basic knowledge

Represents computations as graphs.
Executes graphs in the context of Sessions.
Represents data as tensors.
Maintains state with Variables.
Uses feeds and fetches to get data into and out of arbitrary operations.



In [1]:

    
import tensorflow as tf
import numpy as np
import pandas as pd
import matplotlib
%matplotlib inline

build graph

So we should build the graph first To build the graph start with ops that do not need any input and then pass their output to other ops that do computation



In [15]:

    
# the ops add the default graph by default
matrix1 = tf.constant([[2.,3.]])

matrix2 = tf.constant([[2.,3.],[2.,3.]])

product = tf.matmul(matrix1,matrix2)

launching graph

To launch a graph ,create a session object.Without arguments session launch the default graph.



In [16]:

    
sess = tf.Session()

result = sess.run(product)

print(result)

sess.close()









    



[[ 10.  15.]]

Interactive usage

use it so that can avoids having to keep a variable holding the session
use eval() for tensor and use run() for ops



In [19]:

    
sess = tf.InteractiveSession()

x = tf.Variable([1.,2.])

a = tf.constant([3.,4.])

# must initilize the variable
x.initializer.run()

# sub a from x
sub = tf.sub(a,x)

sub.eval()

sess.close()









    



Exception AssertionError: AssertionError("Nesting violated for default stack of <type 'weakref'> objects",) in <bound method InteractiveSession.__del__ of <tensorflow.python.client.session.InteractiveSession object at 0x7fb3def36dd0>> ignored

Variables

Variables maintain state across executions of the graph.



In [31]:

    
# reset the graph
tf.reset_default_graph()
# counter
state = tf.Variable(0,name="counter")

# constant
one = tf.constant(1)

# new state
new_state = tf.add(state,one)

# assign new_state to state
# note that assign return the ref of state
update = tf.assign(state,new_state)

# init ops
init = tf.initialize_all_variables()

with tf.Session() as sess:
#   initialize variable
    sess.run(init)
#   print the initial state  
    
    print(sess.run(state))

#     update three times
    for _ in range(3):
        print(sess.run(update))

you can also get multiple tensors just by pass in the array of tensors

feeds

feeds temporarily replaces the output of the op with the tensor value.The common case is that load the data to train



In [38]:

    
tf.reset_default_graph()
input1 = tf.placeholder(tf.float32,shape=[None,2])

input2 = tf.placeholder(tf.float32,shape=[2,1])

output = tf.matmul(input1,input2)

with tf.Session() as sess:
    res = sess.run([output],feed_dict={input1:[[4.,5.]],input2:[[3.],[5.]]})
    print(res)









    



[array([[ 37.]], dtype=float32)]

Others

do some experiments



In [5]:

    
with tf.Session() as sess:
    logits = tf.Variable([[0.2,0.3,0.5],[0.3,0.3,0.4]],name='logits')
    labels = [0,1]
    init = tf.initialize_all_variables()
    sess.run(init)
    
    correct_number = tf.nn.in_top_k(logits,labels,1)
    true = tf.reduce_sum(tf.cast(correct_number,tf.int32))
    cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits,labels,
                                                   name='xentropy')
#     print(true)
    print(sess.run(true))
    print(sess.run(cross_entropy))









    



0
[ 1.23983097  1.13306868]