Tensorflow实战 LeNet-5神经网络进行手写体数字识别
完整代码：https://github.com/jiang4869/LeNet5-Tensorflow-MNIST
关于LeNet5网络的介绍，可以参考我的上一篇博客浅谈LeNet-5
本文基于Tensorflow对LeNet5进行复现并进行手写体数字识别。
运行环境：win10+tensorflow1.8.0+cuda9.0+cudnn7.0
一、项目结构
文件介绍
| 文件 | 功能 |
|---|---|
| lenet.py | 定义lenent模型 |
| layer_util.py | 封装一些常用函数 |
| train.py | 配置一些训练参数以及数据的读取 |
二、代码详解
layer_util.py
为了降低代码耦合度，所以我将卷积，池化，全连接过程，抽象封装成函数，这样便于定义网络的时候使用。现在我们在定义卷积层、池化层，全连接层的时候，只需要关注输出是多少就好，不需要在进行繁琐计算。
参数的获取
定义了获取变量和获取常量的函数
def get_variable(shape, stddev=0.1):
initial = tf.truncated_normal(shape, stddev=stddev)
return tf.Variable(initial)
def get_constant_variable(shape, value=0.1):
initial = tf.constant(value, shape=shape)
return tf.Variable(initial)
conv2d
具体参数含义见注释。
def conv2d(
inputs,
out_channels,
kernel_size,
scope,
stride=[1, 1],
padding='SAME',
stddev=1e-1,
activation_fn=tf.nn.relu):
"""
Args:
inputs: 4-D tensor variable BxHxWxC
output_channels: int
kernel_size: a list of 2 ints
scope: string
stride: a list of 2 ints
padding: 'SAME' or 'VALID'
stddev: float, stddev for truncated_normal init
activation_fn: function
Returns:
Variable tensor
"""
with tf.variable_scope(scope) as sc:
kernel_h, kernel_w = kernel_size
in_channels = inputs.shape[-1].value
kernel_shape = [kernel_h, kernel_w, in_channels, out_channels]
kernel = get_variable(kernel_shape, stddev)
stride_h, stride_w = stride
outputs = tf.nn.conv2d(inputs, kernel, strides=[1, stride_h, stride_w, 1], padding=padding)
if activation_fn is not None:
outputs = activation_fn(outputs)
return outputs
max_pool2d
def max_pool2d(inputs,
kernel_size,
scope,
stride=[2, 2],
padding='SAME'):
""" 2D max pooling.
Args:
inputs: 4-D tensor BxHxWxC
kernel_size: a list of 2 ints
stride: a list of 2 ints
Returns:
Variable tensor
"""
with tf.variable_scope(scope) as sc:
kernel_h, kernel_w = kernel_size
stride_h, stride_w = stride
outputs = tf.nn.max_pool(inputs,
ksize=[1, kernel_h, kernel_w, 1],
strides=[1, stride_h, stride_w, 1],
padding=padding,
)
return outputs
full_connection
def full_connection(
inputs,
num_outputs,
scope,
stddev=1e-1,
activation_fn=tf.nn.relu):
"""
Args:
inputs: 2-D tensor BxN
num_outputs: int
Returns:
Variable tensor of size B x num_outputs.
"""
with tf.variable_scope(scope) as sc:
num_inputs = inputs.shape[-1].value
weights = get_variable(shape=[num_inputs, num_outputs], stddev=stddev)
outputs = tf.matmul(inputs, weights)
biases = get_constant_variable(shape=[num_outputs])
outputs = tf.nn.bias_add(outputs, biases)
tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(0.01)(weights))
if activation_fn is not None:
outputs = activation_fn(outputs)
return outputs
dropout
虽然本项目中没用到，不过还是写一下。
def dropout(inputs, scope, keep_prob, is_training=False):
with tf.variable_scope(scope) as sc:
outputs = tf.nn.dropout(inputs, keep_prob=keep_prob)
return outputs
layer_util.py完整代码
import tensorflow as tf
def get_variable(shape, stddev=0.1):
initial = tf.truncated_normal(shape, stddev=stddev)
return tf.Variable(initial)
def get_constant_variable(shape, value=0.1):
initial = tf.constant(value, shape=shape)
return tf.Variable(initial)
def conv2d(
inputs,
out_channels,
kernel_size,
scope,
stride=[1, 1],
padding='SAME',
stddev=1e-1,
activation_fn=tf.nn.relu):
"""
Args:
inputs: 4-D tensor variable BxHxWxC
output_channels: int
kernel_size: a list of 2 ints
scope: string
stride: a list of 2 ints
padding: 'SAME' or 'VALID'
stddev: float, stddev for truncated_normal init
activation_fn: function
Returns:
Variable tensor
"""
with tf.variable_scope(scope) as sc:
kernel_h, kernel_w = kernel_size
in_channels = inputs.shape[-1].value
kernel_shape = [kernel_h, kernel_w, in_channels, out_channels]
kernel = get_variable(kernel_shape, stddev)
stride_h, stride_w = stride
outputs = tf.nn.conv2d(inputs, kernel, strides=[1, stride_h, stride_w, 1], padding=padding)
if activation_fn is not None:
outputs = activation_fn(outputs)
return outputs
def max_pool2d(inputs,
kernel_size,
scope,
stride=[2, 2],
padding='SAME'):
""" 2D max pooling.
Args:
inputs: 4-D tensor BxHxWxC
kernel_size: a list of 2 ints
stride: a list of 2 ints
Returns:
Variable tensor
"""
with tf.variable_scope(scope) as sc:
kernel_h, kernel_w = kernel_size
stride_h, stride_w = stride
outputs = tf.nn.max_pool(inputs,
ksize=[1, kernel_h, kernel_w, 1],
strides=[1, stride_h, stride_w, 1],
padding=padding,
)
return outputs
def full_connection(
inputs,
num_outputs,
scope,
stddev=1e-1,
activation_fn=tf.nn.relu):
"""
Args:
inputs: 2-D tensor BxN
num_outputs: int
Returns:
Variable tensor of size B x num_outputs.
"""
with tf.variable_scope(scope) as sc:
num_inputs = inputs.shape[-1].value
weights = get_variable(shape=[num_inputs, num_outputs], stddev=stddev)
outputs = tf.matmul(inputs, weights)
biases = get_constant_variable(shape=[num_outputs])
outputs = tf.nn.bias_add(outputs, biases)
tf.add_to_collection('losses', tf.contrib.layers.l2_regularizer(0.01)(weights))
if activation_fn is not None:
outputs = activation_fn(outputs)
return outputs
def dropout(inputs, scope, keep_prob, is_training=False):
with tf.variable_scope(scope) as sc:
outputs = tf.nn.dropout(inputs, keep_prob=keep_prob)
return outputs
lenet.py
根据leNent5网络定义，并返回输出和经过softmax后的结果
import tensorflow as tf
from utils import layer_util
def LeNet(inputs, keep_prob=None):
with tf.variable_scope('Conv1'):
conv1 = layer_util.conv2d(inputs, 6, [5, 5], 'conv1', padding='VALID')
with tf.variable_scope('S2'):
s2 = layer_util.max_pool2d(conv1, [2, 2], 'S2')
with tf.variable_scope('Conv3'):
conv3 = layer_util.conv2d(s2, 16, [5, 5], 'conv3', padding='VALID')
with tf.variable_scope('S4'):
s4 = layer_util.max_pool2d(conv3, [2, 2], 's4')
with tf.variable_scope('Conv5'):
conv5 = layer_util.conv2d(s4, 120, [5, 5], 'conv5')
flattened_shape = conv5.shape[1].value * conv5.shape[2].value * conv5.shape[3].value
conv5 = tf.reshape(conv5, [-1, flattened_shape])
with tf.variable_scope('F6'):
f6 = layer_util.full_connection(conv5, 84, 'f6')
with tf.variable_scope('output'):
outputs = layer_util.full_connection(f6, 10, 'outputs', activation_fn=None)
prediction = tf.nn.softmax(outputs)
return outputs,prediction
def get_model(inputs, keep_prob=None):
return LeNet(inputs, keep_prob)
train.py
加载数据。如果加载目录不存在数据会自动下载。如果网络问题下载失败或者下载速度很慢，可以从我的GitHub项目中获取。
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("./data/MNIST_data", one_hot=True) #one_hot为独热编码
定义超参数
BATCH_SIZE = 100
N_BATCH = mnist.train.num_examples // BATCH_SIZE
定义输入
with tf.variable_scope('inputs'):
x = tf.placeholder(tf.float32, [None, 784])
y = tf.placeholder(tf.float32, [None, 10])
因为网络输入为图片形式，所以需要对数据进行reshape
x_image = tf.reshape(x, [-1, 28, 28, 1])
获取模型的输出
outputs, prediction = lenet.get_model(x_image)
损失函数的定义，这里用了交叉熵损失
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=prediction))
使用梯度下降优化器
train_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss)
进行准确率的计算
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(prediction, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
train.py完整代码
import tensorflow as tf
from moudle import lenet
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("./data/MNIST_data", one_hot=True)
BATCH_SIZE = 100
N_BATCH = mnist.train.num_examples // BATCH_SIZE
def train():
with tf.variable_scope('inputs'):
x = tf.placeholder(tf.float32, [None, 784])
y = tf.placeholder(tf.float32, [None, 10])
x_image = tf.reshape(x, [-1, 28, 28, 1])
outputs, prediction = lenet.get_model(x_image)
loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=prediction))
train_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss)
init = tf.global_variables_initializer()
correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(prediction, 1))
accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
saver = tf.train.Saver()
with tf.Session() as sess:
writer = tf.summary.FileWriter('logs/', sess.graph)
sess.run(init)
for epoch in range(20):
for batch in range(N_BATCH):
batch_xs, batch_ys = mnist.train.next_batch(BATCH_SIZE)
sess.run(train_step, feed_dict={x: batch_xs, y: batch_ys})
pre = sess.run(prediction, feed_dict={x: mnist.test.images, y: mnist.test.labels})
acc = sess.run(accuracy, feed_dict={x: mnist.test.images, y: mnist.test.labels})
print('Iter' + str(epoch) + ",Testing Accuracy " + str(acc))
saver.save(sess, 'logs/train.ckpt')
if __name__ == '__main__':
train()
网络结构图
三、训练结果
可以进行一些消融实验，换个优化器或者多训练几轮，准确率最高应该可以到99%甚至100%。
