利用TFRcord文件完成建立AlexNex tensorflow与深度学习

数据来源于猫狗大战，我在这里用分好了，如下图所示

文章图片

首先是建立TFRcord文件，如下所示：

import os import tensorflow as tf from PIL import Image#注意Image,后面会用到 import matplotlib.pyplot as plt import numpy as np cwd='D:\比赛\猫狗大战\cat_vs_dog\\train\\' classes={'cat','dog'} #人为设定 2 类 writer= tf.python_io.TFRecordWriter("D:\比赛\猫狗大战\cat_vs_dog\\train\\train.tfrecords") #要生成的文件 for index,name in enumerate(classes): class_path=cwd+name+'\\' for img_name in os.listdir(class_path): img_path=class_path+img_name #每一个图片的地址 img=Image.open(img_path) img= img.resize((208,208)) img_raw=img.tobytes()#将图片转化为二进制格式 example = tf.train.Example(features=tf.train.Features(feature={ "label": tf.train.Feature(int64_list=tf.train.Int64List(value=https://www.it610.com/article/[index])),'image_raw': tf.train.Feature(bytes_list=tf.train.BytesList(value=https://www.it610.com/article/[img_raw])) })) #example对象对label和image数据进行封装 writer.write(example.SerializeToString())#序列化为字符串 writer.close()

这样就可以建立tfrecords文件了

文章图片

如何读入文件？如下所示：

def read_and_decode(tfrecords_file, batch_size=5,epoch_num=3): '''read and decode tfrecord file, generate (image, label) batches Args: tfrecords_file:tfrecord 文件地址 batch_size: 每个batch大小，我计算机性能差，设置的比较小 epoch_num:重复多少次 Returns: image: 4D tensor - [batch_size, width, height, channel] label: 1D tensor - [batch_size] ''' # make an input queue from the tfrecord file filename_queue = tf.train.string_input_producer([tfrecords_file])reader = tf.TFRecordReader() _, serialized_example = reader.read(filename_queue)img_features = tf.parse_single_example( serialized_example, features={ 'label': tf.FixedLenFeature([], tf.int64), 'image_raw': tf.FixedLenFeature([], tf.string), }) image = tf.decode_raw(img_features['image_raw'], tf.uint8) label = tf.cast(img_features['label'], tf.float32) ########################################################## # you can put data augmentation here, I didn't use it ########################################################## # all the images of notMNIST are 28*28, you need to change the image size if you use other dataset.image = tf.reshape(image, [208, 208,3]) #label=tf.reshape(label, [batch_size,2]) #image.set_shape([208*208*3])image_batch, label_batch = tf.train.batch([image, label], batch_size= batch_size, num_threads= 64, capacity = 2000)#image = tf.image.per_image_standardization(image) #从tensor列表中按顺序或随机抽取一个tensor准备放入文件名称队列 input_queue = tf.train.slice_input_producer([image_batch, label_batch], num_epochs=epoch_num, shuffle=False) image_batch, label_batch = tf.train.batch(input_queue, batch_size, num_threads=2, capacity=10, allow_smaller_final_batch=False) image_batch=tf.cast(image_batch, tf.float32) label_batch=tf.cast(label_batch, tf.int32) return image_batch, tf.reshape(label_batch, [batch_size])

通过以上即可将文件读取数据了，那怎么使用这种tensor呢，我看到很多博客建议的是将tensor变为numpy的格式，再通过feed_dict填补数据，这种方法非常愚蠢，完全没有利用到TFRcord文件的优势。我接下来就利用猫狗大战的数据，写个简单的方法实现AlexNet。
首先是定义网格参数以及基础函数：

#定义网络超参数 learning_rate = 0.0001 train_epochs = 5 batch_size = 5 display = 10 t = 50# 每轮随机抽的次数 # 网络参数 n_input = 208*208 n_classes = 2keep_prob= 0.75 # 卷积层使用截断正太分布初始化,shape为列表对象def conv_weight_init(shape,stddev): weight = tf.Variable(tf.truncated_normal(dtype=tf.float32, shape=shape, stddev=stddev)) return weight# 全连接层使用xavier初始化 def xavier_init(layer1, layer2, constant = 1): Min = -constant * np.sqrt(6.0 / (layer1 + layer2)) Max = constant * np.sqrt(6.0 / (layer1 + layer2)) weight = tf.random_uniform((layer1, layer2), minval = Min, maxval = Max, dtype = tf.float32) return tf.Variable(weight)# 偏置 def biases_init(shape): biases = tf.Variable(tf.random_normal(shape, dtype=tf.float32)) return biases #卷积 +池化＋ＬＲＮ def conv2d(image, weight, stride=1): return tf.nn.conv2d(image, weight, strides=[1,stride,stride,1],padding='SAME')def max_pool(tensor, k = 2, stride=2): return tf.nn.max_pool(tensor, ksize=[1,k,k,1], strides=[1,stride,stride,1],padding='SAME')def lrnorm(tensor): return tf.nn.lrn(tensor,depth_radius=4, bias=1.0, alpha=0.001/9.0, beta=0.75) # AlexNet中的ＬＲＮ不知道是否适合 ##卷积 def inference(images,batch_size=5): # 转换image的shape imgs = tf.reshape(images,[-1, 208, 208, 3]) imgs=tf.cast(imgs, tf.float32)# 初始化权重 # wc = weight convolution wc1 = conv_weight_init([5, 5, 3, 12], 0.05) wc2 = conv_weight_init([5, 5, 12, 32], 0.05) wc3 = conv_weight_init([3, 3, 32, 48], 0.05) wc4 = conv_weight_init([3, 3, 48, 48], 0.05) wc5 = conv_weight_init([3, 3, 48, 32], 0.05) # wf : weight fullconnection wf1 = xavier_init(26*26*32, 512) wf2 = xavier_init(512, 512) wf3 = xavier_init(512, 2)# 初始化偏置 bc1 = biases_init([12]) bc2 = biases_init([32]) bc3 = biases_init([48]) bc4 = biases_init([48]) bc5 = biases_init([32]) # full connection bf1 = biases_init([512]) bf2 = biases_init([512]) bf3 = biases_init([2]) # 卷积１ c1= conv2d(imgs, wc1)# 未激活 conv1 = tf.nn.relu(c1 + bc1) lrn1= lrnorm(conv1) pool1= max_pool(lrn1) # 卷积２ conv2 = tf.nn.relu(conv2d(pool1, wc2) + bc2) lrn2= lrnorm(conv2) pool2 = max_pool(lrn2) # 卷积３－－５，不在进行ＬＲＮ，会严重影响网络前馈，反馈的速度（效果还不明显） conv3 = tf.nn.relu(conv2d(pool2, wc3) + bc3) #pool3 = max_pool(conv3) # 卷积４ conv4 = tf.nn.relu(conv2d(conv3, wc4) + bc4) #pool4 = max_pool(conv4) # 卷积５ conv5 = tf.nn.relu(conv2d(conv4, wc5) + bc5) pool5 = max_pool(conv5)##?*26*26*32#pool5#全连接 # 转换pool5的shape reshape_p5 = tf.reshape(pool5, [-1,26*26*32]) fc1 = tf.nn.relu(tf.matmul(reshape_p5, wf1) + bf1) # 正则化 drop_fc1 = tf.nn.dropout(fc1, keep_prob) # full connect 2 fc2 = tf.nn.relu(tf.matmul(drop_fc1, wf2) + bf2) drop_fc2 = tf.nn.dropout(fc2, keep_prob) # full connect 3 (输出层)未激活 output = tf.matmul(drop_fc2, wf3) + bf3#?,2 y=tf.reshape(output, [-1,2]) return y#损失函数和优化器 def cost(x,y): pred=inference(x) #loss=tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred,labels=y))with tf.variable_scope('loss') as scope: cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits\ (logits=pred, labels=y, name='xentropy_per_example') loss = tf.reduce_mean(cross_entropy, name='loss') tf.summary.scalar(scope.name+'/loss', loss)return lossdef trainning(loss, learning_rate):with tf.name_scope('optimizer'): optimizer = tf.train.AdamOptimizer(learning_rate= learning_rate) global_step = tf.Variable(0, name='global_step', trainable=False) train_op = optimizer.minimize(loss, global_step= global_step) return train_op

【利用TFRcord文件完成建立AlexNex】以上基础函数建立完毕，我们读入数据以及运行以上函数

filename="D:\比赛\猫狗大战\cat_vs_dog\\train\\train.tfrecords" x,y = read_and_decode(filename, batch_size=5) train_logits = inference(x,batch_size=5) train_loss = cost(x, y) train_op = trainning(train_loss, learning_rate)with tf.Session() as sess:# 先执行初始化工作 sess.run(tf.global_variables_initializer()) sess.run(tf.local_variables_initializer())# 开启一个协调器 coord = tf.train.Coordinator() # 使用start_queue_runners 启动队列填充 threads = tf.train.start_queue_runners(sess, coord)try: while not coord.should_stop(): print ('******1******') # 获取每一个batch中batch_size个样本和标签 _, tra_loss = sess.run([train_op, train_loss]) print('train loss = %.2f,' %( tra_loss)) #image_batch_v, label_batch_v = sess.run([image_batch, label_batch]) #print(image_batch_v.shape, label_batch_v) except tf.errors.OutOfRangeError:#如果读取到文件队列末尾会抛出此异常 print("done! now lets kill all the threads……")finally: # 协调器coord发出所有线程终止信号 coord.request_stop() print('all threads are asked to stop!') #coord.join(threads) #把开启的线程加入主线程，等待threads结束 print('all threads are stopped!')

运行结果如下所示：