4用于cifar10的卷积神经网络-4.22为计算图中的非线性全连接层的权重添加L2损失
来源:互联网 发布:低空突防航迹规划算法 编辑:程序博客网 时间:2024/06/05 20:26
使用不同的L2权重损失系数观察网络的性能曲线变化
首先在网络的非线性全连接层为权重加入L2正则化损失
然后计算 总体损失 == 源自样本的经验损失(交叉熵损失) + 源自模型的正则化损失(权重的L2损失)
同时汇总多个不同的损失
总体损失 == 源自样本的经验损失(交叉熵损失) + ratio * 源自模型的正则化损失(权重的L2损失)
不断改变正则化损失在总体损失中的占比系数ratio,观察网络的性能曲线如何变化?
代码:
#-*- coding:utf-8 -*-#实现简单卷积神经网络对MNIST数据集进行分类:conv2d + activation + pool + fcimport csvimport tensorflow as tfimport osfrom tensorflow.examples.tutorials.mnist import input_dataos.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'import sysfrom six.moves import urllibimport tarfileimport cifar10_inputimport numpy as np# 设置算法超参数learning_rate_init = 0.001# learning_rate_init = 0.01# learning_rate_init = 0.1training_epochs = 6batch_size = 100display_step = 20conv1_kernel_num = 32conv2_kernel_num = 32fc1_units_num = 192fc2_units_num = 96activation_func = tf.nn.reluactivation_name = 'relu'l2loss_ratio = 0.005# Network Parametersn_input = 784 # MNIST data input (img shape: 28*28)#数据集中输入图像的参数dataset_dir='../CIFAR10_dataset'num_examples_per_epoch_for_train = cifar10_input.NUM_EXAMPLES_PER_EPOCH_FOR_TRAIN#50000num_examples_per_epoch_for_eval = cifar10_input.NUM_EXAMPLES_PER_EPOCH_FOR_EVAL#10000image_size = cifar10_input.IMAGE_SIZE#24image_channel = 3n_classes = cifar10_input.NUM_CLASSES #10个分类:CiFar10 中类的数量#从网址下载数据集存放到data_dir指定的目录中def maybe_download_and_extract(data_dir): """下载并解压缩数据集 from Alex's website.""" dest_directory = data_dir DATA_URL = 'http://www.cs.toronto.edu/~kriz/cifar-10-binary.tar.gz' if not os.path.exists(dest_directory): os.makedirs(dest_directory) filename = DATA_URL.split('/')[-1] #'cifar-10-binary.tar.gz' filepath = os.path.join(dest_directory, filename)#'../CIFAR10_dataset\\cifar-10-binary.tar.gz' if not os.path.exists(filepath): def _progress(count, block_size, total_size): sys.stdout.write('\r>> Downloading %s %.1f%%' % (filename, float(count * block_size) / float(total_size) * 100.0)) sys.stdout.flush() filepath, _ = urllib.request.urlretrieve(DATA_URL, filepath, _progress) print() statinfo = os.stat(filepath) print('Successfully downloaded', filename, statinfo.st_size, 'bytes.') extracted_dir_path = os.path.join(dest_directory, 'cifar-10-batches-bin')#'../CIFAR10_dataset\\cifar-10-batches-bin' if not os.path.exists(extracted_dir_path): tarfile.open(filepath, 'r:gz').extractall(dest_directory)def get_distorted_train_batch(data_dir,batch_size): """Construct distorted input for CIFAR training using the Reader ops. Returns: images: Images. 4D tensor of [batch_size, IMAGE_SIZE, IMAGE_SIZE, 3] size. labels: Labels. 1D tensor of [batch_size] size. Raises: ValueError: If no data_dir """ if not data_dir: raise ValueError('Please supply a data_dir') data_dir = os.path.join(data_dir, 'cifar-10-batches-bin') images, labels = cifar10_input.distorted_inputs(data_dir=data_dir,batch_size=batch_size) return images,labelsdef get_undistorted_eval_batch(data_dir,eval_data, batch_size): """Construct input for CIFAR evaluation using the Reader ops. Args: eval_data: bool, indicating if one should use the train or eval data set. Returns: images: Images. 4D tensor of [batch_size, IMAGE_SIZE, IMAGE_SIZE, 3] size. labels: Labels. 1D tensor of [batch_size] size. Raises: ValueError: If no data_dir """ if not data_dir: raise ValueError('Please supply a data_dir') data_dir = os.path.join(data_dir, 'cifar-10-batches-bin') images, labels = cifar10_input.inputs(eval_data=eval_data,data_dir=data_dir,batch_size=batch_size) return images,labels#根据指定的维数返回初始化好的指定名称的权重 Variabledef WeightsVariable(shape, name_str, stddev=0.1): # initial = tf.random_normal(shape=shape, stddev=stddev, dtype=tf.float32) initial = tf.truncated_normal(shape=shape, stddev=stddev, dtype=tf.float32) return tf.Variable(initial, dtype=tf.float32, name=name_str)#根据指定的维数返回初始化好的指定名称的偏置 Variabledef BiasesVariable(shape, name_str, init_value=0.00001): initial = tf.constant(init_value, shape=shape) return tf.Variable(initial, dtype=tf.float32, name=name_str)# 二维卷积层activation(conv2d+bias)的封装def Conv2d(x, W, b, stride=1, padding='SAME',activation=tf.nn.relu,act_name='relu'): with tf.name_scope('conv2d_bias'): y = tf.nn.conv2d(x, W, strides=[1, stride, stride, 1], padding=padding) y = tf.nn.bias_add(y, b) with tf.name_scope(act_name): y = activation(y) return y# 二维池化层pool的封装def Pool2d(x, pool= tf.nn.max_pool, k=2, stride=2,padding='SAME'): return pool(x, ksize=[1, k, k, 1], strides=[1, stride, stride, 1], padding=padding)# 全连接层activate(wx+b)的封装def FullyConnected(x, W, b, activation=tf.nn.relu, act_name='relu'): with tf.name_scope('Wx_b'): y = tf.matmul(x, W) y = tf.add(y, b) with tf.name_scope(act_name): y = activation(y) return y#为每一层的激活输出添加汇总节点def AddActivationSummary(x): tf.summary.histogram('/activations',x) tf.summary.scalar('/sparsity',tf.nn.zero_fraction(x))#为所有损失节点添加(滑动平均)标量汇总操作def AddLossesSummary(losses): #计算所有(individual losses)和(total loss)的滑动平均 loss_averages = tf.train.ExponentialMovingAverage(0.9,name='avg') loss_averages_op = loss_averages.apply(losses) #为所有individual losses 和 total loss 绑定标量汇总节点 #为所有平滑处理过的individual losses 和 total loss也绑定标量汇总节点 for loss in losses: #没有平滑过的loss名字后面加上‘(raw)’,平滑以后的loss使用其原来的名称 tf.summary.scalar(loss.op.name + '(raw)',loss) tf.summary.scalar(loss.op.name + '(avg)',loss_averages.average(loss)) return loss_averages_op#修改了4处激活函数:Conv2d_1、Conv2d_2、FC1_nonlinear、FC2_nonlineardef Inference(image_holder): # 第一个卷积层activate(conv2d + biase) with tf.name_scope('Conv2d_1'): # conv1_kernel_num = 64 weights = WeightsVariable(shape=[5, 5, image_channel, conv1_kernel_num], name_str='weights',stddev=5e-2) biases = BiasesVariable(shape=[conv1_kernel_num], name_str='biases',init_value=0.0) conv1_out = Conv2d(image_holder, weights, biases, stride=1, padding='SAME',activation=activation_func,act_name=activation_name) AddActivationSummary(conv1_out) # 第一个池化层(pool 2d) with tf.name_scope('Pool2d_1'): pool1_out = Pool2d(conv1_out, pool=tf.nn.max_pool, k=3, stride=2,padding='SAME') # 第二个卷积层activate(conv2d + biase) with tf.name_scope('Conv2d_2'): # conv2_kernels_num = 64 weights = WeightsVariable(shape=[5, 5, conv1_kernel_num, conv2_kernel_num], name_str='weights', stddev=5e-2) biases = BiasesVariable(shape=[conv2_kernel_num], name_str='biases', init_value=0.0) conv2_out = Conv2d(pool1_out, weights, biases, stride=1, padding='SAME',activation=activation_func,act_name=activation_name) AddActivationSummary(conv2_out) # 第二个池化层(pool 2d) with tf.name_scope('Pool2d_2'): pool2_out = Pool2d(conv2_out, pool=tf.nn.max_pool, k=3, stride=2, padding='SAME') #将二维特征图变换为一维特征向量 with tf.name_scope('FeatsReshape'): features = tf.reshape(pool2_out, [batch_size,-1]) feats_dim = features.get_shape()[1].value # 第一个全连接层(fully connected layer) with tf.name_scope('FC1_nonlinear'): weights = WeightsVariable(shape=[feats_dim, fc1_units_num],name_str='weights',stddev=4e-2) biases = BiasesVariable(shape=[fc1_units_num], name_str='biases',init_value=0.1) fc1_out = FullyConnected(features, weights, biases, activation=activation_func,act_name=activation_name) AddActivationSummary(fc1_out) with tf.name_scope('L2_loss'): weight_loss = tf.multiply(tf.nn.l2_loss(weights),l2loss_ratio,name="fc1_weight_loss") tf.add_to_collection('losses',weight_loss) # 第二个全连接层(fully connected layer) with tf.name_scope('FC2_nonlinear'): weights = WeightsVariable(shape=[fc1_units_num, fc2_units_num],name_str='weights',stddev=4e-2) biases = BiasesVariable(shape=[fc2_units_num], name_str='biases',init_value=0.1) fc2_out = FullyConnected(fc1_out, weights, biases, activation=activation_func,act_name=activation_name) AddActivationSummary(fc2_out) with tf.name_scope('L2_loss'): weight_loss = tf.multiply(tf.nn.l2_loss(weights), l2loss_ratio, name="fc2_weight_loss") tf.add_to_collection('losses', weight_loss) # 第三个全连接层(fully connected layer) with tf.name_scope('FC3_linear'): fc3_units_num = n_classes weights = WeightsVariable(shape=[fc2_units_num, fc3_units_num],name_str='weights',stddev=1.0/fc2_units_num) biases = BiasesVariable(shape=[fc3_units_num], name_str='biases',init_value=0.0) logits = FullyConnected(fc2_out, weights, biases,activation=tf.identity, act_name='linear') AddActivationSummary(logits) return logitsdef TrainModel(): #调用上面写的函数构造计算图 with tf.Graph().as_default(): # 计算图输入 with tf.name_scope('Inputs'): image_holder = tf.placeholder(tf.float32, [batch_size, image_size,image_size,image_channel], name='images') labels_holder = tf.placeholder(tf.int32, [batch_size], name='labels') # 计算图前向推断过程 with tf.name_scope('Inference'): logits = Inference(image_holder) # 定义损失层(loss layer) with tf.name_scope('Loss'): cross_entropy = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels_holder,logits=logits) cross_entropy_loss = tf.reduce_mean(cross_entropy,name='xentropy_loss') tf.add_to_collection('losses',cross_entropy_loss) #总体损失(total loss)= 交叉熵损失 + 所有权重的L2损失 total_loss = tf.add_n(tf.get_collection('losses'),name='total_loss') average_losses = AddLossesSummary(tf.get_collection('losses') + [total_loss]) # 定义优化训练层(train layer) with tf.name_scope('Train'): learning_rate = tf.placeholder(tf.float32) global_step = tf.Variable(0, name='global_step', trainable=False, dtype=tf.int64) optimizer = tf.train.RMSPropOptimizer(learning_rate=learning_rate) # optimizer = tf.train.MomentumOptimizer(learning_rate=learning_rate,momentum=0.9) # optimizer = tf.train.GradientDescentOptimizer(learning_rate=learning_rate) # optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate) # optimizer = tf.train.AdagradOptimizer(learning_rate=learning_rate) # optimizer = tf.train.FtrlOptimizer(learning_rate=learning_rate) train_op = optimizer.minimize(total_loss,global_step=global_step) # 定义模型评估层(evaluate layer) with tf.name_scope('Evaluate'): top_K_op = tf.nn.in_top_k(predictions=logits,targets=labels_holder,k=1) #定义获取训练样本批次的计算节点 with tf.name_scope('GetTrainBatch'): image_train,labels_train = get_distorted_train_batch(data_dir=dataset_dir,batch_size=batch_size) # 定义获取测试样本批次的计算节点 with tf.name_scope('GetTestBatch'): image_test, labels_test = get_undistorted_eval_batch(eval_data=True,data_dir=dataset_dir, batch_size=batch_size) merged_summaries = tf.summary.merge_all() # 添加所有变量的初始化节点 init_op = tf.global_variables_initializer() print('把计算图写入事件文件,在TensorBoard里面查看') summary_writer = tf.summary.FileWriter(logdir='evaluate_results/weights/L2_loss') summary_writer.add_graph(graph=tf.get_default_graph()) summary_writer.flush() # 将评估结果保存到文件 results_list = list() # 写入参数配置 results_list.append(['learning_rate', learning_rate_init, 'training_epochs', training_epochs, 'batch_size', batch_size, 'conv1_kernel_num', conv1_kernel_num, 'conv2_kernel_num', conv2_kernel_num, 'fc1_units_num', fc1_units_num, 'fc2_units_num', fc2_units_num]) results_list.append(['train_step', 'train_loss','train_step', 'train_accuracy']) with tf.Session() as sess: sess.run(init_op) print('===>>>>>>>==开始训练集上训练模型==<<<<<<<=====') total_batches = int(num_examples_per_epoch_for_train / batch_size) print('Per batch Size:,',batch_size) print('Train sample Count Per Epoch:',num_examples_per_epoch_for_train) print('Total batch Count Per Epoch:', total_batches) #启动数据读取队列 tf.train.start_queue_runners() #记录模型被训练的步数 training_step = 0 # 训练指定轮数,每一轮的训练样本总数为:num_examples_per_epoch_for_train for epoch in range(training_epochs): #每一轮都要把所有的batch跑一遍 for batch_idx in range(total_batches): #运行获取训练数据的计算图,取出一个批次数据 images_batch ,labels_batch = sess.run([image_train,labels_train]) #运行优化器训练节点 _,loss_value,avg_losses = sess.run([train_op,total_loss,average_losses], feed_dict={image_holder:images_batch, labels_holder:labels_batch, learning_rate:learning_rate_init}) #每调用一次训练节点,training_step就加1,最终==training_epochs * total_batch training_step = sess.run(global_step) #每训练display_step次,计算当前模型的损失和分类准确率 if training_step % display_step == 0: #运行accuracy节点,计算当前批次的训练样本的准确率 predictions = sess.run([top_K_op], feed_dict={image_holder:images_batch, labels_holder:labels_batch}) #当前批次上的预测正确的样本量 batch_accuracy = np.sum(predictions)/batch_size results_list.append([training_step,loss_value,training_step,batch_accuracy]) print("Training Step:" + str(training_step) + ",Training Loss = " + "{:.6f}".format(loss_value) + ",Training Accuracy = " + "{:.5f}".format(batch_accuracy) ) #运行汇总节点 summaries_str = sess.run(merged_summaries,feed_dict= {image_holder:images_batch, labels_holder:labels_batch}) summary_writer.add_summary(summary=summaries_str,global_step=training_step) summary_writer.flush() summary_writer.close() print('训练完毕') print('===>>>>>>>==开始在测试集上评估模型==<<<<<<<=====') total_batches = int(num_examples_per_epoch_for_eval / batch_size) total_examples = total_batches * batch_size print('Per batch Size:,', batch_size) print('Test sample Count Per Epoch:', total_examples) print('Total batch Count Per Epoch:', total_batches) correct_predicted = 0 for test_step in range(total_batches): #运行获取测试数据的计算图,取出一个批次测试数据 images_batch,labels_batch = sess.run([image_test,labels_test]) #运行accuracy节点,计算当前批次的测试样本的准确率 predictions = sess.run([top_K_op], feed_dict={image_holder:images_batch, labels_holder:labels_batch}) #累计每个批次上的预测正确的样本量 correct_predicted += np.sum(predictions) accuracy_score = correct_predicted / total_examples print('---------->Accuracy on Test Examples:',accuracy_score) results_list.append(['Accuracy on Test Examples:',accuracy_score]) # 将评估结果保存到文件 results_file = open('evaluate_results/weights/evaluate_results(L2_loss).csv', 'w', newline='') csv_writer = csv.writer(results_file, dialect='excel') for row in results_list: csv_writer.writerow(row)def main(argv=None): maybe_download_and_extract(data_dir=dataset_dir) train_dir='/logs' if tf.gfile.Exists(train_dir): tf.gfile.DeleteRecursively(train_dir) tf.gfile.MakeDirs(train_dir) TrainModel()if __name__ =='__main__': tf.app.run()
阅读全文
0 0
- 4用于cifar10的卷积神经网络-4.22为计算图中的非线性全连接层的权重添加L2损失
- 3用于MNIST的卷积神经网络-3.10为非线性全连接层添加正则化损失
- 4用于cifar10的卷积神经网络-4.2/4.3cifar10卷积神经网络的计算图设计(上下)
- 4用于cifar10的卷积神经网络-4.8/4.9为cifar10卷积网络添加汇总操作上/下
- 3用于MNIST的卷积神经网络-3.9增加一个非线性全连接层并观察过拟合现象
- 4用于cifar10的卷积神经网络-4.24Tensorflow中LRN层的工作原理以及添加方法
- 4用于cifar10的卷积神经网络-4.1用于cifar10的卷积神经网络结构设计
- 4用于cifar10的卷积神经网络-4.23正则化损失的比重对模型正确率曲线的影响分析
- 4用于cifar10的卷积神经网络-4.17cifar10卷积网络在不同激活函数下的性能分析
- 4用于cifar10的卷积神经网络-4.25/4.26Tensorflow中的四种学习率衰减方法(上/下)
- 4用于cifar10的卷积神经网络-4.6设计模型训练和评估的会话流程
- 4用于cifar10的卷积神经网络-4.7TensorFlow的汇总类Summary和FileWriter用法
- 4用于cifar10的卷积神经网络-4.11优化器GradientDescenOptimizer的汇总结果分析
- 4用于cifar10的卷积神经网络-4.12优化器MomentumOptimizer的汇总结果分析
- 4用于cifar10的卷积神经网络-4.13优化器FtrlOptimizer的汇总结果分析
- 4用于cifar10的卷积神经网络-4.14优化器AdagradOptimizer的汇总结果分析
- 4用于cifar10的卷积神经网络-4.15优化器RMSProbOptimizer的汇总结果分析
- 4用于cifar10的卷积神经网络-4.16优化器AdamOptimizer的汇总结果分析
- 用PHP快速开发一个爬虫,用到的开源库
- powerdesigner comment 写到 name 和 name 写到 comment
- github命令大全
- 封装一个小JS模块
- rand和srand随机函数的用法
- 4用于cifar10的卷积神经网络-4.22为计算图中的非线性全连接层的权重添加L2损失
- 总结echarts的常用手法
- SAP 修改字段长度
- 能看懂就看吧
- Java transient 关键字解释
- Java作业杨枝11.25/11.26
- 10款强大的开源 Javascript 图表库
- JQuery 对节点数据的遍历
- 内容中的 position : fixed 竟然没相对窗口来定位