tensorflow learning

Part 1: Prepare data

This is common for all deep learning work. And it depends on the data. In python, many people like to use pickle to store data.

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("MNIST_data/",one_hot = True)

In general, data include inputs and labels.

Part 2: build model

At most time, we could create a python file named model.py to define our deep learning model.

In general, the model include the input layers, hidden layers and loss function.

We use placeholder to set the input layers in tensorflow. (Take mnist as example)

inputs = tf.placeholder(dtype='float32',shape=[None,784],name='inputs')
labels = tf.placeholder(dtype='float32',shape=[None,10],name='labels')

And we define every layers by the code below:

in_size = 784
hid_size = 10
with tf.name_scope('layer1'):
    Weights= tf.Variable(tf.random_normal([in_size,hid_size]),name='Weights')
    biases = tf.Variable(tf.random_normal(1,hid_size),name = 'biases')
    l1 = tf.add(tf.matmul(inputs,Weights),biases)

define loss:

loss = tf.loss.mean_square(l1,labels)

define summary:

tf.summary.scalar(loss,name='loss')

Part 3: train and save

define the training size

batch_size = 100
all_size = 50000

define the Session

with tf.Session(config=tf.ConfigProto(log_device_placement=True)) as sess:

initialize the parameters

tf.global_variables_initializer().run()
merged = tf.summary.merge_all()
train_writer = tf.summary.FileWriter('logs/',sess.graph)
for epoch in range(1000):
    begin = epoch*batch_size%(all_size-batch_size)
    end = begin + batch_size
    batch = inputs[begin:end]
    print(batch.shape)
    batch_labels = labels[epoch*batch_size:(epoch+1) * batch_size]

input the training data

    train_summary,loss_value=sess.run([train_step,loss],feed_dict={image_inputs:batch,points:batch_labels})
    if (epoch%10==0):
        train_writer.add_summary(train_summary,epoch)
        print(loss_value)
train_writer.close()