浅入浅出TensorFlow 4 — 训练CIFAR数据

1git clone https://github.com/tensorflow/models.git  

2cd models/tutorials/image/cifar10

  1#coding=utf-8  
  2import cifar10,cifar10_input  
  3import tensorflow as tf  
  4import numpy as np  
  5import time  
  6
  7# define max_iter_step  batch_size  
  8max_iter_step = 1000  
  9batch_size = 128  
 10
 11# define variable_with_weight_loss  
 12# 和之前定义的weight有所不同，   
 13# 这里定义附带loss的weight，通过权重惩罚避免部分权重系数过大，导致overfitting   
 14def variable_with_weight_loss(shape,stddev,w1):  
 15    var = tf.Variable(tf.truncated_normal(shape,stddev=stddev))  
 16    if w1 is not None:  
 17        weight_loss = tf.multiply(tf.nn.l2_loss(var),w1,name='weight_loss')  
 18        tf.add_to_collection('losses',weight_loss)  
 19    return var  
 20
 21# 下载数据集 － 调用cifar10函数下载并解压  
 22cifar10.maybe_download_and_extract()  
 23# 注意路径  
 24cifar_dir = './cifar-10-batches-bin'  
 25
 26# 采用 data augmentation进行数据处理  
 27# 生成训练数据，训练数据通过cifar10_input的distort变化   
 28images_train, labels_train = cifar10_input.distorted_inputs(data_dir=cifar_dir,batch_size=batch_size)  
 29# 测试数据（eval_data 测试数据）  
 30images_test,labels_test = cifar10_input.inputs(eval_data=True,data_dir=cifar_dir,batch_size=batch_size)  
 31
 32




    
# 创建输入数据，采用 placeholder  
 33x_input = tf.placeholder(tf.float32,[batch_size,24,24,3])  
 34y_input = tf.placeholder(tf.int32,[batch_size])  
 35
 36# 创建第一个卷积层 input:3(channel) kernel:64 size:5*5  
 37weight1 = variable_with_weight_loss(shape=[5,5,3,64],stddev=5e-2,w1=0.0)  
 38bias1 = tf.Variable(tf.constant(0.0,shape=[64]))  
 39conv1 = tf.nn.conv2d(x_input,weight1,[1,1,1,1],padding='SAME')  
 40relu1 = tf.nn.relu(tf.nn.bias_add(conv1,bias1))  
 41pool1 = tf.nn.max_pool(conv1,ksize=[1,3,3,1],strides=[1,2,2,1],padding='SAME')  
 42norm1 = tf.nn.lrn(pool1,4,bias=1.0,alpha=0.001/9.0,beta=0.75)  
 43
 44# 创建第二个卷积层 input:64 kernel:64 size:5*5  
 45weight2 = variable_with_weight_loss(shape=[5,5,64,64],stddev=5e-2,w1=0.0)  
 46bias2 = tf.Variable(tf.constant(0,1,shape=[64]))  
 47conv2 = tf.nn.conv2d(norm1,weight2,[1,1,1,1],padding='SAME')  
 48relu2 = tf.nn.relu(tf.nn.bias_add(conv2,bias2))  
 49norm2 = tf.nn.lrn(relu2,4,bias=1.0,alpha=0.001/9.0,beta=0.75)  
 50pool2 = tf.nn.max_pool(norm2,ksize=[1,3,3,1],strides=[1,2,2,1],padding='SAME')  
 51
 52# 创建第三个层－全连接层  output:384  
 53reshape = tf.reshape(pool2,[batch_size,-1])  
 54dim = reshape.get_shape()[1].value  
 55weight3 = variable_with_weight_loss(shape=[dim,384],stddev=0.04,w1=0.004)  
 56bias3 = tf.Variable(tf.constant(0.1,shape=[384]))  
 57local3 = tf.nn.relu(tf.matmul(reshape,weight3)+bias3)  
 58
 59# 创建第四个层－全连接层  output:192  
 60weight4 = variable_with_weight_loss(shape=[384,192],stddev=0.04,w1=0.004)  
 61bias4 = tf.Variable(tf.constant(0.1,shape=[192]))  
 62local4 = tf.nn.relu(tf.matmul(local3,weight4)+bias4)  
 63
 64# 最后一层  output:10  
 65weight5 = variable_with_weight_loss(shape=[192,10],stddev=1/192.0,w1=0.0)  
 66bias5 = tf.Variable(tf.constant(0.0,shape=[10]))  
 67results = tf.add(tf.matmul(local4,weight5),bias5)  
 68
 69# 定义loss  
 70def loss(results,labels):  
 71    labels = tf.cast(labels,tf.int64)  
 72    cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=results,labels=labels,name='cross_entropy_per_example')  
 73    cross_entropy_mean = tf.reduce_mean(cross_entropy,name='cross_entropy')  
 74    tf.add_to_collection('losses',cross_entropy_mean)  
 75    return tf.add_n(tf.get_collection('losses'),name='total_loss')  
 76
 77# 计算loss  
 78loss = loss(results,y_input)  
 79
 80train_op = tf.train.AdamOptimizer(1e-3).minimize(loss)  # Adam  
 81top_k_op = tf.nn.in_top_k(results,y_input,1)  # top1 准确率  
 82
 83sess = tf.InteractiveSession()         # 创建session  
 84# tf.global_variable_initializer().run() # 初始化全部模型  
 85tf.initialize_all_variables().run()
 86
 87tf.train.start_queue_runners()  # 启动多线程加速  
 88