
[1. 2. 3.]
[1. 2. 3.]
[1. 2. 3.]
[1. 2. 3.]

虽然循环体内 sess.run(X) 了9次，但 read_data 实际却只调用了一次
真的好神奇！！！！

---

换成random，效果更明显

def read_data():
    print("read data ...")
    return tf.random_uniform(shape=(3,), maxval=1.0)
    
read data ...
[0.6201165  0.70080805 0.04186273]
[0.04440641 0.27251375 0.35242593]
[0.58647656 0.6420467  0.47325552]
[0.4188739 0.9272245 0.3594923]
[0.49990058 0.930122   0.30738378]
[0.6659864  0.48257875 0.2847129 ]
[0.64794695 0.0111196  0.00765169]
[0.9654126  0.05055571 0.20485735]
[0.7216649  0.483734   0.49638057]

只调用一次函数，但随机数竟然每次run都不一样...

继续实验

import tensorflow as tf

# def read_data():
#     print("read data ...")
#     return tf.constant(value=[1.0, 2.0, 3.0], dtype=tf.float32)

def hello():
    print("hello")

def read_data():
    print("read data ...")
    print("加载10个G数据")
    print("[1,2,3,4,...,99999]")
    hello()
    return tf.random_uniform(shape=(3,), maxval=1.0)

X = read_data()
X_train = tf.placeholder(dtype=tf.float32)
# with tf.Session() as sess:
#     for epoch in range(3):
#         for batch in range(3):
#             x = sess.run(X)
#             print(sess.run(X_train, feed_dict={X_train: x}))

read data ...
加载10个G数据
[1,2,3,4,...,99999]
hello

除了tensorflow本身的操作在sess.run才执行，其他操作都会提前执行！！！