卷积神经网络0-9数字识别

西瓜味的苹果 · 发表于 2020-12-30 13:50:01

马上注册，结交更多好友，享用更多功能^_^

您需要登录才可以下载或查看，没有账号？立即注册

x

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
import numpy as np
from PIL import Image
#算法超参数
batch_size=100
training_epochs=10
learning_rate_init=0.01
display_step=100
#网络参数
n_input=784
n_classes=10
#根据指定的维数返回指定的权数
def WeightsVariable(shape,name_str,stddev=0.1):
"""
:param shape: 形状，n行n列
:param name_str: 变量名称
:param stddev: 正态分布标准差
:return:
"""
initial=tf.random_normal(shape=shape,stddev=stddev,dtype=tf.float32)
return tf.Variable(initial_value=initial,dtype=tf.float32,name=name_str)
#根据指定维数返回指定的偏置
def BiasesVariable(shape,name_str,stddev=0.1):
"""
:param shape: 偏置的形状，nxn
:param name_str: 变量名称
:param stddev: 正态分布标准差
:return:
"""
initial=tf.random_normal(shape=shape,stddev=stddev,dtype=tf.float32)
return tf.Variable(initial_value=initial,dtype=tf.float32,name=name_str)
#二维卷积层的封装(conv2d+bias)
def Conv2d(X,w,b,stride=1,padding="SAME"):
"""
:param X: 特征图输入n*n*None
:param w: 滤波器(权重参数)
:param b: 修正（偏置参数）
:param stride: 滑动步长（默认为1）
:param padding: 外填充
:return:
"""
with tf.name_scope("Conv2d"):
y=tf.nn.conv2d(X,w,strides=[1,stride,stride,1],padding=padding)
y=tf.nn.bias_add(y,b)
return y
#非线性激活层的封装
def Activation(x,activation=tf.nn.relu,name="relu"):
"""
:param x: 特征图输入
:param activation: 激活函数，类似于看待事物的一种模式
:param name: 名称
:return:
"""
with tf.name_scope(name):
y=activation(x)
return y
#最大池化层的封装
def Pool2d(x,pool=tf.nn.max_pool,k=2,stride=2):
return pool(x,ksize=[1,k,k,1],strides=[1,stride,stride,1],padding="VALID")
#全连接层封装
def FullyConnect(x,w,b,activate=tf.identity,act_name="identity"):
with tf.name_scope("Wx_b"):
y=tf.add(tf.matmul(x,w),b)
with tf.name_scope("SoftMax"):
y=tf.nn.softmax(logits=y)
with tf.name_scope(act_name):
y=activate(y)
return y
#通用的评估函数，用来评估模型在给定的数据集上的损失和准确率
def EvaluateMode10nDataset(sess,images,labels):
n_samples=images.shape[0] #传入的样本数量
per_batch_size=batch_size
loss=0
acc=0
if(n_samples<=per_batch_size):
batch_count=1
loss,acc=sess.run([cross_entropy_loss,accuracy],
feed_dict={X_origin:images,
Y_true:labels,
learning_rate:learning_rate_init})
else:
batch_count=int(n_samples/per_batch_size)
batch_start=0
for idx in range(batch_count):
batch_loss,batch_acc=sess.run([cross_entropy_loss,accuracy],
feed_dict={X_origin:images[batch_start:batch_start+per_batch_size,:],
Y_true:labels[batch_start:batch_start+per_batch_size,:],
learning_rate:learning_rate_init})
batch_start+=per_batch_size
#累计所有批次上的损失和准确率
loss+=batch_loss
acc+=batch_acc
return loss/batch_count,acc/batch_count
'''计算图绘制'''
with tf.Graph().as_default():
#输入层:28*28*1
with tf.name_scope("Input"):
X_origin=tf.placeholder(tf.float32,shape=[None,n_input],name="X_origin")
Y_true=tf.placeholder(tf.float32,shape=[None,n_classes],name="Y_true")
X_image=tf.reshape(X_origin,[-1,28,28,1],name="X_image")
#前向推断过程
with tf.name_scope("Inference"):
#卷积层第一层:24*24*16
with tf.name_scope("Conv2d_1"):
weight_1=WeightsVariable(shape=[5,5,1,16],name_str="weight_1")
bias_1=BiasesVariable(shape=[16],name_str="bias_1")
conv1_out=Conv2d(X_image,weight_1,bias_1,stride=1,padding="VALID")
#非线性激活层
with tf.name_scope("Active"):
activation_out=Activation(x=conv1_out,activation=tf.nn.relu,name="relu")
#池化层（默认采用最大池化）12*12*16
with tf.name_scope("Pool2d_1"):
pool1_out=Pool2d(x=activation_out,pool=tf.nn.max_pool,k=2,stride=2)
#将二维特征图转变为1维,2304
with tf.name_scope("FeatsReshape"):
features=tf.reshape(pool1_out,[-1,12*12*16])
#全连接层
with tf.name_scope("FC_linear1"):
weight_fc=WeightsVariable(shape=[12*12*16,n_classes],name_str="weight_fc")
bias_fc=BiasesVariable(shape=[n_classes],name_str="bias_fc")
Ypre_logits=FullyConnect(features,weight_fc,bias_fc,activate=tf.identity,act_name="identity")
#定义损失层
with tf.name_scope("Loss"):
cross_entropy_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(
labels=Y_true, logits=Ypre_logits
))
#定义训练优化层
with tf.name_scope("Train"):
global_step = tf.Variable(0, name='global_step', trainable=False)
learning_rate=tf.placeholder(tf.float32)
optimizer=tf.train.AdamOptimizer(learning_rate=learning_rate)
trainer=optimizer.minimize(cross_entropy_loss,global_step=global_step)
#定义模型评估层
with tf.name_scope("Evaluate"):
correct_pred=tf.equal(tf.argmax(Ypre_logits,1),tf.argmax(Y_true,1))
accuracy=tf.reduce_mean(tf.cast(correct_pred,tf.float32))
#定义初始化所有变量节点
init=tf.initialize_all_variables()
print("将计算图写入log事件文件中，并在tensorboard中查看！！")
writer=tf.summary.FileWriter("logs/",graph=tf.get_default_graph())
writer.close()
# 导入mnist数据集
mnist=input_data.read_data_sets("MNIST_data/",one_hot=True)
'''启动会话'''
with tf.Session() as sess:
sess.run(init) #初始化variable变量
total_batches=int(mnist.train.num_examples/batch_size)
print("每批次的样本数量：",batch_size)
print("总共的批次数量:",total_batches)
print("总共的训练数据:",mnist.train.num_examples)
#保存和载入网络
saver=tf.train.Saver()
checkpoint=tf.train.get_checkpoint_state("saver_network")
if checkpoint and checkpoint.model_checkpoint_path:
saver.restore(sess,checkpoint.model_checkpoint_path)
print("Successfully loaded:", checkpoint.model_checkpoint_path)
else:
print("Could not find old network weights")
training_step=0 #记录模型被训练的步数
#指定训练轮数，将所有的样本都训练一遍
for epoch in range(training_epochs):
#把一轮所有的batch都跑一遍
for batch_idx in range(total_batches):
# 取出数据
batch_x, batch_y = mnist.train.next_batch(batch_size)
#训练优化器训练节点
sess.run(trainer,feed_dict={
X_origin:batch_x,
Y_true:batch_y,
learning_rate:learning_rate_init,
})
#每调用一次训练节点，training_step就加1
training_step=training_step+1
# 每训练display_step次，就计算当前模型的损失和准确率
if training_step % display_step == 0:
start_idx = max(0, (batch_idx - display_step) * batch_size)
end_idx = batch_idx * batch_size
train_loss, train_acc = EvaluateMode10nDataset(sess,
mnist.train.images[start_idx:end_idx, :],
mnist.train.labels[start_idx:end_idx, :])
print("Training Step:" + str(training_step) +
",Training_Loss={:.6f}".format(train_loss) +
",Training_accuracy={:.5f}".format(train_acc))
# 计算当前模型在验证集上的损失和准确率
validation_loss, validation_acc = EvaluateMode10nDataset(sess,
mnist.validation.images,
mnist.validation.labels)
print("Training Step:" + str(training_step) +
",Validation_Loss={:.6f}".format(validation_loss) +
",Validation_accuracy={:.5f}".format(validation_acc))
if training_step%display_step==0:
saver.save(sess,"./saver_network/check_mnist",global_step=global_step)
print("训练完毕！！！")
img = Image.open("5.png")
img = img.convert('L') # 灰度化
img = np.reshape(img, (784, 1)).reshape(1, 784)
img = ((255 - np.array(img, dtype=np.uint8)) / 255.0).reshape(1, 784)
logits=sess.run(Ypre_logits,feed_dict={
X_origin:img
})
print("预测的图像数字为:",np.argmax(logits))
if __name__ == '__main__':
pass

复制代码

西瓜味的苹果 · 发表于 2020-12-30 13:53:31

因为gui界面功能尚未完善，因此将主要逻辑上传至此（可以运行），后续持续更新代码完善

tryhi · 发表于 2020-12-30 16:45:47

高手啊

江晓夜 · 发表于 2021-1-9 15:13:48

为什么我运行提示没有名叫“tensorflow.examples”的模块

西瓜味的苹果 · 发表于 2021-1-12 22:09:14

江晓夜发表于 2021-1-9 15:13
为什么我运行提示没有名叫“tensorflow.examples”的模块

因为你用的2.0,要用1.0的版本

账号		自动登录	找回密码
密码			立即注册

[作品展示] 卷积神经网络0-9数字识别

马上注册，结交更多好友，享用更多功能^_^