Tensorflow概念

AI将会变得更加民主化, AI程序开发迟早会变为程序员的必备技能,还犹豫什么,赶紧来学习. Google的Tensorflow无疑是目前最有前景的框架, 那么Tensorflow到底好不好学呢?我们拭目以待. 本篇介绍Tensorflow的基本概念.

1. 基本元素

1.1 constant

const的原型是tf.constant(value, dtype=None, shape=None, name='Const', verify_shape=False),可以是常数,向量,矩阵等.例子如下:

import tensorflow as tf


def const_literal():
    a = tf.constant(2, name='a')
    b = tf.constant(3, name='b')
    x = tf.add(a, b, name='add')

    with tf.Session() as sess:
        writer = tf.summary.FileWriter('./graphs', sess.graph)
        print(sess.run(x))
    writer.close()


def const_tensor():
    a = tf.constant([2, 2], name='a')
    b = tf.constant([[0, 1], [2, 3]], name='b')
    x = tf.add(a, b, name='add')
    y = tf.multiply(a, b, name='mul')  # element wise multiply

    with tf.Session() as sess:
        x, y = sess.run([x, y])
        print('x:')
        print(x)
        print('y:')
        print(y)


def const_zeros():
    """tf.zeros and tf.ones has same API"""
    a = tf.zeros([2, 3], tf.int32)
    b = tf.zeros_like(a, tf.float32)
    with tf.Session() as sess:
        print(sess.run(a))
        print(sess.run(b))


def const_fill(val):
    """fill the tensor with a value"""
    a = tf.fill([2, 3], val)
    with tf.Session() as sess:
        print(sess.run(a))


def const_linear(start, stop, num):
    """linear space numbers in [start, stop], only float32, float64 permited"""
    a = tf.linspace(start, stop, num)
    b = tf.range(start, stop, 1.0)
    with tf.Session() as sess:
        print(sess.run(a))
        print(sess.run(b))


def const_random():
    """
    tf.random_normal(shape, mean=0.0, stddev=1.0, dtype=tf.float32, seed=None, name=None)
    tf.truncated_normal(shape, mean=0.0, stddev=1.0, dtype=tf.float32, seed=None, name=None)
    tf.random_uniform(shape, minval=0, maxval=None, dtype=tf.float32, seed=None, name=None)
    tf.random_shuffle(value, seed=None, name=None)
    tf.random_crop(value, size, seed=None, name=None)
    tf.multinomial(logits, num_samples, seed=None, name=None)
    tf.random_gamma(shape, alpha, beta=None, dtype=tf.float32, seed=None, name=None)
    """
    pass


def const_graph():
    """don't need to use my_const, it has already been in the compute graph"""
    my_const = tf.constant([1.0, 2.0], name='my_const')
    with tf.Session() as sess:
        print(sess.graph.as_graph_def())


if __name__ == '__main__':
    # const_linear(0.0, 99.0, 99)
    const_graph()

Tensorboard

Tensorboard查看Graph.

Variable

constant是一个operation,子graph构建的时候定义,Variable是一个类,代表变量.constant在图的的定义里边,Variable可以在参数服务器.

变量在使用前要进行显示的初始化,否则报未初始化的错.

可以使用eval()进行求知,只有operation和tensor有eval()函数,Tensor.eval()相当于get_default_session().run(t).

每一个Variable都有一个initializer,只有Variable被初始化了或者赋值成功了,才可以eval()

import tensorflow as tf


def test_eval():
    W = tf.constant(10)
    with tf.Session():
        print(W.eval())         # 10


def test_eval_Variable():
    W = tf.Variable(10)
    with tf.Session() as sess:
        print(sess.run(W.initializer))  # None <--- 1.
        print(W.eval())                 # 10


def test_eval_Variable_all():
    W = tf.Variable(10)
    with tf.Session():
        print(W.initializer.eval())  # error: object has no attribute 'eval'
        print(W.eval())

def initialize_properly():
    W = tf.Variable(10)
    with tf.Session() as sess:
        #This way
        tf.global_variables_initializer().run()

        print(W.eval()) 
        print(sess.run(W))


def run_multiple_times():
    W = tf.Variable(10)
    a_times_two = W.assign(2 * W)
    with tf.Session():
        tf.global_variables_initializer().run()
        print(W.eval())         # 10
        print(a_times_two.eval())  # 20
        print(a_times_two.eval())  # 40


if __name__ == '__main__':
    test_eval()
    test_eval_Variable()
    test_eval_Variable_all()

Placeholders

placeholder和Variable在普通的编程意义上差不多,不过在tensorflow里边,placeholder用来表示输入输出的数据,相当于C/C++的io, Variable代表在学习中可以更新,迭代,存储的参数,更接近于普通意义上的变量. 具体来说有一下不同:

• Variable需要用Tensor初始化; placeholder不用,也不能初始化
• Variable的数据可以在训练中更新
• Variable可以共享,并且可以是nontrainble
• Variable学习好的参数可以保存在磁盘中
• Variable创建的时候有三个op自动创建: variable op, initializer op, ops for the initial value
• Variable是一个class, placeholder是一个function
• 在分布式环境下,Variable在参数服务器里边,并且在不同的worker里边共享
• Variable使用前要初始化,在使用的过程中shape是固定的, placeholder在使用的时候要feed数据.

Session

import tensorflow as tf

x = tf.Variable(3, name='x')
y = tf.Variable(4, name='y')

f = x*x*y + y + 2

with tf.Session() as sess:
    x.initializer.run()
    y.initializer.run()
#     result = f.eval()
#     result = sess.run(f)
    result = tf.get_default_session().run(f)

    
tf.reset_default_graph()
print(result)
result = None

• 这段代码里边用了三种方法求值
• InteractiveSession自动创建一个Session并且是default Session,不需要with block

Graph操作

x1 = tf.Variable(1)
x1.graph is tf.get_default_graph() # True

graph = tf.Graph()
with graph.as_default():
    x2 = tf.Variable(2)
    
x2.graph is graph # True

x2.graph is tf.get_default_graph() # False

• 任何一个创建的node都会自动放到default graph里边
• 当然也可以给node指定graph,尤其是程序中有多个graph的情形.

Node的生命周期

w = tf.constant(3)
x = w + 2
y = x + 5
z = x * 3

with tf.Session() as sess:
    print(y.eval())
    print(z.eval())

• 在这个代码中求y的时候需要先求x和w,求z的时候也是x和w,但是第二次不能复用第一次的结果
• 在graph的run函数调用后,除了variable之外的数据都会被丢弃,variable的声明周期始于初始化,止于session关闭.
• 为了让求值更搞笑,需要在一个run中计算y和z:

with  tf.Session() as sesss:
    y_val, z_val = sess.run([y, z])
    print(y_val)
    print(z_val)