Tensorflow memo [updated from time to time]

Tensorflow memo

It is a memo that I have left behind

Structure of Tensorflow

Concept of graphs, arithmetic nodes, and tensors

What is a graph?

A figure representing one layer consisting of multiple neurons φ(X * W + b)

Consists of nodes and edges connecting the nodes There are arithmetic nodes, variable nodes, press holder nodes, etc.

What is a tensor?

Amount flowing through the graph

A tensor is an n-dimensional array or list

Main notation

Variable tf.Variable Matrix multiplication tf.matmul Application of φ tf.nn.relu

Graph definition

graph.py

#Definition of variable to put W
weights = tf.Variable()
#Definition of variable to put b
bias = tf.Variable()
#The function of the layer φ(X*W+b)Defined in
#Here φ uses relu
#images is the input that this layer receives
#hidden1 is the output of this layer

#1st layer
hidden1 = tf.nn.relu(tf.matmul(images, weights) + bias)

#2nd layer
hidden2 = tf.nn.relu(tf.matmul(hidden1,weights) + bias)
#images, weights, bias,All hidden1 are tensors

Variable definition

Variable initialization

init.py

w = tf.Variable(tf.random_norml([784, 200], stddev = 0.35), name = "weights")
b  =tf.Variable(tf.zeros([200], name = "biases")

#Operation to initialize this variable
#Caution! It hasn't been executed yet, just a node has been added.
init_op = tf.initialize_all_variables()


#Call this initialization after launching the model
#The defined model works for the first time in Session.
#Use run to call
with tf.Session() as sess:
    # Run the init operation.
    sess.run(init_op)

** Saving and restoring variables

save.py

#Create a variable
v1 = tf.variable(..., name = "v1")
v2 = tf.variable(..., name = "v2")

#Initialize variables init_op node
init_op = tf.initalize_all_variables()
#Save all variables,Add save node to restore
saver = tf.train.Saver()

#After launching the model, initializing variables, doing some work
#Save variables to disk
with tf.Session() as sess:
  sess.run(init_op)
  #Do something with the model
  ##########
  
  #Save variables to disk
  save_path = saver.save(sess, "/tmp/model.ckpt")
  print("Modef saved in file: %s" % save_path)

  #Variable restore
  saver.retore(sess, "/tmp/model.ckpt")
  print ("Modell restored")

Parameter optimization (training)

Parameter (weight and bias) optimization

1. Build a graph and calculate the output with the input data in the graph
2. Compare the output with the correct answer. Use loss function for comparison
3. Use Gradient Descent to modify the value of the loss function to a smaller value
4. Calculate the output with the new parameters. Repeat until the loss function is small

Main functions

GradientDescentOptimizer() Optimization operation for parameter optimization. Optimize the loss function with this Optimizer

opt.py

###For numerical prediction###
 loss = tf.reduce_mean(tf.square(y - y_data))
#Learning rate 0.Gradient descent at 5
optimizer = tf.train.GradientDescentOptimizer(0.5)
train = optimizer.minimize(loss)

###In case of classification###
y_ = tf.placeholder("float", [None , 10])
cross_enttopy = -tf.reduce_sum(y_ * tf.log(y))
optimizer = tf.train.GradientDescentOptimizer(0.01)
train = optimizer.minimize(cross_entropy)