用TensorFlow搭建一个万能的神经网络框架(持续更新)
博客作者:凌逆战
博客地址:https://www.cnblogs.com/LXP-Never/p/12774058.html
文章代码:https://github.com/LXP-Never/blog_data/tree/master/tensorflow_model
我一直觉得TensorFlow的深度神经网络代码非常困难且繁琐,对TensorFlow搭建模型也十分困惑,所以我近期阅读了大量的神经网络代码,终于找到了搭建神经网络的规律,各位要是觉得我的文章对你有帮助不妨点个赞,点个关注吧。
我个人把深度学习分为以下步骤:数据处理 --> 模型搭建 --> 构建损失 --> 模型训练 --> 模型评估
我先把代码放出来,然后一点一点来讲
# Author:凌逆战 # -*- encoding:utf-8 -*- # 修改时间:2020年5月31日 import time from tensorflow.examples.tutorials.mnist import input_data from nets.my_alex import alexNet from ops import * tf.flags.DEFINE_integer('batch_size', 50, 'batch size, default: 1') tf.flags.DEFINE_integer('class_num', 10, 'batch size, default: 1') tf.flags.DEFINE_integer('epochs', 10, 'batch size, default: 1') tf.flags.DEFINE_float('learning_rate', 1e-4, '初始学习率, 默认: 0.0002') tf.flags.DEFINE_string('checkpoints_dir', "checkpoints", '保存检查点的地址') FLAGS = tf.flags.FLAGS # 从MNIST_data/中读取MNIST数据。当数据不存在时,会自动执行下载 mnist = input_data.read_data_sets('./data', one_hot=True, reshape=False) # reshape=False (None, 28,28,1) # 用于第一层是卷积层 # reshape=False (None, 784) # 用于第一层是全连接层 # 我们看一下数据的shape print(mnist.train.images.shape) # 训练数据图片(55000, 28, 28, 1) print(mnist.train.labels.shape) # 训练数据标签(55000, 10) print(mnist.test.images.shape) # 测试数据图片(10000, 28, 28, 1) print(mnist.test.labels.shape) # 测试数据图片(10000, 10) print(mnist.validation.images.shape) # 验证数据图片(5000, 28, 28, 1) print(mnist.validation.labels.shape) # 验证数据图片(5000, 784) def train(): batch_size = FLAGS.batch_size # 一个batch训练多少个样本 batch_nums = mnist.train.images.shape[0] // batch_size # 一个epoch中应该包含多少batch数据 class_num = FLAGS.class_num # 分类类别数 epochs = FLAGS.epochs # 训练周期数 learning_rate = FLAGS.learning_rate # 初始学习率 ############ 保存检查点的地址 ############ checkpoints_dir = FLAGS.checkpoints_dir # checkpoints # 如果检查点不存在,则创建 if not os.path.exists(checkpoints_dir): os.makedirs(FLAGS.checkpoints_dir) ###################################################### # 创建图 # ###################################################### graph = tf.Graph() # 自定义图 # 在自己的图中定义数据和操作 with graph.as_default(): inputs = tf.placeholder(dtype="float", shape=[None, 28, 28, 1], name='inputs') labels = tf.placeholder(dtype="float", shape=[None, class_num], name='labels') # 看个人喜欢,有的人在初始化定义中就定义了learning_rate,有的人喜欢通过feed传learning_rate learning_rate = tf.placeholder("float", None, name='learning_rate') # 如果网络结构有dropout层,需要定义keep_probn,如果没有则不需要 # 训练的时候需要,测试的时候需要设置成1 keep_prob = tf.placeholder(dtype="float", name='keep_prob') ############ 搭建模型 ############ logits = alexNet(inputs, class_num, keep_prob=keep_prob) # 使用placeholder搭建模型 ############ 损失函数 ############ loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=labels, logits=logits)) tf.add_to_collection('losses', loss) total_loss = tf.add_n(tf.get_collection("losses")) # total_loss=模型损失+权重正则化损失 ############ 模型精度 ############ predict = tf.argmax(logits, 1) # 模型预测结果 accuracy = tf.reduce_mean(tf.cast(tf.equal(predict, tf.argmax(labels, 1)), tf.float32)) ############ 优化器 ############ variable_to_train = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES) # 可训练变量列表 # 创建优化器,更新网络参数,最小化loss, global_step = tf.Variable(0, trainable=False) learning_rate = tf.train.exponential_decay(learning_rate=learning_rate, # 初始学习率 global_step=global_step, decay_steps=batch_nums, # 多少步衰减一次 decay_rate=0.1, # 衰减率 staircase=True) # 以阶梯的形式衰减 # 移动平均值更新参数 # train_op = moving_average(loss, learning_rate, global_step) # adam优化器,adam算法好像会自动衰减学习率, train_op = tf.train.AdamOptimizer(learning_rate).minimize(loss=total_loss, global_step=global_step, var_list=variable_to_train) ############ TensorBoard可视化 summary ############ summary_writer = tf.summary.FileWriter(logdir="./logs", graph=graph) # 创建事件文件 tf.summary.scalar(name="losses", tensor=total_loss) # 收集损失值变量 tf.summary.scalar(name="acc", tensor=accuracy) # 收集精度值变量 tf.summary.scalar(name='learning_rate', tensor=learning_rate) merged_summary_op = tf.summary.merge_all() # 将所有的summary合并为一个op ############ 模型保存和恢复 Saver ############ saver = tf.train.Saver(max_to_keep=5) ###################################################### # 创建会话 # ###################################################### max_acc = 0. config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=True) with tf.Session(config=config, graph=graph) as sess: # 加载模型,如果模型存在返回 是否加载成功和训练步数 could_load, checkpoint_step = load_model(sess, saver, FLAGS.checkpoints_dir) if could_load: print(" [*] 模型加载成功") else: print(" [!] 模型加载失败") try: tf.global_variables_initializer().run() except: tf.initialize_all_variables().run() for epoch in range(epochs): for i in range(batch_nums): start_time = time.time() # batch_images = data_X[i * batch_size:(i + 1) * batch_size] # batch_labels = data_y[i * batch_size:(i + 1) * batch_size] train_batch_x, train_batch_y = mnist.train.next_batch(batch_size) # 使用真实数据填充placeholder,运行训练模型和合并变量操作 _, summary, loss, step = sess.run([train_op, merged_summary_op, total_loss, global_step], feed_dict={inputs: train_batch_x, labels: train_batch_y, keep_prob: 0.5}) if step % 100 == 0: summary_writer.add_summary(summary, step) # 将每次迭代后的变量写入事件文件 summary_writer.flush() # 强制summary_writer将缓存中的数据写入到日志文件中(可选) ############ 可视化打印 ############ print("Epoch:[%2d] [%4d/%4d] time:%4.4f,loss:%.8f" % ( epoch, i, batch_nums, time.time() - start_time, loss)) # 打印一些可视化的数据,损失... if step % 100 == 0: acc = sess.run(accuracy, feed_dict={inputs: mnist.validation.images, labels: mnist.validation.labels, keep_prob: 1.0}) print("Epoch:[%2d] [%4d/%4d] accuracy:%.8f" % (epoch, i, batch_nums, acc)) ############ 保存模型 ############ if acc > max_acc: max_acc = acc save_path = saver.save(sess, save_path=os.path.join(checkpoints_dir, "model.ckpt"), global_step=step) tf.logging.info("模型保存在: %s" % save_path) print("优化完成!") def main(argv=None): train() if __name__ == '__main__': # logging.basicConfig(level=logging.INFO) tf.logging.set_verbosity(tf.logging.INFO) tf.app.run()
# Author:凌逆战 # -*- encoding:utf-8 -*- # 修改时间:2020年5月31日 import time from tensorflow.examples.tutorials.mnist import input_data from nets.my_vgg import VGG16Net from ops import * tf.flags.DEFINE_integer('batch_size', 100, 'batch size, default: 1') tf.flags.DEFINE_integer('class_num', 10, 'batch size, default: 1') tf.flags.DEFINE_integer('epochs', 10, 'batch size, default: 1') tf.flags.DEFINE_float('learning_rate', 2e-4, '初始学习率, 默认: 0.0001') tf.flags.DEFINE_string('checkpoints_dir', "checkpoint", '保存检查点的地址') FLAGS = tf.flags.FLAGS # 从MNIST_data/中读取MNIST数据。当数据不存在时,会自动执行下载 mnist = input_data.read_data_sets('./MNIST_data', one_hot=True, reshape=False) # reshape=False (None, 28,28,1) # 用于第一层是卷积层 # reshape=False (None, 784) # 用于第一层是全连接层 # 我们看一下数据的shape print(mnist.train.images.shape) # 训练数据图片(55000, 28, 28, 1) print(mnist.train.labels.shape) # 训练数据标签(55000, 10) print(mnist.test.images.shape) # 测试数据图片(10000, 28, 28, 1) print(mnist.test.labels.shape) # 测试数据图片(10000, 10) print(mnist.validation.images.shape) # 验证数据图片(5000, 28, 28, 1) print(mnist.validation.labels.shape) # 验证数据图片(5000, 784) def train(): batch_size = FLAGS.batch_size batch_nums = mnist.train.images.shape[0] // batch_size # 一个epoch中应该包含多少batch数据 class_num = FLAGS.class_num epochs = FLAGS.epochs learning_rate = FLAGS.learning_rate ############ 保存检查点的地址 ############ checkpoints_dir = FLAGS.checkpoints_dir # checkpoints # 如果检查点不存在,则创建 if not os.path.exists(checkpoints_dir): os.makedirs(FLAGS.checkpoints_dir) ###################################################### # 创建图 # ###################################################### graph = tf.Graph() # 自定义图 # 在自己的图中定义数据和操作 with graph.as_default(): inputs = tf.placeholder(dtype="float", shape=[None, 28, 28, 1], name='inputs') labels = tf.placeholder(dtype="float", shape=[None, class_num], name='labels') ############ 搭建模型 ############ logits = VGG16Net(inputs, class_num) # 使用placeholder搭建模型 ############ 损失函数 ############ # 计算预测值和真实值之间的误差 loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=labels, logits=logits)) tf.add_to_collection('losses', loss) total_loss = tf.add_n(tf.get_collection("losses")) # total_loss=模型损失+权重正则化损失 ############ 模型精度 ############ predict = tf.argmax(logits, axis=1) accuracy = tf.reduce_mean(tf.cast(tf.equal(predict, tf.argmax(labels, axis=1)), tf.float32)) ############ 优化器 ############ variable_to_train = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES) # 可训练变量列表 # 创建优化器,更新网络参数,最小化loss, train_op = tf.train.AdamOptimizer(learning_rate).minimize(loss=total_loss, var_list=variable_to_train) ############ TensorBoard可视化 summary ############ summary_writer = tf.summary.FileWriter("./logs", graph=graph) # 创建事件文件 tf.summary.scalar(name="loss", tensor=total_loss) # 收集损失值变量 tf.summary.scalar(name='accuracy', tensor=accuracy) # 收集精度值变量 tf.summary.scalar(name='learning_rate', tensor=learning_rate) merged_summary_op = tf.summary.merge_all() # 将所有的summary合并为一个op ############ 模型保存和恢复 Saver ############ saver = tf.train.Saver(max_to_keep=5) ###################################################### # 创建会话 # ###################################################### max_acc = 0. config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=True) with tf.Session(config=config, graph=graph) as sess: # 加载模型,如果模型存在返回 是否加载成功和训练步数 could_load, checkpoint_step = load_model(sess, saver, FLAGS.checkpoints_dir) if could_load: step = checkpoint_step print(" [*] 模型加载成功") else: print(" [!] 模型加载失败") try: tf.global_variables_initializer().run() except: tf.initialize_all_variables().run() step = 0 for epoch in range(epochs): for i in range(batch_nums): start_time = time.time() # 记录一下开始训练的时间 # batch_images = data_X[i * batch_size:(i + 1) * batch_size] # batch_labels = data_y[i * batch_size:(i + 1) * batch_size] train_batch_x, train_batch_y = mnist.train.next_batch(batch_size) # 使用真实数据填充placeholder,运行训练模型和合并变量操作 _, summary, loss = sess.run([train_op, merged_summary_op, total_loss], feed_dict={inputs: train_batch_x, labels: train_batch_y}) if step % 100 == 0: summary_writer.add_summary(summary, step) # 将每次迭代后的变量写入事件文件 summary_writer.flush() # 强制summary_writer将缓存中的数据写入到日志文件中(可选) ############ 可视化打印 ############ print("Epoch:[%2d] [%4d/%4d] time:%4.4f,loss:%.8f" % ( epoch, i, batch_nums, time.time() - start_time, loss)) # 打印一些可视化的数据,损失... # if np.mod(step, 100) == 1 if step % 100 == 0: acc = sess.run(accuracy, {inputs: mnist.validation.images, labels: mnist.validation.labels}) print("Epoch:[%2d] [%4d/%4d],acc:%.8f" % (epoch, i, batch_nums, acc)) ############ 保存模型 ############ if acc > max_acc: max_acc = acc save_path = saver.save(sess, save_path=os.path.join(checkpoints_dir, "model.ckpt"), global_step=step) # logging.info("模型保存在: %s" % save_path) tf.logging.info("模型保存在: %s" % save_path) step += 1 print("优化完成!") def main(argv=None): train() if __name__ == '__main__': # logging.basicConfig(level=logging.INFO) tf.logging.set_verbosity(tf.logging.INFO) tf.app.run()
数据处理
数据处理因为每个专业领域的原因各不相同,而这不同点也是各位论文创新点的新方向。不同的我没法讲,但我总结了几点相同的地方——batch数据生成。因为深度学习模型需要一个batch一个batch的喂数据进行训练,所以我们的数据必须是batch的形式,这里衍生了三点问题
- 通过代码批量读取数据,
- 如何生成batch数据:由于篇幅过长,实在有很多地方要介绍和详述,我把这一块内容移到了这篇文章《TensorFlow读取数据的三种方法》中
- 数据的shape:我举两个例子让大家理解:图片数据为4维 (batch_size, height,width, channels),序列数据为3维 (batch_size, time_steps, input_size),
- 不同的shape处理方法不同,选择神经网络模型单元也不同。我会在后面细讲
模型搭建
阅读这一节我默认大家已经学会了数据的batch读取了。
模型搭建这一步很像我们小时候玩的搭积木,我这里以经典神经网络模型VGG、Alex、ResNet、Google Inception Net为例讲解,大家看代码看多了也会很简单的就找到,当然我是有一点私心的,我想把这些经典的网络在这篇文章做一个tensorflow实现汇总,我细讲第一个,大家可能看一个例子就懂了,看懂了就直接往下看,看不懂就多看几个。
LeNet5模型
论文:1998_LeNet_Gradient-Based Learning Applied to Document Recognition
下面我们定义一个LeNet5模型,我们先定义需要用到的神经网络单元,相同的代码尽量封装成函数的形式以节省代码量和简洁代码
def conv(input, kernel_size, output_size, stride, init_bias=0.0, padding="SAME", name=None, wd=None): input_size = input.shape[-1] conv_weights = tf.get_variable(name='weights', shape=[kernel_size, kernel_size, input_size, output_size], initializer=tf.truncated_normal_initializer(stddev=0.1), dtype=tf.float32) conv_biases = tf.get_variable(name='biases', shape=[output_size], initializer=tf.constant_initializer(init_bias), dtype=tf.float32) if wd is not None: # wd 0.004 # tf.nn.l2_loss(var)=sum(t**2)/2 weight_decay = tf.multiply(tf.nn.l2_loss(conv_weights), wd, name='weight_loss') tf.add_to_collection('losses', weight_decay) conv_layer = tf.nn.conv2d(input, conv_weights, [1, stride, stride, 1], padding=padding, name=name) # 卷积操作 conv_layer = tf.nn.bias_add(conv_layer, conv_biases) # 加上偏置项 conv_layer = tf.nn.relu(conv_layer) # relu激活函数 return conv_layer def fc(input, output_size, init_bias=0.0, activeation_func=True, wd=None): input_shape = input.get_shape().as_list() # 创建 全连接权重 变量 fc_weights = tf.get_variable(name="weights", shape=[input_shape[-1], output_size], initializer=tf.truncated_normal_initializer(stddev=0.1), dtype=tf.float32) if wd is not None: # wd 0.004 # tf.nn.l2_loss(var)=sum(t**2)/2 weight_decay = tf.multiply(tf.nn.l2_loss(fc_weights), wd, name='weight_loss') tf.add_to_collection('losses', weight_decay) # 创建 全连接偏置 变量 fc_biases = tf.get_variable(name="biases", shape=[output_size], initializer=tf.constant_initializer(init_bias), dtype=tf.float32) fc_layer = tf.matmul(input, fc_weights) # 全连接计算 fc_layer = tf.nn.bias_add(fc_layer, fc_biases) # 加上偏置项 if activeation_func: fc_layer = tf.nn.relu(fc_layer) # rele激活函数 return fc_layer
然后利用我们搭建的神经网络单元,搭建LeNet5神经网络模型
# 训练时:keep_prob=0.5 # 测试时:keep_prob=1.0 def leNet(inputs, class_num, keep_prob=0.5): # 第一层 卷积层 conv1 with tf.variable_scope('layer1-conv1'): conv1 = conv(input=inputs, kernel_size=5, output_size=32, stride=1, init_bias=0.0, name="layer1-conv1", padding="SAME") # 第二层 池化层 with tf.name_scope('layer2-pool1'): pool1 = tf.nn.max_pool(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') # 第三层 卷积层 conv2 with tf.variable_scope('layer3-conv2'): conv2 = conv(input=pool1, kernel_size=5, output_size=64, stride=1, init_bias=0.0, name="layer3-conv2", padding="SAME") # 第四层 池化层 with tf.name_scope('layer4-pool2'): pool2 = tf.nn.max_pool(conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') # 后面要做全连接,因此要把数据变成2维 # pool_shape = pool2.get_shape().as_list() pool_shape = pool2.shape flatten = tf.reshape(pool2, [-1, pool_shape[1] * pool_shape[2] * pool_shape[3]]) with tf.variable_scope('layer5-fcl'): fc1 = fc(input=flatten, output_size=512, init_bias=0.1, activeation_func=tf.nn.relu, wd=None) fc1 = tf.nn.dropout(fc1, keep_prob=keep_prob, name="dropout1") with tf.variable_scope('layer6-fc2'): logit = fc(input=fc1, output_size=class_num, init_bias=0.1, activeation_func=False, wd=None) return logit
Alex模型
论文:2012_Alex_ImageNet Classification with Deep Convolutional Neural Networks
下面我们定义一个Alex模型,我们先定义需要用到的神经网络单元,相同的代码尽量封装成函数的形式以节省代码量和简洁代码
def conv(input, kernel_size, output_size, stride, init_bias=0.0, padding="SAME", name=None, wd=None): input_size = input.shape[-1] conv_weights = tf.get_variable(name='weights', shape=[kernel_size, kernel_size, input_size, output_size], initializer=tf.random_normal_initializer(mean=0, stddev=0.01), dtype=tf.float32) if wd is not None: # wd 0.004 # tf.nn.l2_loss(var)=sum(t**2)/2 weight_decay = tf.multiply(tf.nn.l2_loss(conv_weights), wd, name='weight_loss') tf.add_to_collection('losses', weight_decay) conv_biases = tf.get_variable(name='biases', shape=[output_size], initializer=tf.constant_initializer(init_bias), dtype=tf.float32) conv_layer = tf.nn.conv2d(input, conv_weights, [1, stride, stride, 1], padding=padding, name=name) # 卷积操作 conv_layer = tf.nn.bias_add(conv_layer, conv_biases) # 加上偏置项 conv_layer = tf.nn.relu(conv_layer) # relu激活函数 return conv_layer
def fc(input, output_size, init_bias=0.0, activeation_func=True, wd=None): input_shape = input.get_shape().as_list() # 创建 全连接权重 变量 fc_weights = tf.get_variable(name="weights", shape=[input_shape[-1], output_size], initializer=tf.random_normal_initializer(mean=0.0, stddev=0.01), dtype=tf.float32) if wd is not None: # wd 0.004 # tf.nn.l2_loss(var)=sum(t**2)/2 weight_decay = tf.multiply(tf.nn.l2_loss(fc_weights), wd, name='weight_loss') tf.add_to_collection('losses', weight_decay) # 创建 全连接偏置 变量 fc_biases = tf.get_variable(name="biases", shape=[output_size], initializer=tf.constant_initializer(init_bias), dtype=tf.float32) fc_layer = tf.matmul(input, fc_weights) # 全连接计算 fc_layer = tf.nn.bias_add(fc_layer, fc_biases) # 加上偏置项 if activeation_func: fc_layer = tf.nn.relu(fc_layer) # rele激活函数 return fc_layer
def LRN(input, depth_radius=2, alpha=0.0001, beta=0.75, bias=1.0): """Local Response Normalization 局部响应归一化""" return tf.nn.local_response_normalization(input, depth_radius=depth_radius, alpha=alpha, beta=beta, bias=bias)
然后利用我们搭建的神经网络单元,搭建Alex神经网络模型
def alexNet(inputs, class_num, keep_prob=0.5): # 第一层卷积层 conv1 with tf.variable_scope("conv1"): conv1 = conv(input=inputs, kernel_size=7, output_size=96, stride=3, init_bias=0.0, name="conv1", padding="SAME") conv1 = LRN(conv1) conv1 = tf.nn.max_pool(conv1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID', name="pool1") # 第二层卷积层 conv2 with tf.variable_scope("conv2"): conv2 = conv(input=conv1, kernel_size=7, output_size=96, stride=3, init_bias=1.0, name="conv2", padding="SAME") conv2 = LRN(conv2) conv2 = tf.nn.max_pool(conv2, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID', name="pool2") # 第三层卷积层 conv3 with tf.variable_scope("conv3"): conv3 = conv(input=conv2, kernel_size=7, output_size=96, stride=3, init_bias=0.0, name="conv3", padding="SAME") # 第四层卷积层 conv4 with tf.variable_scope("conv4"): conv4 = conv(input=conv3, kernel_size=7, output_size=96, stride=3, init_bias=1.0, name="conv4", padding="SAME") # 第五层卷积层 conv5 with tf.variable_scope("conv5"): conv5 = conv(input=conv4, kernel_size=3, output_size=256, stride=1, init_bias=1.0, name="conv5") conv5 = tf.nn.max_pool(conv5, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='VALID', name="pool5") conv5_shape = conv5.shape # 后面做全连接,所以要把shape改成2维 # shape=[batch, dim] flatten = tf.reshape(conv5, [-1, conv5_shape[1] * conv5_shape[2] * conv5_shape[3]]) # 第一层全连接层 fc1 with tf.variable_scope("fc1"): fc1 = fc(input=flatten, output_size=4096, init_bias=1.0, activeation_func=tf.nn.relu, wd=None) fc1 = tf.nn.dropout(fc1, keep_prob=keep_prob, name="dropout1") # 第一层全连接层 fc2 with tf.variable_scope("fc2"): fc2 = fc(input=fc1, output_size=4096, init_bias=1.0, activeation_func=tf.nn.relu, wd=None) fc2 = tf.nn.dropout(fc2, keep_prob=keep_prob, name="dropout1") # 第一层全连接层 fc3 with tf.variable_scope("fc3"): logit = fc(input=fc2, output_size=class_num, init_bias=1.0, activeation_func=False, wd=None) return logit # 模型输出
VGG模型
论文:2014_VGG_Very Deep Convolutional Networks for Large-Scale Image Recognition
VGG有两个比较有名的网络:VGG16、VGG19,我在这里搭建VGG16,有兴趣的朋友可以按照上面的模型结构自己用TensorFlow搭建VGG19模型
下面我们定义一个VGG16模型,和前面一样,我们先定义需要用到的神经网络单元,相同的代码尽量封装成函数的形式以节省代码量和简洁代码
因为模型中同一个变量域中包含多个卷积操作,因此在卷积函数中套一层变量域
def conv(inputs, scope_name, kernel_size, output_size, stride, init_bias=0.0, padding="SAME", wd=None): input_size = int(inputs.get_shape()[-1]) with tf.variable_scope(scope_name): conv_weights = tf.get_variable(name='weights', shape=[kernel_size, kernel_size, input_size, output_size], dtype=tf.float32, initializer=tf.truncated_normal_initializer(mean=0.0, stddev=1e-1)) if wd is not None: # tf.nn.l2_loss(var)=sum(t**2)/2 weight_decay = tf.multiply(tf.nn.l2_loss(conv_weights), wd, name='weight_loss') tf.add_to_collection('losses', weight_decay) conv_biases = tf.get_variable(name='biases', shape=[output_size], dtype=tf.float32, initializer=tf.constant_initializer(init_bias)) conv_layer = tf.nn.conv2d(inputs, conv_weights, [1, stride, stride, 1], padding=padding, name=scope_name) conv_layer = tf.nn.bias_add(conv_layer, conv_biases) conv_layer = tf.nn.relu(conv_layer) return conv_layer
def fc(inputs, scope_name, output_size, init_bias=0.0, activeation_func=True, wd=None): input_shape = inputs.get_shape().as_list() with tf.variable_scope(scope_name): # 创建 全连接权重 变量 fc_weights = tf.get_variable(name="weights", shape=[input_shape[-1], output_size], dtype=tf.float32, initializer=tf.truncated_normal_initializer(mean=0.0, stddev=1e-1)) if wd is not None: # wd 0.004 # tf.nn.l2_loss(var)=sum(t**2)/2 weight_decay = tf.multiply(tf.nn.l2_loss(fc_weights), wd, name='weight_loss') tf.add_to_collection('losses', weight_decay) # 创建 全连接偏置 变量 fc_biases = tf.get_variable(name="biases", shape=[output_size], dtype=tf.float32, initializer=tf.constant_initializer(init_bias), trainable=True) fc_layer = tf.matmul(inputs, fc_weights) # 全连接计算 fc_layer = tf.nn.bias_add(fc_layer, fc_biases) # 加上偏置项 if activeation_func: fc_layer = tf.nn.relu(fc_layer) # rele激活函数 return fc_layer
然后利用我们搭建的神经网络单元,搭建VGG16神经网络模型
def VGG16Net(inputs, class_num): with tf.variable_scope("conv1"): # conv1_1 [conv3_64] conv1_1 = conv(inputs=inputs, scope_name="conv1_1", kernel_size=3, output_size=64, stride=1, init_bias=0.0, padding="SAME") # conv1_2 [conv3_64] conv1_2 = conv(inputs=conv1_1, scope_name="conv1_2", kernel_size=3, output_size=64, stride=1, init_bias=0.0, padding="SAME") pool1 = tf.nn.max_pool(conv1_2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool1') with tf.variable_scope("conv2"): # conv2_1 conv2_1 = conv(inputs=pool1, scope_name="conv2_1", kernel_size=3, output_size=128, stride=1, init_bias=0.0, padding="SAME") # conv2_2 conv2_2 = conv(inputs=conv2_1, scope_name="conv2_2", kernel_size=3, output_size=128, stride=1, init_bias=0.0, padding="SAME") pool2 = tf.nn.max_pool(conv2_2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool2') with tf.variable_scope("conv3"): # conv3_1 conv3_1 = conv(inputs=pool2, scope_name="conv3_1", kernel_size=3, output_size=256, stride=1, init_bias=0.0, padding="SAME") # conv3_2 conv3_2 = conv(inputs=conv3_1, scope_name="conv3_2", kernel_size=3, output_size=256, stride=1, init_bias=0.0, padding="SAME") # conv3_3 conv3_3 = conv(inputs=conv3_2, scope_name="conv3_3", kernel_size=3, output_size=256, stride=1, init_bias=0.0, padding="SAME") pool3 = tf.nn.max_pool(conv3_3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool3') with tf.variable_scope("conv4"): # conv4_1 conv4_1 = conv(inputs=pool3, scope_name="conv4_1", kernel_size=3, output_size=512, stride=1, init_bias=0.0, padding="SAME") # conv4_2 conv4_2 = conv(inputs=conv4_1, scope_name="conv4_2", kernel_size=3, output_size=512, stride=1, init_bias=0.0, padding="SAME") # conv4_3 conv4_3 = conv(inputs=conv4_2, scope_name="conv4_3", kernel_size=3, output_size=512, stride=1, init_bias=0.0, padding="SAME") pool4 = tf.nn.max_pool(conv4_3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool4') with tf.variable_scope("conv5"): # conv5_1 conv5_1 = conv(inputs=pool4, scope_name="conv4_1", kernel_size=3, output_size=512, stride=1, init_bias=0.0, padding="SAME") # conv5_2 conv5_2 = conv(inputs=conv5_1, scope_name="conv4_2", kernel_size=3, output_size=512, stride=1, init_bias=0.0, padding="SAME") # conv5_3 conv5_3 = conv(inputs=conv5_2, scope_name="conv4_3", kernel_size=3, output_size=512, stride=1, init_bias=0.0, padding="SAME") pool5 = tf.nn.max_pool(conv5_3, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME', name='pool4') input_shape = pool5.get_shape().as_list() # 后面做全连接,所以要把shape改成2维 # shape=[batch, dim] flatten = tf.reshape(pool5, [-1, input_shape[1] * input_shape[2] * input_shape[3]]) fc1 = fc(inputs=flatten, scope_name="fc1", output_size=4096, init_bias=1.0, activeation_func=True) fc2 = fc(inputs=fc1, scope_name="fc2", output_size=4096, init_bias=1.0, activeation_func=True) fc3 = fc(inputs=fc2, scope_name="fc3", output_size=class_num, init_bias=1.0, activeation_func=True) return fc3
上图中有一个softmax层,我们也可以定义出来
class_num = 1000 # placeholder 定义 inputs = tf.placeholder(dtype="float", shape=[None, 28, 28, 3], name='inputs') labels = tf.placeholder(dtype="float", shape=[None, class_num], name='labels') learning_rate = tf.placeholder("float", None, name='learning_rate') logits = VGG16Net(inputs) probs = tf.nn.softmax(logits)
ResNet模型
论文:
- 2016_ResNet_Deep Residual Learning for Image Recognition
- 2016_ResNet_Identity Mappings in Deep Residual Networks
ResNet的网络结构如下图所示
我们先定义需要用到的神经网络单元
def batch_normalization(inputs, is_training, epsilon=0.001, decay=0.9): # 计算公式为: # inputs = (inputs-mean)/tf.sqrt(variance+epilon) # inputs = inputs * gamma + beta input_size = inputs.get_shape().as_list()[-1] # 扩大参数 gamma = tf.get_variable('gamma', input_size, tf.float32, initializer=tf.ones_initializer) # 也叫scale # 平移参数 beta = tf.get_variable('beta', input_size, tf.float32, initializer=tf.zeros_initializer) # 也叫shift # 移动均值 moving_mean = tf.get_variable('moving_mean', input_size, tf.float32, initializer=tf.zeros_initializer, trainable=False) # 移动方差 moving_variance = tf.get_variable('moving_variance', input_size, tf.float32, initializer=tf.ones_initializer, trainable=False) def mean_and_var_update(): # 这些op只有在训练时才能进行 # 因为image是4维数据, 我们需要对[batch, height, width]求取均值和方差,[0, 1, 2] axes = list(range(len(inputs.get_shape()) - 1)) # [0, 1, 2] mean, variance = tf.nn.moments(inputs, axes=axes, name="moments") # 计算均值和方差 # 用滑动平均值来统计整体的均值和方差,在训练阶段并用不上,在测试阶段才会用,这里是保证在训练阶段计算了滑动平均值 update_moving_mean = moving_averages.assign_moving_average(moving_mean, mean, decay=decay) # 应用滑动平均 操作 # 也可以用:moving_average_mean = tf.assign(pop_mean, pop_mean * decay + batch_mean * (1 - decay)) update_moving_variance = moving_averages.assign_moving_average(moving_variance, variance, decay=decay) # 也可以用:moving_average_var = tf.assign(pop_var, pop_var * decay + batch_var * (1 - decay)) with tf.control_dependencies([update_moving_mean, update_moving_variance]): return tf.identity(mean), tf.identity(variance) is_training = tf.cast(is_training, tf.bool) mean, variance = tf.cond(is_training, mean_and_var_update, lambda: (moving_mean, moving_variance)) bn_layer = tf.nn.batch_normalization(inputs, mean=mean, variance=variance, offset=beta, scale=gamma, variance_epsilon=epsilon) return bn_layer
def conv(input, kernel_size, output_size, stride, padding="SAME", wd=None): input_size = input.shape[-1] conv_weights = tf.get_variable(name='weights', shape=[kernel_size, kernel_size, input_size, output_size], dtype=tf.float32, initializer=tf.truncated_normal_initializer(mean=0.0, stddev=0.1), regularizer=tf.contrib.layers.l2_regularizer(0.00004)) # 正则损失衰减率0.000004 conv_layer = tf.nn.conv2d(input, conv_weights, [1, stride, stride, 1], padding=padding) # 卷积操作 batch_norm = batch_normalization(conv_layer, output_size) conv_output = tf.nn.relu(batch_norm) # relu激活函数 return conv_output
def fc(input, output_size, activeation_func=True): input_shape = input.shape[-1] # 创建 全连接权重 变量 fc_weights = tf.get_variable(name="weights", shape=[input_shape, output_size], initializer=tf.truncated_normal_initializer(stddev=0.01), dtype=tf.float32, regularizer=tf.contrib.layers.l2_regularizer(0.01)) # 创建 全连接偏置 变量 fc_biases = tf.get_variable(name="biases", shape=[output_size], initializer=tf.zeros_initializer, dtype=tf.float32) fc_layer = tf.matmul(input, fc_weights) # 全连接计算 fc_layer = tf.nn.bias_add(fc_layer, fc_biases) # 加上偏置项 if activeation_func: fc_layer = tf.nn.relu(fc_layer) # rele激活函数 return fc_layer
def block(input, n, output_size, change_first_stride, bottleneck): if n == 0 and change_first_stride: stride = 2 else: stride = 1 if bottleneck: with tf.variable_scope('a'): conv_a = conv(input=input, kernel_size=1, output_size=output_size, stride=stride, padding="SAME") conv_a = batch_normalization(conv_a, output_size) conv_a = tf.nn.relu(conv_a) with tf.variable_scope('b'): conv_b = conv(input=conv_a, kernel_size=3, output_size=output_size, stride=1, padding="SAME") conv_b = batch_normalization(conv_b, output_size) conv_b = tf.nn.relu(conv_b) with tf.variable_scope('c'): conv_c = conv(input=conv_b, kernel_size=1, output_size=output_size * 4, stride=1, padding="SAME") output = batch_normalization(conv_c, output_size * 4) else: with tf.variable_scope('A'): conv_A = conv(input=input, kernel_size=3, output_size=output_size, stride=stride, padding="SAME") conv_A = batch_normalization(conv_A, output_size) conv_A = tf.nn.relu(conv_A) with tf.variable_scope('B'): conv_B = conv(input=conv_A, kernel_size=3, output_size=output_size, stride=1, padding="SAME") output = batch_normalization(conv_B, output_size) if input.shape == output.shape: with tf.variable_scope('shortcut'): shortcut = input # shortcut else: with tf.variable_scope('shortcut'): shortcut = conv(input=input, kernel_size=1, output_size=output_size * 4, stride=1, padding="SAME") shortcut = batch_normalization(shortcut, output_size * 4) return tf.nn.relu(output + shortcut)
然后我们定义神经网络框架
def inference(inputs, class_num, num_blocks=[3, 4, 6, 3], bottleneck=True): # data[1, 224, 224, 3] # 我们尝试搭建50层ResNet with tf.variable_scope('conv1'): conv1 = conv(input=inputs, kernel_size=7, output_size=64, stride=2, padding="SAME") conv1 = batch_normalization(inputs=conv1, output_size=64) conv1 = tf.nn.relu(conv1) with tf.variable_scope('conv2_x'): conv_output = tf.nn.max_pool(conv1, ksize=[1, 3, 3, 1], strides=[1, 2, 2, 1], padding='SAME') for n in range(num_blocks[0]): with tf.variable_scope('block%d' % (n + 1)): conv_output = block(conv_output, n, output_size=64, change_first_stride=False, bottleneck=bottleneck) with tf.variable_scope('conv3_x'): for n in range(num_blocks[1]): with tf.variable_scope('block%d' % (n + 1)): conv_output = block(conv_output, n, output_size=128, change_first_stride=True, bottleneck=bottleneck) with tf.variable_scope('conv4_x'): for n in range(num_blocks[2]): with tf.variable_scope('block%d' % (n + 1)): conv_output = block(conv_output, n, output_size=256, change_first_stride=True, bottleneck=bottleneck) with tf.variable_scope('conv5_x'): for n in range(num_blocks[3]): with tf.variable_scope('block%d' % (n + 1)): conv_output = block(conv_output, n, output_size=512, change_first_stride=True, bottleneck=bottleneck) output = tf.reduce_mean(conv_output, reduction_indices=[1, 2], name="avg_pool") with tf.variable_scope('fc'): output = fc(output, class_num, activeation_func=False) return output
Google Inception Net模型
Inception Net模型 以后再更新吧,如果这篇文章对大家有用,欢迎大家催促我。
RNN模型
Tensorflow中的CNN变数很少,而RNN却丰富多彩,不仅在RNN Cell上有很多种、在实现上也有很多种,在用法上更是花样百出。
五个基本的RNN Cell:RNNCell、BasicRNNCell、LSTMCell、BasicLSTMCell、GRUCell
RNN Cell的封装和变形:MultiRNNCell(多层RNN)、DropoutWrapper、ResidualWrapper、DeviceWrapper
四种架构 (static+dynamic)*(单向+双向)=4:static_rnn(静态RNN)、dynamic_rnn(动态RNN)、static_bidirectional_rnn(静态双向RNN)、bidirectional_dynamic_rnn(动态双向RNN)
五种手法 (one+many)*(one+many) +1=5:
- one to one(1 vs 1):输入一个,输出一个。其实和全连接神经网络并没有什么区别,这一类别算不得是 RNN。
- one to many(1 vs N):输入一个,输出多个。图像标注,输入一个图片,得到对图片的语言描述
- many to one(N vs 1):输入多个,输出一个。序列分类,把序列压缩成一个向量
- many to many(N vs N):输入多个,输出多个。两者长度可以不一样。翻译任务
- many to many(N vs N):输入多个,输出多个。两者长度一样。char RNN
我们先定义需要用到的神经网络单元
全连接层
def fc(input, output_size, activeation_func=tf.nn.relu): input_shape = input.shape[-1] # 创建 全连接权重 变量 fc_weights = tf.get_variable(name="weights", shape=[input_shape, output_size], initializer=tf.truncated_normal_initializer(stddev=0.01), dtype=tf.float32, regularizer=tf.contrib.layers.l2_regularizer(0.01)) # 创建 全连接偏置 变量 fc_biases = tf.get_variable(name="biases", shape=[output_size], initializer=tf.zeros_initializer, dtype=tf.float32) fc_layer = tf.matmul(input, fc_weights) # 全连接计算 fc_layer = tf.nn.bias_add(fc_layer, fc_biases) # 加上偏置项 if activeation_func: fc_layer = activeation_func(fc_layer) # rele激活函数 return fc_layer
单层 静态/动态 LSTM/GRU
####################################### # 单层 静态/动态 LSTM/GRU # ####################################### # 单层静态LSTM def single_layer_static_lstm(input_x, time_steps, hidden_size): """ :param input_x: 输入张量 形状为[batch_size, n_steps, input_size] :param n_steps: 时序总数 :param n_hidden: LSTM单元输出的节点个数 即隐藏层节点数 """ # 把输入input_x按列拆分,并返回一个有n_steps个张量组成的list # 如batch_sizex28x28的输入拆成[(batch_size,28),((batch_size,28))....] # 如果是调用的是静态rnn函数,需要这一步处理 即相当于把序列作为第一维度 input_x1 = tf.unstack(input_x, num=time_steps, axis=1) lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(num_units=hidden_size) # 创建LSTM_cell # 静态rnn函数传入的是一个张量list 每一个元素都是一个(batch_size,input_size)大小的张量 output, states = tf.nn.static_rnn(cell=lstm_cell, inputs=input_x1, dtype=tf.float32) # 通过cell类构建RNN return output, states # 单层静态gru def single_layer_static_gru(input, time_steps, hidden_size): """ :param input: 输入张量 形状为[batch_size, n_steps, input_size] :param n_steps: 时序总数 :param n_hidden: gru单元输出的节点个数 即隐藏层节点数 :return: 返回静态单层GRU单元的输出,以及cell状态 """ # 把输入input_x按列拆分,并返回一个有n_steps个张量组成的list # 如batch_sizex28x28的输入拆成[(batch_size,28),((batch_size,28))....] # 如果是调用的是静态rnn函数,需要这一步处理 即相当于把序列作为第一维度 input_x = tf.unstack(input, num=time_steps, axis=1) gru_cell = tf.nn.rnn_cell.GRUCell(num_units=hidden_size) # 创建GRU_cell # 静态rnn函数传入的是一个张量list 每一个元素都是一个(batch_size,input_size)大小的张量 output, states = tf.nn.static_rnn(cell=gru_cell, inputs=input_x, dtype=tf.float32) # 通过cell类构建RNN return output, states # 单层动态LSTM def single_layer_dynamic_lstm(input, time_steps, hidden_size): """ :param input_x: 输入张量 形状为[batch_size, time_steps, input_size] :param time_steps: 时序总数 :param hidden_size: LSTM单元输出的节点个数 即隐藏层节点数 :return: 返回动态单层LSTM单元的输出,以及cell状态 """ lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(num_units=hidden_size) # 创建LSTM_cell # 动态rnn函数传入的是一个三维张量,[batch_size,time_steps, input_size] 输出也是这种形状 output, states = tf.nn.dynamic_rnn(cell=lstm_cell, inputs=input, dtype=tf.float32) # 通过cell类构建RNN output = tf.transpose(output, [1, 0, 2]) # 注意这里输出需要转置 转换为时序优先的 return output, states # 单层动态gru def single_layer_dynamic_gru(input, time_steps, hidden_size): """ :param input: 输入张量 形状为[batch_size, time_steps, input_size] :param time_steps: 时序总数 :param hidden_size: GRU单元输出的节点个数 即隐藏层节点数 :return: 返回动态单层GRU单元的输出,以及cell状态 """ gru_cell = tf.nn.rnn_cell.GRUCell(num_units=hidden_size) # 创建GRU_cell # 动态rnn函数传入的是一个三维张量,[batch_size,n_steps,input_size] 输出也是这种形状 output, states = tf.nn.dynamic_rnn(cell=gru_cell, inputs=input, dtype=tf.float32) # 通过cell类构建RNN output = tf.transpose(output, [1, 0, 2]) # 注意这里输出需要转置 转换为时序优先的 return output, states
多层 静态/动态 LSTM/GRU
####################################### # 多层 静态/动态 LSTM/GRU # ####################################### # 多层静态LSTM网络 def multi_layer_static_lstm(input, time_steps, hidden_size): """ :param input: 输入张量 形状为[batch_size,time_steps,input_size] :param time_steps: 时序总数 :param n_hidden: LSTM单元输出的节点个数 即隐藏层节点数 :return: 返回静态多层LSTM单元的输出,以及cell状态 """ # 把输入input_x按列拆分,并返回一个有n_steps个张量组成的list # 如batch_sizex28x28的输入拆成[(batch_size,28),((batch_size,28))....] # 如果是调用的是静态rnn函数,需要这一步处理 即相当于把序列作为第一维度 input_x1 = tf.unstack(input, num=time_steps, axis=1) # 多层RNN的实现 例如cells=[cell1,cell2,cell3],则表示一共有三层 mcell = tf.nn.rnn_cell.MultiRNNCell( [tf.nn.rnn_cell.LSTMCell(num_units=hidden_size) for _ in range(3)]) # 静态rnn函数传入的是一个张量list 每一个元素都是一个(batch_size,input_size)大小的张量 output, states = tf.nn.static_rnn(cell=mcell, inputs=input_x1, dtype=tf.float32) return output, states # 多层静态GRU def multi_layer_static_gru(input, time_steps, hidden_size): """ :param input_x: 输入张量 形状为[batch_size,n_steps,input_size] :param time_steps: 时序总数 :param hidden_size: gru单元输出的节点个数 即隐藏层节点数 :return: 返回静态多层GRU单元的输出,以及cell状态 """ # 把输入input_x按列拆分,并返回一个有n_steps个张量组成的list # 如batch_sizex28x28的输入拆成[(batch_size,28),((batch_size,28))....] # 如果是调用的是静态rnn函数,需要这一步处理 即相当于把序列作为第一维度 input_x = tf.unstack(input, num=time_steps, axis=1) # 多层RNN的实现 例如cells=[cell1,cell2,cell3],则表示一共有三层 mcell = tf.nn.rnn_cell.MultiRNNCell( [tf.nn.rnn_cell.GRUCell(num_units=hidden_size) for _ in range(3)]) # 静态rnn函数传入的是一个张量list 每一个元素都是一个(batch_size,input_size)大小的张量 output, states = tf.nn.static_rnn(cell=mcell, inputs=input_x, dtype=tf.float32) return output, states # 多层静态GRU和LSTM 混合 def multi_layer_static_mix(input, time_steps, hidden_size): """ :param input: 输入张量 形状为[batch_size,n_steps,input_size] :param time_steps: 时序总数 :param hidden_size: gru单元输出的节点个数 即隐藏层节点数 :return: 返回静态多层GRU和LSTM混合单元的输出,以及cell状态 """ # 把输入input_x按列拆分,并返回一个有n_steps个张量组成的list # 如batch_sizex28x28的输入拆成[(batch_size,28),((batch_size,28))....] # 如果是调用的是静态rnn函数,需要这一步处理 即相当于把序列作为第一维度 input_x = tf.unstack(input, num=time_steps, axis=1) # 可以看做2个隐藏层 lstm_cell = tf.nn.rnn_cell.LSTMCell(num_units=hidden_size) gru_cell = tf.nn.rnn_cell.GRUCell(num_units=hidden_size * 2) # 多层RNN的实现 例如cells=[cell1,cell2],则表示一共有两层,数据经过cell1后还要经过cells mcell = tf.nn.rnn_cell.MultiRNNCell(cells=[lstm_cell, gru_cell]) # 静态rnn函数传入的是一个张量list 每一个元素都是一个(batch_size,input_size)大小的张量 output, states = tf.nn.static_rnn(cell=mcell, inputs=input_x, dtype=tf.float32) return output, states # 多层动态LSTM def multi_layer_dynamic_lstm(input, time_steps, hidden_size): """ :param input: 输入张量 形状为[batch_size,n_steps,input_size] :param time_steps: 时序总数 :param hidden_size: LSTM单元输出的节点个数 即隐藏层节点数 :return: 返回动态多层LSTM单元的输出,以及cell状态 """ # 多层RNN的实现 例如cells=[cell1,cell2],则表示一共有两层,数据经过cell1后还要经过cells mcell = tf.nn.rnn_cell.MultiRNNCell( [tf.nn.rnn_cell.LSTMCell(num_units=hidden_size) for _ in range(3)]) # 动态rnn函数传入的是一个三维张量,[batch_size,n_steps,input_size] 输出也是这种形状 output, states = tf.nn.dynamic_rnn(cell=mcell, inputs=input, dtype=tf.float32) # 注意这里输出需要转置 转换为时序优先的 output = tf.transpose(output, [1, 0, 2]) return output, states # 多层动态GRU def multi_layer_dynamic_gru(input, time_steps, hidden_size): """ :param input: 输入张量 形状为[batch_size,n_steps,input_size] :param time_steps: 时序总数 :param hidden_size: gru单元输出的节点个数 即隐藏层节点数 :return: 返回动态多层GRU单元的输出,以及cell状态 """ # 多层RNN的实现 例如cells=[cell1,cell2],则表示一共有两层,数据经过cell1后还要经过cells mcell = tf.nn.rnn_cell.MultiRNNCell( [tf.nn.rnn_cell.GRUCell(num_units=hidden_size) for _ in range(3)]) # 动态rnn函数传入的是一个三维张量,[batch_size,n_steps,input_size] 输出也是这种形状 output, states = tf.nn.dynamic_rnn(cell=mcell, inputs=input, dtype=tf.float32) # 注意这里输出需要转置 转换为时序优先的 output = tf.transpose(output, [1, 0, 2]) return output, states # 多层动态GRU和LSTM 混合 def multi_layer_dynamic_mix(input, time_steps, hidden_size): """ :param input: 输入张量 形状为[batch_size,n_steps,input_size] :param time_steps: 时序总数 :param hidden_size: gru单元输出的节点个数 即隐藏层节点数 :return: 返回动态多层GRU和LSTM混合单元的输出,以及cell状态 """ # 可以看做2个隐藏层 gru_cell = tf.nn.rnn_cell.GRUCell(num_units=hidden_size * 2) lstm_cell = tf.nn.rnn_cell.LSTMCell(num_units=hidden_size) # 多层RNN的实现 例如cells=[cell1,cell2],则表示一共有两层,数据经过cell1后还要经过cells mcell = tf.nn.rnn_cell.MultiRNNCell(cells=[lstm_cell, gru_cell]) # 动态rnn函数传入的是一个三维张量,[batch_size,n_steps,input_size] 输出也是这种形状 output, states = tf.nn.dynamic_rnn(cell=mcell, inputs=input, dtype=tf.float32) # 注意这里输出需要转置 转换为时序优先的 output = tf.transpose(output, [1, 0, 2]) return output, states
单层/多层 双向 静态/动态 LSTM/GRU
####################################### # 单层/多层 双向 静态/动态 LSTM/GRU # ####################################### # 单层静态双向LSTM def single_layer_static_bi_lstm(input, time_steps, hidden_size): """ :param input: 输入张量 形状为[batch_size,time_steps,input_size] :param time_steps: 时序总数 :param hidden_size: LSTM单元输出的节点个数 即隐藏层节点数 :return: 返回单层静态双向LSTM单元的输出,以及cell状态 """ # 把输入input_x按列拆分,并返回一个有n_steps个张量组成的list # 如batch_sizex28x28的输入拆成[(batch_size,28),((batch_size,28))....] # 如果是调用的是静态rnn函数,需要这一步处理 即相当于把序列作为第一维度 input_x = tf.unstack(input, num=time_steps, axis=1) lstm_fw_cell = tf.nn.rnn_cell.BasicLSTMCell(num_units=hidden_size) # 正向 lstm_bw_cell = tf.nn.rnn_cell.BasicLSTMCell(num_units=hidden_size) # 反向 # 静态rnn函数传入的是一个张量list 每一个元素都是一个(batch_size,input_size)大小的张量 # 这里的输出output是一个list 每一个元素都是前向输出,后向输出的合并 output, fw_state, bw_state = tf.nn.static_bidirectional_rnn(cell_fw=lstm_fw_cell, cell_bw=lstm_bw_cell, inputs=input_x, dtype=tf.float32) print(type(output)) # <class 'list'> print(len(output)) # 28 print(output[0].shape) # (?, 256) return output, fw_state, bw_state # 单层动态双向LSTM def single_layer_dynamic_bi_lstm(input, time_steps, hidden_size): """ :param input: 输入张量 形状为[batch_size,time_steps,input_size] :param time_steps: 时序总数 :param hidden_size: gru单元输出的节点个数 即隐藏层节点数 :return: 返回单层动态双向LSTM单元的输出,以及cell状态 """ lstm_fw_cell = tf.nn.rnn_cell.BasicLSTMCell(num_units=hidden_size) # 正向 lstm_bw_cell = tf.nn.rnn_cell.BasicLSTMCell(num_units=hidden_size) # 反向 # 动态rnn函数传入的是一个三维张量,[batch_size,time_steps,input_size] 输出是一个元组 每一个元素也是这种形状 output, state = tf.nn.bidirectional_dynamic_rnn(cell_fw=lstm_fw_cell, cell_bw=lstm_bw_cell, inputs=input, dtype=tf.float32) print(type(output)) # <class 'tuple'> print(len(output)) # 2 print(output[0].shape) # (?, 28, 128) print(output[1].shape) # (?, 28, 128) output = tf.concat(output, axis=2) # 按axis=2合并 (?,28,128) (?,28,128)按最后一维合并(?,28,256) output = tf.transpose(output, [1, 0, 2]) # 注意这里输出需要转置 转换为时序优先的 return output, state # 多层静态双向LSTM def multi_layer_static_bi_lstm(input, time_steps, hidden_size): """ :param input: 输入张量 形状为[batch_size,time_steps,input_size] :param time_steps: 时序总数 :param hidden_size: LSTM单元输出的节点个数 即隐藏层节点数 :return: 返回多层静态双向LSTM单元的输出,以及cell状态 """ # 把输入input_x按列拆分,并返回一个有n_steps个张量组成的list # 如batch_sizex28x28的输入拆成[(batch_size,28),((batch_size,28))....] # 如果是调用的是静态rnn函数,需要这一步处理 即相当于把序列作为第一维度 input_x = tf.unstack(input, num=time_steps, axis=1) stacked_fw_rnn = [] stacked_bw_rnn = [] for i in range(3): stacked_fw_rnn.append(tf.nn.rnn_cell.BasicLSTMCell(num_units=hidden_size)) # 正向 stacked_bw_rnn.append(tf.nn.rnn_cell.BasicLSTMCell(num_units=hidden_size)) # 反向 # 静态rnn函数传入的是一个张量list 每一个元素都是一个(batch_size,input_size)大小的张量 # 这里的输出output是一个list 每一个元素都是前向输出,后向输出的合并 output, fw_state, bw_state = tf.contrib.rnn.stack_bidirectional_rnn(stacked_fw_rnn, stacked_bw_rnn, inputs=input_x, dtype=tf.float32) print(type(output)) # <class 'list'> print(len(output)) # 28 print(output[0].shape) # (?, 256) return output, fw_state, bw_state # 多层动态双向LSTM def multi_layer_dynamic_bi_lstm(input, time_steps, hidden_size): """ :param input: 输入张量 形状为[batch_size,n_steps,input_size] :param time_steps: 时序总数 :param hidden_size: gru单元输出的节点个数 即隐藏层节点数 :return: 返回多层动态双向LSTM单元的输出,以及cell状态 """ stacked_fw_rnn = [] stacked_bw_rnn = [] for i in range(3): stacked_fw_rnn.append(tf.nn.rnn_cell.BasicLSTMCell(num_units=hidden_size)) # 正向 stacked_bw_rnn.append(tf.nn.rnn_cell.BasicLSTMCell(num_units=hidden_size)) # 反向 # 动态rnn函数传入的是一个三维张量,[batch_size,n_steps,input_size] 输出也是这种形状, # input_size变成了正向和反向合并之后的 即input_size*2 output, fw_state, bw_state = tf.contrib.rnn.stack_bidirectional_dynamic_rnn(stacked_fw_rnn, stacked_bw_rnn, inputs=input, dtype=tf.float32) print(type(output)) # <class 'tensorflow.python.framework.ops.Tensor'> print(output.shape) # (?, 28, 256) output = tf.transpose(output, [1, 0, 2]) # 注意这里输出需要转置 转换为时序优先的 return output, fw_state, bw_state
然后我们定义神经网络框架
def RNN_inference(inputs, class_num, time_steps, hidden_size): """ :param inputs: [batch_size, n_steps, input_size] :param class_num: 类别数 :param time_steps: 时序总数 :param n_hidden: LSTM单元输出的节点个数 即隐藏层节点数 """ ####################################### # 单层 静态/动态 LSTM/GRU # ####################################### # outputs, states = single_layer_static_lstm(inputs, time_steps, hidden_size) # 单层静态LSTM # outputs, states = single_layer_static_gru(inputs, time_steps, hidden_size) # 单层静态gru # outputs, states = single_layer_dynamic_lstm(inputs, time_steps, hidden_size) # 单层动态LSTM # outputs, states = single_layer_dynamic_gru(inputs, time_steps, hidden_size) # 单层动态gru ####################################### # 多层 静态/动态 LSTM/GRU # ####################################### # outputs, states = multi_layer_static_lstm(inputs, time_steps, hidden_size) # 多层静态LSTM网络 # outputs, states = multi_layer_static_gru(inputs, time_steps, hidden_size) # 多层静态GRU # outputs, states = multi_layer_static_mix(inputs, time_steps, hidden_size) # 多层静态GRU和LSTM 混合 # outputs, states = multi_layer_dynamic_lstm(inputs, time_steps, hidden_size) # 多层动态LSTM # outputs, states = multi_layer_dynamic_gru(inputs, time_steps, hidden_size) # 多层动态GRU # outputs, states = multi_layer_dynamic_mix(inputs, time_steps, hidden_size) # 多层动态GRU和LSTM 混合 ####################################### # 单层/多层 双向 静态/动态 LSTM/GRU # ####################################### # outputs, fw_state, bw_state = single_layer_static_bi_lstm(inputs, time_steps, hidden_size) # 单层静态双向LSTM # outputs, state = single_layer_dynamic_bi_lstm(inputs, time_steps, hidden_size) # 单层动态双向LSTM # outputs, fw_state, bw_state = multi_layer_static_bi_lstm(inputs, time_steps, hidden_size) # 多层静态双向LSTM outputs, fw_state, bw_state = multi_layer_dynamic_bi_lstm(inputs, time_steps, hidden_size) # 多层动态双向LSTM # output静态是 time_step=28个(batch=128, output=128)组成的列表 # output动态是 (time_step=28, batch=128, output=128) print('hidden:', outputs[-1].shape) # 最后一个时序的shape(128,128) # 取LSTM最后一个时序的输出,然后经过全连接网络得到输出值 fc_output = fc(input=outputs[-1], output_size=class_num, activeation_func=tf.nn.relu) return fc_output
设置全局变量和超参数
在模型训练之前我们首先会定义一些超参数:batch_size、batch_nums、class_num、epochs、learning_rate
batch_size = FLAGS.batch_size batch_nums = mnist.train.images.shape[0] // batch_size # 一个epoch中应该包含多少batch数据 class_num = FLAGS.class_num epochs = FLAGS.epochs learning_rate = FLAGS.learning_rate
保存检查点的地址
############ 保存检查点的地址 ############ checkpoints_dir = FLAGS.checkpoints_dir # checkpoints # 如果检查点不存在,则创建 if not os.path.exists(checkpoints_dir): os.makedirs(FLAGS.checkpoints_dir)
创建图
这一步可以不设置,因为tensorflow有一个默认图,我们定义的操作都是在默认图上的,当然我们也可以定义自己的,方便管理。
###################################################### # 创建图 # ###################################################### graph = tf.Graph() # 自定义图 # 在自己的图中定义数据和操作 with graph.as_default():
占位符
一般我们会把input和label做成placeholder,方便我们使用把不同的batch数据传入网络,一些其他的超参数也可以做成placeholder,比如learning_rate、dorpout_keep_prob。一般在搭建模型的时候把placeholder的变量传入模型,在训练模型sess.run(train_op, feed_dict)的时候通过参数feed_dict={input:真实数据,label:真实标签} 把真实的数据传入神经网络。
inputs = tf.placeholder(dtype="float", shape=[None, 28, 28, 1], name='inputs') labels = tf.placeholder(dtype="float", shape=[None, class_num], name='labels') # 看个人喜欢,有的人在初始化定义中就定义了learning_rate,有的人喜欢通过feed传learning_rate learning_rate = tf.placeholder("float", None, name='learning_rate') # 如果网络结构有dropout层,需要定义keep_probn,如果没有则不需要 # 训练的时候需要,测试的时候需要设置成1 keep_prob = tf.placeholder(dtype="float", name='keep_prob')
搭建模型
传进入的都是placeholder数据,不是我们之前整理好的batch数据。
############ 搭建模型 ############ logits = alexNet(inputs, class_num, keep_prob=keep_prob) # 使用placeholder搭建模型
构建损失
分类任务一般输出的是每个类别的概率向量,因此模型输出最后都要经过softmax转换成概率。一般经过softmax的输出损失函数都是交叉熵损失函数,tensorflow有将以上两步合在一起的现成函数 tf.nn.softmax_cross_entropy_with_logits
############ 损失函数 ############ loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=labels, logits=logits)) tf.add_to_collection('losses', loss) total_loss = tf.add_n(tf.get_collection("loss")) # total_loss=模型损失+权重正则化损失
自定义损失
以后更新,欢迎大家催我。
模型精度
在测试数据集上的精度
############ 模型精度 ############ predict = tf.argmax(logits, 1) # 模型预测结果 accuracy = tf.reduce_mean(tf.cast(tf.equal(predict, tf.argmax(labels, 1)), tf.float32))
自定义度量
以后更新,欢迎大家催我。
优化器
创建优化器,更新网络参数,最小化loss
优化器的种类有很多种,但是用法都差不多,常用的优化器有:
- tf.train.AdamOptimizer
-
tf.train.GradientDescentOptimizer
- tf.train.RMSPropOptimizer
下面以Adam优化器为例
############ 优化器 ############ variable_to_train = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES) # 可训练变量列表 global_step = tf.Variable(0, trainable=False) # 训练step # 设置学习率衰减 learning_rate = tf.train.exponential_decay(learning_rate=learning_rate, # 初始学习率 global_step=global_step, decay_steps=batch_nums, # 多少步衰减一次 decay_rate=0.1, # 衰减率 staircase=True) # 以阶梯的形式衰减 # 创建Adam优化器,更新模型参数,最小化损失函数 train_op = tf.train.AdamOptimizer(learning_rate).minimize(loss=total_loss, # 损失函数 global_step=global_step, var_list=variable_to_train) # 通过训练需要更新的参数列表
讲解:
- variable_to_train:上面的代码定义了可训练变量,我只是把列出了模型默认的可训练变量,这一个步是tensorflow默认的,如果不设置也没有关系。我写出来的原因是,有的大牛会这么写,对不同的可训练变量分别进行不同的优化,希望大家看到我的代码,下次看到别人的不会觉得陌生。
- global_step:大多数人会用step=0,然后在训练的时候step+=1的方式更新step,但是本文介绍的是另一种方式,以tf.Variable的方式定义step,在模型训练的时候传入sess.run,global_step会自动+1更新
- learning_rate:本文还设置了学习率衰减,大家也可以不设置,以固定的学习率训练模型,但是对于大型项目,还是推荐设置。
移动平均值更新参数
采用移动平均值的方式更新损失值和模型参数
def train(total_loss, global_step): lr = tf.train.exponential_decay(0.01, global_step, decay_steps=350, decay_rate=0.1, staircase=True) # 采用滑动平均的方法更新损失值 loss_averages = tf.train.ExponentialMovingAverage(decay=0.9, name='avg') losses = tf.get_collection('losses') # losses的列表 loss_averages_op = loss_averages.apply(losses + [total_loss]) # 计算损失值的影子变量op # 计算梯度 with tf.control_dependencies([loss_averages_op]): # 控制计算指定,只有执行了括号中的语句才能执行下面的语句 opt = tf.train.GradientDescentOptimizer(lr) # 创建优化器 grads = opt.compute_gradients(total_loss) # 计算梯度 # 应用梯度 apply_gradient_op = opt.apply_gradients(grads, global_step=global_step) # 采用滑动平均的方法更新参数 variable_averages = tf.train.ExponentialMovingAverage(0.999, num_updates=global_step) variables_averages_op = variable_averages.apply(tf.trainable_variables()) with tf.control_dependencies([apply_gradient_op, variables_averages_op]): # tf.no_op()表示执行完apply_gradient_op, variable_averages_op操作之后什么都不做 train_op = tf.no_op(name='train') return train_op
TensorBoard可视化 summary
############ TensorBoard可视化 summary ############ summary_writer = tf.summary.FileWriter(logdir="./logs", graph=graph) # 创建事件文件 tf.summary.scalar(name="losses", tensor=total_loss) # 收集损失值变量 tf.summary.scalar(name="acc", tensor=accuracy) # 收集精度值变量 tf.summary.scalar(name='learning_rate', tensor=learning_rate) merged_summary_op = tf.summary.merge_all() # 将所有的summary合并为一个op
模型保存和恢复 Saver
saver = tf.train.Saver(max_to_keep=5) # 保存最新的5个检查点
创建会话
配置会话
在创建会话之前我们一般都要配置会话,比如使用GPU还是CPU,用多少GPU等等。
我们一般使用 tf.ConfigProto()配置Session运行参数&&GPU设备指定
config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=True) config.gpu_options.per_process_gpu_memory_fraction = 0.4 # 占用40%显存 sess = tf.Session(config=config) # 或者 config = tf.ConfigProto() config.allow_soft_placement = True config.log_device_placement = True with tf.Session(config=config) as sess: # 或者 sess = tf.Session(config=config)
tf.ConfigProto(log_device_placement=True):记录设备指派情况
设置tf.ConfigProto()中参数log_device_placement = True,获取 operations 和 Tensor 被指派到哪个设备(几号CPU或几号GPU)上运行,会在终端打印出各项操作是在哪个设备上运行的。
tf.ConfigProto(allow_soft_placement=True):自动选择运行设备
在TensorFlow中,通过命令 "with tf.device('/cpu:0'):",允许手动设置操作运行的设备。如果手动设置的设备不存在或者不可用,就会导致tf程序等待或异常,为了防止这种情况,可以设置tf.ConfigProto()中参数allow_soft_placement=True,自动选择一个存在并且可用的设备来运行操作。
config.gpu_options.allow_growth = True
当使用GPU时候,Tensorflow运行自动慢慢达到最大GPU的内存
tf.test.is_built_with_cuda():返回是否能够使用GPU进行运算
为了加快运行效率,TensorFlow在初始化时会尝试分配所有可用的GPU显存资源给自己,这在多人使用的服务器上工作就会导致GPU占用,别人无法使用GPU工作的情况。这时我们需要限制GPU资源使用,详细实现方法请参考我的另一篇博客 tensorflow常用函数 Ctrl+F搜索“限制GPU资源使用”
创建会话Session
Session有两种创建方式:
sess = tf.Session(config=config, graph=graph) # 或通过with的方式创建Session with tf.Session(config=config, graph=graph) as sess:
如果我们之前自定义了graph,则在会话中也要配置graph,如果之前没有自定义graph,使用的是tensorflow默认graph,则在会话不用自己去定义,tensorflow会自动找到默认图。
在训练模型之前我们首先要设置一个高级一点的东西,那就是检查是否有之前保存好的模型,如果有着接着前面的继续训练,如果没有则从头开始训练模型。
恢复/重新训练
定义一个检查模型是否存在的函数,为了美观,可以把这个函数放在最上面,或者其他脚本中,通过import导入。
def load_model(sess, saver, checkpoint_dir): """加载模型,看看还能不能加一个功能,必须现在的检查检点是1000,但是我的train是100,要报错 还有就是读取之前的模型继续训练的问题 checkpoint_dir = checkpoint""" # 通过checkpoint找到模型文件名 ckpt = tf.train.get_checkpoint_state(checkpoint_dir=checkpoint_dir) if ckpt and ckpt.model_checkpoint_path: ckpt_name = os.path.basename(ckpt.model_checkpoint_path) # 返回最新的chechpoint文件名 model.ckpt-1000 print("新的chechpoint文件名", ckpt_name) # model.ckpt-2 saver.restore(sess, os.path.join(checkpoint_dir, ckpt_name)) # 现在不知道checkpoint文件名时怎样的,因此不知道里面如何运行 counter = int(next(re.finditer("(\d+)(?!.*\d)", ckpt_name)).group(0)) # 2 print(" [*] 成功模型 {}".format(ckpt_name)) return True, counter else: print(" [*] 找不到checkpoint") return False, 0
如果大家之前用的是global_step = tf.Variable(0, trainable=False),则使用下面diamante
# 加载模型,如果模型存在返回 是否加载成功和训练步数 could_load, checkpoint_step = load_model(sess, saver, "./log") if could_load: print(" [*] 加载成功") else: print(" [!] 加载失败") try: tf.global_variables_initializer().run() except: tf.initialize_all_variables().run()
如果大家想使用step=0,step+=1,则可以使用下面代码
# 加载模型,如果模型存在返回 是否加载成功和训练步数 could_load, checkpoint_step = load_model(sess, saver, FLAGS.checkpoints_dir) if could_load: step = checkpoint_step print(" [*] 模型加载成功") else: print(" [!] 模型加载失败") try: tf.global_variables_initializer().run() except: tf.initialize_all_variables().run() step = 0
开始训练
for epoch in range(epochs): for i in range(batch_nums): start_time = time.time() # batch_images = data_X[i * batch_size:(i + 1) * batch_size] # batch_labels = data_y[i * batch_size:(i + 1) * batch_size] train_batch_x, train_batch_y = mnist.train.next_batch(batch_size) # 使用真实数据填充placeholder,运行训练模型和合并变量操作 _, summary, loss, step = sess.run([train_op, merged_summary_op, total_loss, global_step], feed_dict={inputs: train_batch_x, labels: train_batch_y, keep_prob: 0.5}) if step % 100 == 0: summary_writer.add_summary(summary, step) # 将每次迭代后的变量写入事件文件 summary_writer.flush() # 强制summary_writer将缓存中的数据写入到日志文件中(可选) ############ 可视化打印 ############ print("Epoch:[%2d] [%4d/%4d] time:%4.4f,loss:%.8f" % ( epoch, i, batch_nums, time.time() - start_time, loss)) # 打印一些可视化的数据,损失... if step % 100 == 0: acc = sess.run(accuracy, feed_dict={inputs: mnist.validation.images, labels: mnist.validation.labels, keep_prob: 1.0}) print("Epoch:[%2d] [%4d/%4d] accuracy:%.8f" % (epoch, i, batch_nums, acc)) ############ 保存模型 ############ if acc > max_acc: max_acc = acc save_path = saver.save(sess, save_path=os.path.join(checkpoints_dir, "model.ckpt"), global_step=step) tf.logging.info("模型保存在: %s" % save_path) print("优化完成!")
模型评估
eval.py
模型评估的代码和模型训练的代码很像,只不过不需要对模型进行训练而已。
from ops import * import tensorflow as tf from nets.my_alex import alexNet from tensorflow.examples.tutorials.mnist import input_data tf.flags.DEFINE_integer('batch_size', 50, 'batch size, default: 1') tf.flags.DEFINE_integer('class_num', 10, 'batch size, default: 1') tf.flags.DEFINE_integer('epochs', 10, 'batch size, default: 1') tf.flags.DEFINE_string('checkpoints_dir', "checkpoints", '保存检查点的地址') FLAGS = tf.flags.FLAGS # 从MNIST_data/中读取MNIST数据。当数据不存在时,会自动执行下载 mnist = input_data.read_data_sets('./data', one_hot=True, reshape=False) # 将数组张换成图片形式 print(mnist.train.images.shape) # 训练数据图片(55000, 28, 28, 1) print(mnist.train.labels.shape) # 训练数据标签(55000, 10) print(mnist.test.images.shape) # 测试数据图片(10000, 28, 28, 1) print(mnist.test.labels.shape) # 测试数据图片(10000, 10) print(mnist.validation.images.shape) # 验证数据图片(5000, 28, 28, 1) print(mnist.validation.labels.shape) # 验证数据图片(5000, 10) def evaluate(): batch_size = FLAGS.batch_size batch_nums = mnist.train.images.shape[0] // batch_size # 一个epoch中应该包含多少batch数据 class_num = FLAGS.class_num test_batch_size = 5000 test_batch_num = mnist.test.images.shape[0] // test_batch_size ############ 保存检查点的地址 ############ checkpoints_dir = FLAGS.checkpoints_dir # checkpoints # 如果检查点不存在,则创建 if not os.path.exists(checkpoints_dir): print("模型文件不存在,无法进行评估") ###################################################### # 创建图 # ###################################################### graph = tf.Graph() # 自定义图 # 在自己的图中定义数据和操作 with graph.as_default(): inputs = tf.placeholder(dtype="float", shape=[None, 28, 28, 1], name='inputs') labels = tf.placeholder(dtype="float", shape=[None, class_num], name='labels') ############ 搭建模型 ############ logits = alexNet(inputs, FLAGS.class_num, keep_prob=1) # 使用placeholder搭建模型 ############ 模型精度 ############ predict = tf.argmax(logits, 1) accuracy = tf.reduce_mean(tf.cast(tf.equal(predict, tf.argmax(labels, 1)), tf.float32)) ############ 模型保存和恢复 Saver ############ saver = tf.train.Saver(max_to_keep=5) ###################################################### # 创建会话 # ###################################################### config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=True) with tf.Session(config=config, graph=graph) as sess: # 加载模型,如果模型存在返回 是否加载成功和训练步数 could_load, checkpoint_step = load_model(sess, saver, FLAGS.checkpoints_dir) if could_load: print(" [*] 加载成功") else: print(" [!] 加载失败") raise ValueError("模型文件不存在,无法进行评估") for i in range(test_batch_num): test_batch_x, test_batch_y = mnist.test.next_batch(test_batch_num) acc = sess.run(accuracy, feed_dict={inputs: test_batch_x, labels: test_batch_y}) print("模型精度为:", acc) one_image = mnist.test.images[1].reshape(1, 28, 28, 1) predict_label = sess.run(predict, feed_dict={inputs: one_image}) # print("123", tf.argmax(pre_yyy, 1).eval()) # [7] # print("123", tf.argmax(yyy, 1).eval()) # 7 def main(argv=None): evaluate() if __name__ == '__main__': tf.app.run()
参考文献
github搜索tensorflow AlexNet
github_finetune_alexnet_with_tensorflow
github_AlexNet_with_tensorflow