tensorflow基本操作（1）

import tensorflow as tf
import numpy as np

点乘，支持broadcasting

• 乘号* 和 multiply等价
• mul已经废弃不用了
• matmul 是矩阵相乘

broadcasting参见：

http://docs.scipy.org/doc/numpy/user/basics.broadcasting.html
原则：
When operating on two arrays, NumPy compares their shapes element-wise. It starts with the trailing dimensions, and works its way forward. Two dimensions are compatible when

• they are equal, or
• one of them is 1

a = tf.placeholder("float")
b = tf.placeholder("float")
y = a * b
y2 = tf.multiply(a,b)
sess = tf.Session()
r, r2 = sess.run([y,y2], feed_dict={a:[3,1], b:[[4],[1]]})
print r
print r2
sess.close()

切片 slice

• begin 起点的坐标
• size 切片的大小
• begin[i] + size[i] <=input.shape[i]

分割 split, 沿着某一维度将tensor分割为n个tensor list

• 分割后list中的每个tensor的维度不降

input = tf.placeholder('float')
begin = tf.placeholder('int32')
size = tf.placeholder('int32')
out = tf.slice(input, begin, size)
s = tf.split(input, num_or_size_splits=3, axis=0)
sess= tf.Session()
input_ = [[[1, 1, 1], [2, 2, 2]],[[3, 3, 3], [4, 4, 4]],[[5, 5, 5], [6, 6, 6]]]
begin_ = [1, 0, 0]
size_ = [1,1,3]
o = sess.run(out, feed_dict={input:input_, begin:begin_, size:size_})
print o
s_ = sess.run(s, feed_dict={input:input_})
print(s_)
print(type(s_))
print(s_[0])
print(type(s_[0]))
print(type(s))
print(type(out))
sess.close()

[[[ 3. 3. 3.]]]
[array([[[ 1., 1., 1.],
[ 2., 2., 2.]]], dtype=float32), array([[[ 3., 3., 3.],
[ 4., 4., 4.]]], dtype=float32), array([[[ 5., 5., 5.],
[ 6., 6., 6.]]], dtype=float32)]
<type 'list'>
[[[ 1. 1. 1.]
[ 2. 2. 2.]]]
<type 'numpy.ndarray'>
<type 'list'>
<class 'tensorflow.python.framework.ops.Tensor'>

tensor concat 沿着某一维度连接tensor

tensor stack 将一系列rank=R的tensor 打包为rank=R+1的tensor

tensor pack 已被废弃，用stack代替

t1 = tf.constant([[1,2,3],[4,5,6]])
t2 = tf.constant([[7,8,9],[10,11,12]])
c1= tf.concat([t1,t2], 0)
c2= tf.concat([t1,t2], 1)
r1 = tf.reshape(c1, [-1])
r2 = tf.reshape(c2, [-1])

p1 = tf.stack([t1,t2],0)
p2 = tf.stack([t1,t2],1)

p3 = [t1, t2]

with tf.Session() as sess:
    print(sess.run([c1, c2]))
    print(sess.run([tf.rank(t1),tf.rank(c1)]))
    print("=======")
    print(sess.run([r1, r2]))
    print("=======")
    print(sess.run([p1, p2]))
    print(sess.run([tf.rank(t1),tf.rank(p1)]))
    print(sess.run(tf.shape([p1, p2])))
    print("=======")
    print(sess.run(p3))

[array([[ 1, 2, 3],
[ 4, 5, 6],
[ 7, 8, 9],
[10, 11, 12]], dtype=int32), array([[ 1, 2, 3, 7, 8, 9],
[ 4, 5, 6, 10, 11, 12]], dtype=int32)]
[2, 2]
=======
[array([ 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12], dtype=int32), array([ 1, 2, 3, 7, 8, 9, 4, 5, 6, 10, 11, 12], dtype=int32)]
=======
[array([[[ 1, 2, 3],
[ 4, 5, 6]],

[[ 7, 8, 9],
[10, 11, 12]]], dtype=int32), array([[[ 1, 2, 3],
[ 7, 8, 9]],

[[ 4, 5, 6],
[10, 11, 12]]], dtype=int32)]
[2, 3]
=======
[array([[1, 2, 3],
[4, 5, 6]], dtype=int32), array([[ 7, 8, 9],
[10, 11, 12]], dtype=int32)]
=======
[2 2 2 3]

c1 = tf.constant([1])
c2 = tf.constant([2])
c3 = tf.constant([3])
con = tf.concat([c1,c2,c3], 0)
with tf.Session() as sess:
    print(sess.run(tf.shape(c1)[0]))
    print(sess.run(con))

1
[1 2 3]

可以认为 tensor 是一个n维的数组（array）或列表（list）, tensor具有静态的type 和动态的维度
一个scalar 就是0维的数组，rank = 0，shape是[]
下面的
c1 rank=0, shap=[]
c2 rank=1, shap=[1]
c3 rank=1, shap=[2]

c1 = tf.constant(1)
c2 = tf.constant([1])
c3 = tf.constant([1,2])
s1 = tf.shape(c1)
s2 = tf.shape(c2)
s3 = tf.shape(c3)
with tf.Session() as sess:
    c1_, c2_, c3_ = sess.run([c1,c2,c3])
    print(c1_, c2_, c3_)
    print(type(c1_), type(c2_), type(c3_))
    print(c1_.shape)
    print("================")
    o1, o2,o3 = sess.run([s1,s2,s3])
    print(o1, o2, o3)
    print(type(o1), type(o2), type(o3))
    print(o1.shape)
    

(1, array([1], dtype=int32), array([1, 2], dtype=int32))
(<type 'numpy.int32'>, <type 'numpy.ndarray'>, <type 'numpy.ndarray'>)
()
================
(array([], dtype=int32), array([1], dtype=int32), array([2], dtype=int32))
(<type 'numpy.ndarray'>, <type 'numpy.ndarray'>, <type 'numpy.ndarray'>)
(0,)

li = np.zeros(5)
li[0] = tf.constant(1) # tensor是个数组，哪怕是纯量

---------------------------------------------------------------------------

ValueError Traceback (most recent call last)

in ()
1 li = np.zeros(5)
----> 2 li[0] = tf.constant(1)

ValueError: setting an array element with a sequence.

placeholder, 不指定shape时候，可以feed任何shape的数据；指定的话，必须按照指定的shape feed 数据

x = tf.placeholder(shape=(2,1), dtype=tf.int32)
m = tf.placeholder(dtype=tf.int32)
y = x*2
n = m*2
with tf.Session() as sess:
    m_ = [[2.0],[2.0]]
    m_o = sess.run(m, feed_dict={m:m_})
    print(m_)
    print(m_o)
    print(m_ - m_o)
    print(type(m_), type(m_o))
    print(sess.run(n, feed_dict={m:[[2,2],[1,1]]}))

[[2.0], [2.0]]
[[2]
[2]]
[[ 0.]
[ 0.]]
(<type 'list'>, <type 'numpy.ndarray'>)
[[4 4]
[2 2]]

li = []
li.append([tf.constant(1)])
li.append([tf.constant(2)])
li2 = tf.concat(li, axis=0)
with tf.Session() as sess:
    print(sess.run(li))
    print(sess.run(li2))

[[1], [2]]
[1 2]

li = []
li.append(tf.constant(1))
li.append(tf.constant(2))
li2 = tf.stack(li, axis=0)
#li2 = tf.concat(li, axis=0) # 错误, 纯量不能使用cancat
with tf.Session() as sess:
    print(sess.run(li))
    print(sess.run(li2))

[1, 2]
[1 2]