tensorflow学习(4)卷积网络 - 个人主页

实现一个简单的卷积神经网络

import tensorflow as tf
import numpy as np
from tensorflow.examples.tutorials.mnist import input_data
mnist=input_data.read_data_sets('MNIST_data',one_hot=True)
#导入手写数字数据集

def weight_variable(shape):
    init=tf.truncated_normal(shape,stddev=0.1)
    return tf.Variable(init)
#定义滤波器也就是权重矩阵

def bias_variable(shape):
    init=tf.constant(0.1,shape=shape)
    return tf.Variable(init)
#定义偏置向量

def conv2d(x,W):
    return tf.nn.conv2d(x,W,strides=[1,1,1,1],padding='SAME')
#定义卷积运算

def max_pool_2x2(x):
    return tf.nn.max_pool(x,ksize=[1,2,2,1],strides=[1,2,2,1],padding='SAME')
#定义池化计算

xs=tf.placeholder(tf.float32,[None,784])#不确定输出图像的个数所以写None
ys=tf.placeholder(tf.float32,[None,10])
keep_prob=tf.placeholder(tf.float32)
#定义输入的数据包括输入输出数据和dropout参数

x_image=tf.reshape(xs,[-1,28,28,1])
W_conv1=weight_variable([5,5,1,32])
b_conv1=bias_variable([32])
h_conv1=tf.nn.relu(conv2d(x_image,W_conv1)+b_conv1)
h_pool1=max_pool_2x2(h_conv1)
#正向传播过程，需要手动计算卷积后的数据大小尺寸

W_conv2=weight_variable([5,5,32,64])
b_conv2=bias_variable([64])
h_conv2=tf.nn.relu(conv2d(h_pool1,W_conv2)+b_conv2)
h_pool2=max_pool_2x2(h_conv2)


h_pool2_flat=tf.reshape(h_pool2,[-1,7*7*64])
#转化成全链接网络
W_fc1=weight_variable([7*7*64,1024])
b_fc1=bias_variable([1024])
h_fc1=tf.nn.relu(tf.matmul(h_pool2_flat,W_fc1)+b_fc1)
h_fc1_drop=tf.nn.dropout(h_fc1,keep_prob=keep_prob)
#在全链接上加dorpout

W_fc2=weight_variable([1024,10])
b_fc2=bias_variable([10])
pr=tf.nn.softmax(tf.matmul(h_fc1_drop,W_fc2)+b_fc2)
#定义输出结果

loss=tf.reduce_mean(-tf.reduce_sum(ys*tf.log(pr+1e-10),reduction_indices=[1]))
#定义损失函数加一个数防止log中的数小于等于0
train_=tf.train.AdamOptimizer(0.001).minimize(loss)

seen=tf.Session()
seen.run(tf.global_variables_initializer())
#启动图初始化

for i in range(100):
    batch_xs,batch_ys=mnist.train.next_batch(100)
    seen.run(train_,feed_dict={xs:batch_xs,ys:batch_ys,keep_prob:0.5})
    print(seen.run(loss,feed_dict={xs:batch_xs,ys:batch_ys,keep_prob:0.5}))
    print(i)

FEATURED TAGS

c语言 c++ 面向对象指针容器 python 函数数据结构回归损失函数神经网络机器学习似然函数极大似然标准化深度学习卷积网络参数估计 beta分布数据处理 gradio 模型工程化网页模型加速工具 c++实现变量占位符 tensorflow 线性回归学习tensorflow HMM RNN 强化学习 LSTM pandas 不定长序列损失 pytorch 目标检测 RPN 非极大值抑制 ROIpooling VGG16 Transformer BERT Python 装饰器方法 Pytorch FPN 图像分类 CNN 多模态生成 GPT Tranformer 生成模型 audio

实现一个简单的卷积神经网络

CATALOG

FEATURED TAGS