Build the simplest CNN framework based on the TensorFlow framework

Time:2022-11-24

Project Description

This article will use python and use the TensorFlow framework to build the simplest CNN framework to recognize handwritten digits.

The CNN framework structure built in this paper

[1] Input layer (the input in this article is a 28*28 single-channel image, so the input layer has 784 nodes)
[2] The first convolutional layer (the convolutional layer contains 32 different 5The convolution kernel of 5, that is, the convolution layer extracts 32 different graphic features, [5,5,1,32] means that the convolution kernel size is 55,1 color channel, 32 different convolution kernels)
[3] The maximum pooling layer after the first convolutional layer
[4] The second convolutional layer (the convolutional layer contains 64 different 55 convolution kernel, that is, the convolution layer extracts 32 different graphic features, [5,5,32,64] means that the convolution kernel size is 55, 64 different convolution kernels)
[5] The maximum pooling layer after the second convolutional layer
[6] Fully connected layer
[7] A Dropout layer (in order to reduce over-fitting, during training, we randomly discard some node data to reduce over-fitting, and when predicting, keep all the data to pursue the best prediction performance)
[8] Softmax layer to get the final probability output.
[9] Define the loss function as cross entropy (cross entropy), and the optimizer uses Adam
[10] Obtain the prediction accuracy of the model

project code

import the corresponding library

from tensorflow.examples.tutorials.mnist import input_data
import tensorflow as tf

Import handwritten digits dataset

mnist = input_data.read_data_sets(“MNIST_data/”,one_hot = True)
sess = tf.InteractiveSession()

Define the function that generates the weights

def weight_variabel(shape):

initial = tf.truncated_normal(shape,stddev = 0.1)
return tf.Variable(initial)

Define the function that generates the bias

def bias_variable(shape):

initial = tf.constant(0.1,shape = shape)
return tf.Variable(initial)

Define the function that generates the convolutional layer

convolutional layer

def conv2d(x,W):

return tf.nn.conv2d(x,W,strides=[1,1,1,1],padding='SAME')

pooling layer

Define the function that generates the max pooling layer

def max_pool_2x2(x):

return tf.nn.max_pool(x,ksize=[1,2,2,1],strides=[1,2,2,1],
                     padding='SAME')

The variable passed into the input

x = tf.placeholder(tf.float32,[None,784])

Variables passed into the label

y_ = tf.placeholder(tf.float32,[None,10])

Convert 1D pictures to 28*28 2D pictures

x_image = tf.reshape(x,[-1,28,28,1])

We define the first convolutional layer

Weights

W_conv1 = weight_variabel([5,5,1,32])

bias

b_conv1 = bias_variable([32])

convolution kernel

h_conv1 = tf.nn.relu(conv2d(x_image,W_conv1)+b_conv1)

max pooling layer

h_pool1 = max_pool_2x2(h_conv1)

Define the second convolutional layer

Weights

W_conv2 = weight_variabel([5,5,32,64])

bias

b_conv2 = bias_variable([64])

convolution kernel

h_conv2 = tf.nn.relu(conv2d(h_pool1,W_conv2)+b_conv2)

max pooling layer

h_pool2 = max_pool_2x2(h_conv2)

Define a fully connected layer, the number of hidden nodes is 1024, and use the ReLU activation function

W_fc1 = weight_variabel([7764,1024])
b_fc1 = bias_variable([1024])
h_pool2_flat = tf.reshape(h_pool2,[-1,7764])
h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat,W_fc1)+b_fc1)

In order to alleviate overfitting, a Dropout layer is used below, which is controlled by the transmission keepr_prob ratio of a placeholder. During training,

We randomly discard a part of the node data to reduce overfitting, and we keep all the data to pursue the best prediction performance when predicting

keep_prob = tf.placeholder(tf.float32)
h_fc1_drop = tf.nn.dropout(h_fc1,keep_prob)

Finally, we connect the output of the Dropout layer to a Softmax layer to get the final probability output

W_fc2 = weight_variabel([1024,10])
b_fc2 = bias_variable([10])
y_conv = tf.nn.softmax(tf.matmul(h_fc1_drop,W_fc2)+b_fc2)

We define the loss function cross entropy as before, but the optimizer uses Adam and gives a small learning rate of 1e-4

cross_entropy = tf.reduce_mean(-tf.reduce_sum(y_*tf.log(y_conv),

                                         reduction_indices = [1]))

train_step = tf.train.AdamOptimizer(1e-4).minimize(cross_entropy)

correct_prediction = tf.equal(tf.argmax(y_conv,1),tf.argmax(y_,1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction,tf.float32))

Let’s start the training process, first of all, initialize all parameters

tf.global_variables_initializer().run()
for i in range(20000):

batch = mnist.train.next_batch(50)
if i % 100 == 0:
    train_accuracy = accuracy.eval(feed_dict= {x:batch[0],y_:batch[1],
                                              keep_prob:1.0})
    print("step %d,trainning accuracy %g"%(i,train_accuracy))
train_step.run(feed_dict={x:batch[0],y_:batch[1],keep_prob:0.5})

print(“test accuracy %g”%accuracy.eval(feed_dict={x:mnist.test.images,y_:mnist.test.labels,keep_prob:1.0}))