博客作者:凌逆戰
博客地址:https://www.cnblogs.com/LXP-Never/p/12774058.html
文章代碼:https://github.com/LXP-Never/blog_data/tree/master/tensorflow_model
我一直覺得TensorFlow的深度神經網絡代碼非常困難且繁瑣,對TensorFlow搭建模型也十分困惑,所以我近期閱讀了大量的神經網絡代碼,終於找到了搭建神經網絡的規律,各位要是覺得我的文章對你有幫助不妨點個贊,點個關注吧。
我個人把深度學習分為以下步驟:數據處理 --> 模型搭建 --> 構建損失 --> 模型訓練 --> 模型評估
我先把代碼放出來,然后一點一點來講

# Author:凌逆戰 # -*- encoding:utf-8 -*- # 修改時間: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, '初始學習率, 默認: 0.0002') tf.flags.DEFINE_string('checkpoints_dir', "checkpoints", '保存檢查點的地址') FLAGS = tf.flags.FLAGS # 從MNIST_data/中讀取MNIST數據。當數據不存在時,會自動執行下載 mnist = input_data.read_data_sets('./data', one_hot=True, reshape=False) # reshape=False (None, 28,28,1) # 用於第一層是卷積層 # reshape=False (None, 784) # 用於第一層是全連接層 # 我們看一下數據的shape print(mnist.train.images.shape) # 訓練數據圖片(55000, 28, 28, 1) print(mnist.train.labels.shape) # 訓練數據標簽(55000, 10) print(mnist.test.images.shape) # 測試數據圖片(10000, 28, 28, 1) print(mnist.test.labels.shape) # 測試數據圖片(10000, 10) print(mnist.validation.images.shape) # 驗證數據圖片(5000, 28, 28, 1) print(mnist.validation.labels.shape) # 驗證數據圖片(5000, 784) def train(): batch_size = FLAGS.batch_size # 一個batch訓練多少個樣本 batch_nums = mnist.train.images.shape[0] // batch_size # 一個epoch中應該包含多少batch數據 class_num = FLAGS.class_num # 分類類別數 epochs = FLAGS.epochs # 訓練周期數 learning_rate = FLAGS.learning_rate # 初始學習率 ############ 保存檢查點的地址 ############ checkpoints_dir = FLAGS.checkpoints_dir # checkpoints # 如果檢查點不存在,則創建 if not os.path.exists(checkpoints_dir): os.makedirs(FLAGS.checkpoints_dir) ###################################################### # 創建圖 # ###################################################### graph = tf.Graph() # 自定義圖 # 在自己的圖中定義數據和操作 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') # 看個人喜歡,有的人在初始化定義中就定義了learning_rate,有的人喜歡通過feed傳learning_rate learning_rate = tf.placeholder("float", None, name='learning_rate') # 如果網絡結構有dropout層,需要定義keep_probn,如果沒有則不需要 # 訓練的時候需要,測試的時候需要設置成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 ############ 損失函數 ############ 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=模型損失+權重正則化損失 ############ 模型精度 ############ predict = tf.argmax(logits, 1) # 模型預測結果 accuracy = tf.reduce_mean(tf.cast(tf.equal(predict, tf.argmax(labels, 1)), tf.float32)) ############ 優化器 ############ variable_to_train = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES) # 可訓練變量列表 # 創建優化器,更新網絡參數,最小化loss, global_step = tf.Variable(0, trainable=False) learning_rate = tf.train.exponential_decay(learning_rate=learning_rate, # 初始學習率 global_step=global_step, decay_steps=batch_nums, # 多少步衰減一次 decay_rate=0.1, # 衰減率 staircase=True) # 以階梯的形式衰減 # 移動平均值更新參數 # train_op = moving_average(loss, learning_rate, global_step) # adam優化器,adam算法好像會自動衰減學習率, train_op = tf.train.AdamOptimizer(learning_rate).minimize(loss=total_loss, global_step=global_step, var_list=variable_to_train) ############ TensorBoard可視化 summary ############ summary_writer = tf.summary.FileWriter(logdir="./logs", graph=graph) # 創建事件文件 tf.summary.scalar(name="losses", tensor=total_loss) # 收集損失值變量 tf.summary.scalar(name="acc", tensor=accuracy) # 收集精度值變量 tf.summary.scalar(name='learning_rate', tensor=learning_rate) merged_summary_op = tf.summary.merge_all() # 將所有的summary合並為一個op ############ 模型保存和恢復 Saver ############ saver = tf.train.Saver(max_to_keep=5) ###################################################### # 創建會話 # ###################################################### max_acc = 0. config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=True) with tf.Session(config=config, graph=graph) as sess: # 加載模型,如果模型存在返回 是否加載成功和訓練步數 could_load, checkpoint_step = load_model(sess, saver, FLAGS.checkpoints_dir) if could_load: print(" [*] 模型加載成功") else: print(" [!] 模型加載失敗") 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) # 使用真實數據填充placeholder,運行訓練模型和合並變量操作 _, 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) # 將每次迭代后的變量寫入事件文件 summary_writer.flush() # 強制summary_writer將緩存中的數據寫入到日志文件中(可選) ############ 可視化打印 ############ print("Epoch:[%2d] [%4d/%4d] time:%4.4f,loss:%.8f" % ( epoch, i, batch_nums, time.time() - start_time, loss)) # 打印一些可視化的數據,損失... 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("優化完成!") def main(argv=None): train() if __name__ == '__main__': # logging.basicConfig(level=logging.INFO) tf.logging.set_verbosity(tf.logging.INFO) tf.app.run()

# Author:凌逆戰 # -*- encoding:utf-8 -*- # 修改時間: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, '初始學習率, 默認: 0.0001') tf.flags.DEFINE_string('checkpoints_dir', "checkpoint", '保存檢查點的地址') FLAGS = tf.flags.FLAGS # 從MNIST_data/中讀取MNIST數據。當數據不存在時,會自動執行下載 mnist = input_data.read_data_sets('./MNIST_data', one_hot=True, reshape=False) # reshape=False (None, 28,28,1) # 用於第一層是卷積層 # reshape=False (None, 784) # 用於第一層是全連接層 # 我們看一下數據的shape print(mnist.train.images.shape) # 訓練數據圖片(55000, 28, 28, 1) print(mnist.train.labels.shape) # 訓練數據標簽(55000, 10) print(mnist.test.images.shape) # 測試數據圖片(10000, 28, 28, 1) print(mnist.test.labels.shape) # 測試數據圖片(10000, 10) print(mnist.validation.images.shape) # 驗證數據圖片(5000, 28, 28, 1) print(mnist.validation.labels.shape) # 驗證數據圖片(5000, 784) def train(): batch_size = FLAGS.batch_size batch_nums = mnist.train.images.shape[0] // batch_size # 一個epoch中應該包含多少batch數據 class_num = FLAGS.class_num epochs = FLAGS.epochs learning_rate = FLAGS.learning_rate ############ 保存檢查點的地址 ############ checkpoints_dir = FLAGS.checkpoints_dir # checkpoints # 如果檢查點不存在,則創建 if not os.path.exists(checkpoints_dir): os.makedirs(FLAGS.checkpoints_dir) ###################################################### # 創建圖 # ###################################################### graph = tf.Graph() # 自定義圖 # 在自己的圖中定義數據和操作 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 ############ 損失函數 ############ # 計算預測值和真實值之間的誤差 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=模型損失+權重正則化損失 ############ 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)) ############ 優化器 ############ variable_to_train = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES) # 可訓練變量列表 # 創建優化器,更新網絡參數,最小化loss, train_op = tf.train.AdamOptimizer(learning_rate).minimize(loss=total_loss, var_list=variable_to_train) ############ TensorBoard可視化 summary ############ summary_writer = tf.summary.FileWriter("./logs", graph=graph) # 創建事件文件 tf.summary.scalar(name="loss", tensor=total_loss) # 收集損失值變量 tf.summary.scalar(name='accuracy', tensor=accuracy) # 收集精度值變量 tf.summary.scalar(name='learning_rate', tensor=learning_rate) merged_summary_op = tf.summary.merge_all() # 將所有的summary合並為一個op ############ 模型保存和恢復 Saver ############ saver = tf.train.Saver(max_to_keep=5) ###################################################### # 創建會話 # ###################################################### max_acc = 0. config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=True) with tf.Session(config=config, graph=graph) as sess: # 加載模型,如果模型存在返回 是否加載成功和訓練步數 could_load, checkpoint_step = load_model(sess, saver, FLAGS.checkpoints_dir) if could_load: step = checkpoint_step print(" [*] 模型加載成功") else: print(" [!] 模型加載失敗") 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() # 記錄一下開始訓練的時間 # 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) # 使用真實數據填充placeholder,運行訓練模型和合並變量操作 _, 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) # 將每次迭代后的變量寫入事件文件 summary_writer.flush() # 強制summary_writer將緩存中的數據寫入到日志文件中(可選) ############ 可視化打印 ############ print("Epoch:[%2d] [%4d/%4d] time:%4.4f,loss:%.8f" % ( epoch, i, batch_nums, time.time() - start_time, loss)) # 打印一些可視化的數據,損失... # 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("優化完成!") def main(argv=None): train() if __name__ == '__main__': # logging.basicConfig(level=logging.INFO) tf.logging.set_verbosity(tf.logging.INFO) tf.app.run()
數據處理
數據處理因為每個專業領域的原因各不相同,而這不同點也是各位論文創新點的新方向。不同的我沒法講,但我總結了幾點相同的地方——batch數據生成。因為深度學習模型需要一個batch一個batch的喂數據進行訓練,所以我們的數據必須是batch的形式,這里衍生了三點問題
- 通過代碼批量讀取數據,
- 如何生成batch數據:由於篇幅過長,實在有很多地方要介紹和詳述,我把這一塊內容移到了這篇文章《TensorFlow讀取數據的三種方法》中
- 數據的shape:我舉兩個例子讓大家理解:圖片數據為4維 (batch_size, height,width, channels),序列數據為3維 (batch_size, time_steps, input_size),
- 不同的shape處理方法不同,選擇神經網絡模型單元也不同。我會在后面細講
模型搭建
閱讀這一節我默認大家已經學會了數據的batch讀取了。
模型搭建這一步很像我們小時候玩的搭積木,我這里以經典神經網絡模型VGG、Alex、ResNet、Google Inception Net為例講解,大家看代碼看多了也會很簡單的就找到,當然我是有一點私心的,我想把這些經典的網絡在這篇文章做一個tensorflow實現匯總,我細講第一個,大家可能看一個例子就懂了,看懂了就直接往下看,看不懂就多看幾個。
LeNet5模型
論文:1998_LeNet_Gradient-Based Learning Applied to Document Recognition
下面我們定義一個LeNet5模型,我們先定義需要用到的神經網絡單元,相同的代碼盡量封裝成函數的形式以節省代碼量和簡潔代碼

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) # 卷積操作 conv_layer = tf.nn.bias_add(conv_layer, conv_biases) # 加上偏置項 conv_layer = tf.nn.relu(conv_layer) # relu激活函數 return conv_layer def fc(input, output_size, init_bias=0.0, activeation_func=True, wd=None): input_shape = input.get_shape().as_list() # 創建 全連接權重 變量 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) # 創建 全連接偏置 變量 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) # 全連接計算 fc_layer = tf.nn.bias_add(fc_layer, fc_biases) # 加上偏置項 if activeation_func: fc_layer = tf.nn.relu(fc_layer) # rele激活函數 return fc_layer
然后利用我們搭建的神經網絡單元,搭建LeNet5神經網絡模型
# 訓練時:keep_prob=0.5 # 測試時:keep_prob=1.0 def leNet(inputs, class_num, keep_prob=0.5): # 第一層 卷積層 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") # 第二層 池化層 with tf.name_scope('layer2-pool1'): pool1 = tf.nn.max_pool(conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') # 第三層 卷積層 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") # 第四層 池化層 with tf.name_scope('layer4-pool2'): pool2 = tf.nn.max_pool(conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') # 后面要做全連接,因此要把數據變成2維 # 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
論文:2012_Alex_ImageNet Classification with Deep Convolutional Neural Networks
下面我們定義一個Alex模型,我們先定義需要用到的神經網絡單元,相同的代碼盡量封裝成函數的形式以節省代碼量和簡潔代碼

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) # 卷積操作 conv_layer = tf.nn.bias_add(conv_layer, conv_biases) # 加上偏置項 conv_layer = tf.nn.relu(conv_layer) # relu激活函數 return conv_layer

def fc(input, output_size, init_bias=0.0, activeation_func=True, wd=None): input_shape = input.get_shape().as_list() # 創建 全連接權重 變量 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) # 創建 全連接偏置 變量 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) # 全連接計算 fc_layer = tf.nn.bias_add(fc_layer, fc_biases) # 加上偏置項 if activeation_func: fc_layer = tf.nn.relu(fc_layer) # rele激活函數 return fc_layer

def LRN(input, depth_radius=2, alpha=0.0001, beta=0.75, bias=1.0): """Local Response Normalization 局部響應歸一化""" return tf.nn.local_response_normalization(input, depth_radius=depth_radius, alpha=alpha, beta=beta, bias=bias)
然后利用我們搭建的神經網絡單元,搭建Alex神經網絡模型
def alexNet(inputs, class_num, keep_prob=0.5): # 第一層卷積層 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") # 第二層卷積層 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") # 第三層卷積層 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") # 第四層卷積層 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") # 第五層卷積層 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 # 后面做全連接,所以要把shape改成2維 # shape=[batch, dim] flatten = tf.reshape(conv5, [-1, conv5_shape[1] * conv5_shape[2] * conv5_shape[3]]) # 第一層全連接層 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") # 第一層全連接層 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") # 第一層全連接層 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 # 模型輸出
VGG模型
論文:2014_VGG_Very Deep Convolutional Networks for Large-Scale Image Recognition
VGG有兩個比較有名的網絡:VGG16、VGG19,我在這里搭建VGG16,有興趣的朋友可以按照上面的模型結構自己用TensorFlow搭建VGG19模型
下面我們定義一個VGG16模型,和前面一樣,我們先定義需要用到的神經網絡單元,相同的代碼盡量封裝成函數的形式以節省代碼量和簡潔代碼
因為模型中同一個變量域中包含多個卷積操作,因此在卷積函數中套一層變量域

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): # 創建 全連接權重 變量 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) # 創建 全連接偏置 變量 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) # 全連接計算 fc_layer = tf.nn.bias_add(fc_layer, fc_biases) # 加上偏置項 if activeation_func: fc_layer = tf.nn.relu(fc_layer) # rele激活函數 return fc_layer
然后利用我們搭建的神經網絡單元,搭建VGG16神經網絡模型
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() # 后面做全連接,所以要把shape改成2維 # 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
上圖中有一個softmax層,我們也可以定義出來
class_num = 1000 # placeholder 定義 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
論文:
- 2016_ResNet_Deep Residual Learning for Image Recognition
- 2016_ResNet_Identity Mappings in Deep Residual Networks
ResNet的網絡結構如下圖所示
我們先定義需要用到的神經網絡單元

def batch_normalization(inputs, is_training, epsilon=0.001, decay=0.9): # 計算公式為: # inputs = (inputs-mean)/tf.sqrt(variance+epilon) # inputs = inputs * gamma + beta input_size = inputs.get_shape().as_list()[-1] # 擴大參數 gamma = tf.get_variable('gamma', input_size, tf.float32, initializer=tf.ones_initializer) # 也叫scale # 平移參數 beta = tf.get_variable('beta', input_size, tf.float32, initializer=tf.zeros_initializer) # 也叫shift # 移動均值 moving_mean = tf.get_variable('moving_mean', input_size, tf.float32, initializer=tf.zeros_initializer, trainable=False) # 移動方差 moving_variance = tf.get_variable('moving_variance', input_size, tf.float32, initializer=tf.ones_initializer, trainable=False) def mean_and_var_update(): # 這些op只有在訓練時才能進行 # 因為image是4維數據, 我們需要對[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") # 計算均值和方差 # 用滑動平均值來統計整體的均值和方差,在訓練階段並用不上,在測試階段才會用,這里是保證在訓練階段計算了滑動平均值 update_moving_mean = moving_averages.assign_moving_average(moving_mean, mean, decay=decay) # 應用滑動平均 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)) # 正則損失衰減率0.000004 conv_layer = tf.nn.conv2d(input, conv_weights, [1, stride, stride, 1], padding=padding) # 卷積操作 batch_norm = batch_normalization(conv_layer, output_size) conv_output = tf.nn.relu(batch_norm) # relu激活函數 return conv_output

def fc(input, output_size, activeation_func=True): input_shape = input.shape[-1] # 創建 全連接權重 變量 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)) # 創建 全連接偏置 變量 fc_biases = tf.get_variable(name="biases", shape=[output_size], initializer=tf.zeros_initializer, dtype=tf.float32) fc_layer = tf.matmul(input, fc_weights) # 全連接計算 fc_layer = tf.nn.bias_add(fc_layer, fc_biases) # 加上偏置項 if activeation_func: fc_layer = tf.nn.relu(fc_layer) # rele激活函數 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)
然后我們定義神經網絡框架
def inference(inputs, class_num, num_blocks=[3, 4, 6, 3], bottleneck=True): # data[1, 224, 224, 3] # 我們嘗試搭建50層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,如果這篇文章對大家有用,歡迎大家催促我。
RNN模型
Tensorflow中的CNN變數很少,而RNN卻豐富多彩,不僅在RNN Cell上有很多種、在實現上也有很多種,在用法上更是花樣百出。
五個基本的RNN Cell:RNNCell、BasicRNNCell、LSTMCell、BasicLSTMCell、GRUCell
RNN Cell的封裝和變形:MultiRNNCell(多層RNN)、DropoutWrapper、ResidualWrapper、DeviceWrapper
四種架構 (static+dynamic)*(單向+雙向)=4:static_rnn(靜態RNN)、dynamic_rnn(動態RNN)、static_bidirectional_rnn(靜態雙向RNN)、bidirectional_dynamic_rnn(動態雙向RNN)
五種手法 (one+many)*(one+many) +1=5:
- one to one(1 vs 1):輸入一個,輸出一個。其實和全連接神經網絡並沒有什么區別,這一類別算不得是 RNN。
- one to many(1 vs N):輸入一個,輸出多個。圖像標注,輸入一個圖片,得到對圖片的語言描述
- many to one(N vs 1):輸入多個,輸出一個。序列分類,把序列壓縮成一個向量
- many to many(N vs N):輸入多個,輸出多個。兩者長度可以不一樣。翻譯任務
- many to many(N vs N):輸入多個,輸出多個。兩者長度一樣。char RNN
我們先定義需要用到的神經網絡單元
全連接層

def fc(input, output_size, activeation_func=tf.nn.relu): input_shape = input.shape[-1] # 創建 全連接權重 變量 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)) # 創建 全連接偏置 變量 fc_biases = tf.get_variable(name="biases", shape=[output_size], initializer=tf.zeros_initializer, dtype=tf.float32) fc_layer = tf.matmul(input, fc_weights) # 全連接計算 fc_layer = tf.nn.bias_add(fc_layer, fc_biases) # 加上偏置項 if activeation_func: fc_layer = activeation_func(fc_layer) # rele激活函數 return fc_layer
單層 靜態/動態 LSTM/GRU

####################################### # 單層 靜態/動態 LSTM/GRU # ####################################### # 單層靜態LSTM def single_layer_static_lstm(input_x, time_steps, hidden_size): """ :param input_x: 輸入張量 形狀為[batch_size, n_steps, input_size] :param n_steps: 時序總數 :param n_hidden: LSTM單元輸出的節點個數 即隱藏層節點數 """ # 把輸入input_x按列拆分,並返回一個有n_steps個張量組成的list # 如batch_sizex28x28的輸入拆成[(batch_size,28),((batch_size,28))....] # 如果是調用的是靜態rnn函數,需要這一步處理 即相當於把序列作為第一維度 input_x1 = tf.unstack(input_x, num=time_steps, axis=1) lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(num_units=hidden_size) # 創建LSTM_cell # 靜態rnn函數傳入的是一個張量list 每一個元素都是一個(batch_size,input_size)大小的張量 output, states = tf.nn.static_rnn(cell=lstm_cell, inputs=input_x1, dtype=tf.float32) # 通過cell類構建RNN return output, states # 單層靜態gru def single_layer_static_gru(input, time_steps, hidden_size): """ :param input: 輸入張量 形狀為[batch_size, n_steps, input_size] :param n_steps: 時序總數 :param n_hidden: gru單元輸出的節點個數 即隱藏層節點數 :return: 返回靜態單層GRU單元的輸出,以及cell狀態 """ # 把輸入input_x按列拆分,並返回一個有n_steps個張量組成的list # 如batch_sizex28x28的輸入拆成[(batch_size,28),((batch_size,28))....] # 如果是調用的是靜態rnn函數,需要這一步處理 即相當於把序列作為第一維度 input_x = tf.unstack(input, num=time_steps, axis=1) gru_cell = tf.nn.rnn_cell.GRUCell(num_units=hidden_size) # 創建GRU_cell # 靜態rnn函數傳入的是一個張量list 每一個元素都是一個(batch_size,input_size)大小的張量 output, states = tf.nn.static_rnn(cell=gru_cell, inputs=input_x, dtype=tf.float32) # 通過cell類構建RNN return output, states # 單層動態LSTM def single_layer_dynamic_lstm(input, time_steps, hidden_size): """ :param input_x: 輸入張量 形狀為[batch_size, time_steps, input_size] :param time_steps: 時序總數 :param hidden_size: LSTM單元輸出的節點個數 即隱藏層節點數 :return: 返回動態單層LSTM單元的輸出,以及cell狀態 """ lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(num_units=hidden_size) # 創建LSTM_cell # 動態rnn函數傳入的是一個三維張量,[batch_size,time_steps, input_size] 輸出也是這種形狀 output, states = tf.nn.dynamic_rnn(cell=lstm_cell, inputs=input, dtype=tf.float32) # 通過cell類構建RNN output = tf.transpose(output, [1, 0, 2]) # 注意這里輸出需要轉置 轉換為時序優先的 return output, states # 單層動態gru def single_layer_dynamic_gru(input, time_steps, hidden_size): """ :param input: 輸入張量 形狀為[batch_size, time_steps, input_size] :param time_steps: 時序總數 :param hidden_size: GRU單元輸出的節點個數 即隱藏層節點數 :return: 返回動態單層GRU單元的輸出,以及cell狀態 """ gru_cell = tf.nn.rnn_cell.GRUCell(num_units=hidden_size) # 創建GRU_cell # 動態rnn函數傳入的是一個三維張量,[batch_size,n_steps,input_size] 輸出也是這種形狀 output, states = tf.nn.dynamic_rnn(cell=gru_cell, inputs=input, dtype=tf.float32) # 通過cell類構建RNN output = tf.transpose(output, [1, 0, 2]) # 注意這里輸出需要轉置 轉換為時序優先的 return output, states
多層 靜態/動態 LSTM/GRU

####################################### # 多層 靜態/動態 LSTM/GRU # ####################################### # 多層靜態LSTM網絡 def multi_layer_static_lstm(input, time_steps, hidden_size): """ :param input: 輸入張量 形狀為[batch_size,time_steps,input_size] :param time_steps: 時序總數 :param n_hidden: LSTM單元輸出的節點個數 即隱藏層節點數 :return: 返回靜態多層LSTM單元的輸出,以及cell狀態 """ # 把輸入input_x按列拆分,並返回一個有n_steps個張量組成的list # 如batch_sizex28x28的輸入拆成[(batch_size,28),((batch_size,28))....] # 如果是調用的是靜態rnn函數,需要這一步處理 即相當於把序列作為第一維度 input_x1 = tf.unstack(input, num=time_steps, axis=1) # 多層RNN的實現 例如cells=[cell1,cell2,cell3],則表示一共有三層 mcell = tf.nn.rnn_cell.MultiRNNCell( [tf.nn.rnn_cell.LSTMCell(num_units=hidden_size) for _ in range(3)]) # 靜態rnn函數傳入的是一個張量list 每一個元素都是一個(batch_size,input_size)大小的張量 output, states = tf.nn.static_rnn(cell=mcell, inputs=input_x1, dtype=tf.float32) return output, states # 多層靜態GRU def multi_layer_static_gru(input, time_steps, hidden_size): """ :param input_x: 輸入張量 形狀為[batch_size,n_steps,input_size] :param time_steps: 時序總數 :param hidden_size: gru單元輸出的節點個數 即隱藏層節點數 :return: 返回靜態多層GRU單元的輸出,以及cell狀態 """ # 把輸入input_x按列拆分,並返回一個有n_steps個張量組成的list # 如batch_sizex28x28的輸入拆成[(batch_size,28),((batch_size,28))....] # 如果是調用的是靜態rnn函數,需要這一步處理 即相當於把序列作為第一維度 input_x = tf.unstack(input, num=time_steps, axis=1) # 多層RNN的實現 例如cells=[cell1,cell2,cell3],則表示一共有三層 mcell = tf.nn.rnn_cell.MultiRNNCell( [tf.nn.rnn_cell.GRUCell(num_units=hidden_size) for _ in range(3)]) # 靜態rnn函數傳入的是一個張量list 每一個元素都是一個(batch_size,input_size)大小的張量 output, states = tf.nn.static_rnn(cell=mcell, inputs=input_x, dtype=tf.float32) return output, states # 多層靜態GRU和LSTM 混合 def multi_layer_static_mix(input, time_steps, hidden_size): """ :param input: 輸入張量 形狀為[batch_size,n_steps,input_size] :param time_steps: 時序總數 :param hidden_size: gru單元輸出的節點個數 即隱藏層節點數 :return: 返回靜態多層GRU和LSTM混合單元的輸出,以及cell狀態 """ # 把輸入input_x按列拆分,並返回一個有n_steps個張量組成的list # 如batch_sizex28x28的輸入拆成[(batch_size,28),((batch_size,28))....] # 如果是調用的是靜態rnn函數,需要這一步處理 即相當於把序列作為第一維度 input_x = tf.unstack(input, num=time_steps, axis=1) # 可以看做2個隱藏層 lstm_cell = tf.nn.rnn_cell.LSTMCell(num_units=hidden_size) gru_cell = tf.nn.rnn_cell.GRUCell(num_units=hidden_size * 2) # 多層RNN的實現 例如cells=[cell1,cell2],則表示一共有兩層,數據經過cell1后還要經過cells mcell = tf.nn.rnn_cell.MultiRNNCell(cells=[lstm_cell, gru_cell]) # 靜態rnn函數傳入的是一個張量list 每一個元素都是一個(batch_size,input_size)大小的張量 output, states = tf.nn.static_rnn(cell=mcell, inputs=input_x, dtype=tf.float32) return output, states # 多層動態LSTM def multi_layer_dynamic_lstm(input, time_steps, hidden_size): """ :param input: 輸入張量 形狀為[batch_size,n_steps,input_size] :param time_steps: 時序總數 :param hidden_size: LSTM單元輸出的節點個數 即隱藏層節點數 :return: 返回動態多層LSTM單元的輸出,以及cell狀態 """ # 多層RNN的實現 例如cells=[cell1,cell2],則表示一共有兩層,數據經過cell1后還要經過cells mcell = tf.nn.rnn_cell.MultiRNNCell( [tf.nn.rnn_cell.LSTMCell(num_units=hidden_size) for _ in range(3)]) # 動態rnn函數傳入的是一個三維張量,[batch_size,n_steps,input_size] 輸出也是這種形狀 output, states = tf.nn.dynamic_rnn(cell=mcell, inputs=input, dtype=tf.float32) # 注意這里輸出需要轉置 轉換為時序優先的 output = tf.transpose(output, [1, 0, 2]) return output, states # 多層動態GRU def multi_layer_dynamic_gru(input, time_steps, hidden_size): """ :param input: 輸入張量 形狀為[batch_size,n_steps,input_size] :param time_steps: 時序總數 :param hidden_size: gru單元輸出的節點個數 即隱藏層節點數 :return: 返回動態多層GRU單元的輸出,以及cell狀態 """ # 多層RNN的實現 例如cells=[cell1,cell2],則表示一共有兩層,數據經過cell1后還要經過cells mcell = tf.nn.rnn_cell.MultiRNNCell( [tf.nn.rnn_cell.GRUCell(num_units=hidden_size) for _ in range(3)]) # 動態rnn函數傳入的是一個三維張量,[batch_size,n_steps,input_size] 輸出也是這種形狀 output, states = tf.nn.dynamic_rnn(cell=mcell, inputs=input, dtype=tf.float32) # 注意這里輸出需要轉置 轉換為時序優先的 output = tf.transpose(output, [1, 0, 2]) return output, states # 多層動態GRU和LSTM 混合 def multi_layer_dynamic_mix(input, time_steps, hidden_size): """ :param input: 輸入張量 形狀為[batch_size,n_steps,input_size] :param time_steps: 時序總數 :param hidden_size: gru單元輸出的節點個數 即隱藏層節點數 :return: 返回動態多層GRU和LSTM混合單元的輸出,以及cell狀態 """ # 可以看做2個隱藏層 gru_cell = tf.nn.rnn_cell.GRUCell(num_units=hidden_size * 2) lstm_cell = tf.nn.rnn_cell.LSTMCell(num_units=hidden_size) # 多層RNN的實現 例如cells=[cell1,cell2],則表示一共有兩層,數據經過cell1后還要經過cells mcell = tf.nn.rnn_cell.MultiRNNCell(cells=[lstm_cell, gru_cell]) # 動態rnn函數傳入的是一個三維張量,[batch_size,n_steps,input_size] 輸出也是這種形狀 output, states = tf.nn.dynamic_rnn(cell=mcell, inputs=input, dtype=tf.float32) # 注意這里輸出需要轉置 轉換為時序優先的 output = tf.transpose(output, [1, 0, 2]) return output, states
單層/多層 雙向 靜態/動態 LSTM/GRU

####################################### # 單層/多層 雙向 靜態/動態 LSTM/GRU # ####################################### # 單層靜態雙向LSTM def single_layer_static_bi_lstm(input, time_steps, hidden_size): """ :param input: 輸入張量 形狀為[batch_size,time_steps,input_size] :param time_steps: 時序總數 :param hidden_size: LSTM單元輸出的節點個數 即隱藏層節點數 :return: 返回單層靜態雙向LSTM單元的輸出,以及cell狀態 """ # 把輸入input_x按列拆分,並返回一個有n_steps個張量組成的list # 如batch_sizex28x28的輸入拆成[(batch_size,28),((batch_size,28))....] # 如果是調用的是靜態rnn函數,需要這一步處理 即相當於把序列作為第一維度 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 # 靜態rnn函數傳入的是一個張量list 每一個元素都是一個(batch_size,input_size)大小的張量 # 這里的輸出output是一個list 每一個元素都是前向輸出,后向輸出的合並 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 # 單層動態雙向LSTM def single_layer_dynamic_bi_lstm(input, time_steps, hidden_size): """ :param input: 輸入張量 形狀為[batch_size,time_steps,input_size] :param time_steps: 時序總數 :param hidden_size: gru單元輸出的節點個數 即隱藏層節點數 :return: 返回單層動態雙向LSTM單元的輸出,以及cell狀態 """ 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 # 動態rnn函數傳入的是一個三維張量,[batch_size,time_steps,input_size] 輸出是一個元組 每一個元素也是這種形狀 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合並 (?,28,128) (?,28,128)按最后一維合並(?,28,256) output = tf.transpose(output, [1, 0, 2]) # 注意這里輸出需要轉置 轉換為時序優先的 return output, state # 多層靜態雙向LSTM def multi_layer_static_bi_lstm(input, time_steps, hidden_size): """ :param input: 輸入張量 形狀為[batch_size,time_steps,input_size] :param time_steps: 時序總數 :param hidden_size: LSTM單元輸出的節點個數 即隱藏層節點數 :return: 返回多層靜態雙向LSTM單元的輸出,以及cell狀態 """ # 把輸入input_x按列拆分,並返回一個有n_steps個張量組成的list # 如batch_sizex28x28的輸入拆成[(batch_size,28),((batch_size,28))....] # 如果是調用的是靜態rnn函數,需要這一步處理 即相當於把序列作為第一維度 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 # 靜態rnn函數傳入的是一個張量list 每一個元素都是一個(batch_size,input_size)大小的張量 # 這里的輸出output是一個list 每一個元素都是前向輸出,后向輸出的合並 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 # 多層動態雙向LSTM def multi_layer_dynamic_bi_lstm(input, time_steps, hidden_size): """ :param input: 輸入張量 形狀為[batch_size,n_steps,input_size] :param time_steps: 時序總數 :param hidden_size: gru單元輸出的節點個數 即隱藏層節點數 :return: 返回多層動態雙向LSTM單元的輸出,以及cell狀態 """ 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 # 動態rnn函數傳入的是一個三維張量,[batch_size,n_steps,input_size] 輸出也是這種形狀, # input_size變成了正向和反向合並之后的 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]) # 注意這里輸出需要轉置 轉換為時序優先的 return output, fw_state, bw_state
然后我們定義神經網絡框架
def RNN_inference(inputs, class_num, time_steps, hidden_size): """ :param inputs: [batch_size, n_steps, input_size] :param class_num: 類別數 :param time_steps: 時序總數 :param n_hidden: LSTM單元輸出的節點個數 即隱藏層節點數 """ ####################################### # 單層 靜態/動態 LSTM/GRU # ####################################### # outputs, states = single_layer_static_lstm(inputs, time_steps, hidden_size) # 單層靜態LSTM # outputs, states = single_layer_static_gru(inputs, time_steps, hidden_size) # 單層靜態gru # outputs, states = single_layer_dynamic_lstm(inputs, time_steps, hidden_size) # 單層動態LSTM # outputs, states = single_layer_dynamic_gru(inputs, time_steps, hidden_size) # 單層動態gru ####################################### # 多層 靜態/動態 LSTM/GRU # ####################################### # outputs, states = multi_layer_static_lstm(inputs, time_steps, hidden_size) # 多層靜態LSTM網絡 # outputs, states = multi_layer_static_gru(inputs, time_steps, hidden_size) # 多層靜態GRU # outputs, states = multi_layer_static_mix(inputs, time_steps, hidden_size) # 多層靜態GRU和LSTM 混合 # outputs, states = multi_layer_dynamic_lstm(inputs, time_steps, hidden_size) # 多層動態LSTM # outputs, states = multi_layer_dynamic_gru(inputs, time_steps, hidden_size) # 多層動態GRU # outputs, states = multi_layer_dynamic_mix(inputs, time_steps, hidden_size) # 多層動態GRU和LSTM 混合 ####################################### # 單層/多層 雙向 靜態/動態 LSTM/GRU # ####################################### # outputs, fw_state, bw_state = single_layer_static_bi_lstm(inputs, time_steps, hidden_size) # 單層靜態雙向LSTM # outputs, state = single_layer_dynamic_bi_lstm(inputs, time_steps, hidden_size) # 單層動態雙向LSTM # outputs, fw_state, bw_state = multi_layer_static_bi_lstm(inputs, time_steps, hidden_size) # 多層靜態雙向LSTM outputs, fw_state, bw_state = multi_layer_dynamic_bi_lstm(inputs, time_steps, hidden_size) # 多層動態雙向LSTM # output靜態是 time_step=28個(batch=128, output=128)組成的列表 # output動態是 (time_step=28, batch=128, output=128) print('hidden:', outputs[-1].shape) # 最后一個時序的shape(128,128) # 取LSTM最后一個時序的輸出,然后經過全連接網絡得到輸出值 fc_output = fc(input=outputs[-1], output_size=class_num, activeation_func=tf.nn.relu) return fc_output
設置全局變量和超參數
在模型訓練之前我們首先會定義一些超參數:batch_size、batch_nums、class_num、epochs、learning_rate
batch_size = FLAGS.batch_size batch_nums = mnist.train.images.shape[0] // batch_size # 一個epoch中應該包含多少batch數據 class_num = FLAGS.class_num epochs = FLAGS.epochs learning_rate = FLAGS.learning_rate
保存檢查點的地址
############ 保存檢查點的地址 ############ checkpoints_dir = FLAGS.checkpoints_dir # checkpoints # 如果檢查點不存在,則創建 if not os.path.exists(checkpoints_dir): os.makedirs(FLAGS.checkpoints_dir)
創建圖
這一步可以不設置,因為tensorflow有一個默認圖,我們定義的操作都是在默認圖上的,當然我們也可以定義自己的,方便管理。
###################################################### # 創建圖 # ###################################################### graph = tf.Graph() # 自定義圖 # 在自己的圖中定義數據和操作 with graph.as_default():
占位符
一般我們會把input和label做成placeholder,方便我們使用把不同的batch數據傳入網絡,一些其他的超參數也可以做成placeholder,比如learning_rate、dorpout_keep_prob。一般在搭建模型的時候把placeholder的變量傳入模型,在訓練模型sess.run(train_op, feed_dict)的時候通過參數feed_dict={input:真實數據,label:真實標簽} 把真實的數據傳入神經網絡。
inputs = tf.placeholder(dtype="float", shape=[None, 28, 28, 1], name='inputs') labels = tf.placeholder(dtype="float", shape=[None, class_num], name='labels') # 看個人喜歡,有的人在初始化定義中就定義了learning_rate,有的人喜歡通過feed傳learning_rate learning_rate = tf.placeholder("float", None, name='learning_rate') # 如果網絡結構有dropout層,需要定義keep_probn,如果沒有則不需要 # 訓練的時候需要,測試的時候需要設置成1 keep_prob = tf.placeholder(dtype="float", name='keep_prob')
搭建模型
傳進入的都是placeholder數據,不是我們之前整理好的batch數據。
############ 搭建模型 ############ logits = alexNet(inputs, class_num, keep_prob=keep_prob) # 使用placeholder搭建模型
構建損失
分類任務一般輸出的是每個類別的概率向量,因此模型輸出最后都要經過softmax轉換成概率。一般經過softmax的輸出損失函數都是交叉熵損失函數,tensorflow有將以上兩步合在一起的現成函數 tf.nn.softmax_cross_entropy_with_logits
############ 損失函數 ############ 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=模型損失+權重正則化損失
自定義損失
以后更新,歡迎大家催我。
模型精度
在測試數據集上的精度
############ 模型精度 ############ predict = tf.argmax(logits, 1) # 模型預測結果 accuracy = tf.reduce_mean(tf.cast(tf.equal(predict, tf.argmax(labels, 1)), tf.float32))
自定義度量
以后更新,歡迎大家催我。
優化器
創建優化器,更新網絡參數,最小化loss
優化器的種類有很多種,但是用法都差不多,常用的優化器有:
- tf.train.AdamOptimizer
-
tf.train.GradientDescentOptimizer
- tf.train.RMSPropOptimizer
下面以Adam優化器為例
############ 優化器 ############ variable_to_train = tf.get_collection(tf.GraphKeys.TRAINABLE_VARIABLES) # 可訓練變量列表 global_step = tf.Variable(0, trainable=False) # 訓練step # 設置學習率衰減 learning_rate = tf.train.exponential_decay(learning_rate=learning_rate, # 初始學習率 global_step=global_step, decay_steps=batch_nums, # 多少步衰減一次 decay_rate=0.1, # 衰減率 staircase=True) # 以階梯的形式衰減 # 創建Adam優化器,更新模型參數,最小化損失函數 train_op = tf.train.AdamOptimizer(learning_rate).minimize(loss=total_loss, # 損失函數 global_step=global_step, var_list=variable_to_train) # 通過訓練需要更新的參數列表
講解:
- variable_to_train:上面的代碼定義了可訓練變量,我只是把列出了模型默認的可訓練變量,這一個步是tensorflow默認的,如果不設置也沒有關系。我寫出來的原因是,有的大牛會這么寫,對不同的可訓練變量分別進行不同的優化,希望大家看到我的代碼,下次看到別人的不會覺得陌生。
- global_step:大多數人會用step=0,然后在訓練的時候step+=1的方式更新step,但是本文介紹的是另一種方式,以tf.Variable的方式定義step,在模型訓練的時候傳入sess.run,global_step會自動+1更新
- learning_rate:本文還設置了學習率衰減,大家也可以不設置,以固定的學習率訓練模型,但是對於大型項目,還是推薦設置。
移動平均值更新參數
采用移動平均值的方式更新損失值和模型參數

def train(total_loss, global_step): lr = tf.train.exponential_decay(0.01, global_step, decay_steps=350, decay_rate=0.1, staircase=True) # 采用滑動平均的方法更新損失值 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]) # 計算損失值的影子變量op # 計算梯度 with tf.control_dependencies([loss_averages_op]): # 控制計算指定,只有執行了括號中的語句才能執行下面的語句 opt = tf.train.GradientDescentOptimizer(lr) # 創建優化器 grads = opt.compute_gradients(total_loss) # 計算梯度 # 應用梯度 apply_gradient_op = opt.apply_gradients(grads, global_step=global_step) # 采用滑動平均的方法更新參數 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()表示執行完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可視化 summary
############ TensorBoard可視化 summary ############ summary_writer = tf.summary.FileWriter(logdir="./logs", graph=graph) # 創建事件文件 tf.summary.scalar(name="losses", tensor=total_loss) # 收集損失值變量 tf.summary.scalar(name="acc", tensor=accuracy) # 收集精度值變量 tf.summary.scalar(name='learning_rate', tensor=learning_rate) merged_summary_op = tf.summary.merge_all() # 將所有的summary合並為一個op
模型保存和恢復 Saver
saver = tf.train.Saver(max_to_keep=5) # 保存最新的5個檢查點
創建會話
配置會話
在創建會話之前我們一般都要配置會話,比如使用GPU還是CPU,用多少GPU等等。
我們一般使用 tf.ConfigProto()配置Session運行參數&&GPU設備指定
config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=True) config.gpu_options.per_process_gpu_memory_fraction = 0.4 # 占用40%顯存 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):記錄設備指派情況
設置tf.ConfigProto()中參數log_device_placement = True,獲取 operations 和 Tensor 被指派到哪個設備(幾號CPU或幾號GPU)上運行,會在終端打印出各項操作是在哪個設備上運行的。
tf.ConfigProto(allow_soft_placement=True):自動選擇運行設備
在TensorFlow中,通過命令 "with tf.device('/cpu:0'):",允許手動設置操作運行的設備。如果手動設置的設備不存在或者不可用,就會導致tf程序等待或異常,為了防止這種情況,可以設置tf.ConfigProto()中參數allow_soft_placement=True,自動選擇一個存在並且可用的設備來運行操作。
config.gpu_options.allow_growth = True
當使用GPU時候,Tensorflow運行自動慢慢達到最大GPU的內存
tf.test.is_built_with_cuda():返回是否能夠使用GPU進行運算
為了加快運行效率,TensorFlow在初始化時會嘗試分配所有可用的GPU顯存資源給自己,這在多人使用的服務器上工作就會導致GPU占用,別人無法使用GPU工作的情況。這時我們需要限制GPU資源使用,詳細實現方法請參考我的另一篇博客 tensorflow常用函數 Ctrl+F搜索“限制GPU資源使用”
創建會話Session
Session有兩種創建方式:
sess = tf.Session(config=config, graph=graph) # 或通過with的方式創建Session with tf.Session(config=config, graph=graph) as sess:
如果我們之前自定義了graph,則在會話中也要配置graph,如果之前沒有自定義graph,使用的是tensorflow默認graph,則在會話不用自己去定義,tensorflow會自動找到默認圖。
在訓練模型之前我們首先要設置一個高級一點的東西,那就是檢查是否有之前保存好的模型,如果有着接着前面的繼續訓練,如果沒有則從頭開始訓練模型。
恢復/重新訓練
定義一個檢查模型是否存在的函數,為了美觀,可以把這個函數放在最上面,或者其他腳本中,通過import導入。

def load_model(sess, saver, checkpoint_dir): """加載模型,看看還能不能加一個功能,必須現在的檢查檢點是1000,但是我的train是100,要報錯 還有就是讀取之前的模型繼續訓練的問題 checkpoint_dir = checkpoint""" # 通過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)) # 現在不知道checkpoint文件名時怎樣的,因此不知道里面如何運行 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),則使用下面diamante
# 加載模型,如果模型存在返回 是否加載成功和訓練步數 could_load, checkpoint_step = load_model(sess, saver, "./log") if could_load: print(" [*] 加載成功") else: print(" [!] 加載失敗") try: tf.global_variables_initializer().run() except: tf.initialize_all_variables().run()
如果大家想使用step=0,step+=1,則可以使用下面代碼
# 加載模型,如果模型存在返回 是否加載成功和訓練步數 could_load, checkpoint_step = load_model(sess, saver, FLAGS.checkpoints_dir) if could_load: step = checkpoint_step print(" [*] 模型加載成功") else: print(" [!] 模型加載失敗") try: tf.global_variables_initializer().run() except: tf.initialize_all_variables().run() step = 0
開始訓練
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) # 使用真實數據填充placeholder,運行訓練模型和合並變量操作 _, 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) # 將每次迭代后的變量寫入事件文件 summary_writer.flush() # 強制summary_writer將緩存中的數據寫入到日志文件中(可選) ############ 可視化打印 ############ print("Epoch:[%2d] [%4d/%4d] time:%4.4f,loss:%.8f" % ( epoch, i, batch_nums, time.time() - start_time, loss)) # 打印一些可視化的數據,損失... 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("優化完成!")
模型評估
eval.py
模型評估的代碼和模型訓練的代碼很像,只不過不需要對模型進行訓練而已。
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", '保存檢查點的地址') FLAGS = tf.flags.FLAGS # 從MNIST_data/中讀取MNIST數據。當數據不存在時,會自動執行下載 mnist = input_data.read_data_sets('./data', one_hot=True, reshape=False) # 將數組張換成圖片形式 print(mnist.train.images.shape) # 訓練數據圖片(55000, 28, 28, 1) print(mnist.train.labels.shape) # 訓練數據標簽(55000, 10) print(mnist.test.images.shape) # 測試數據圖片(10000, 28, 28, 1) print(mnist.test.labels.shape) # 測試數據圖片(10000, 10) print(mnist.validation.images.shape) # 驗證數據圖片(5000, 28, 28, 1) print(mnist.validation.labels.shape) # 驗證數據圖片(5000, 10) def evaluate(): batch_size = FLAGS.batch_size batch_nums = mnist.train.images.shape[0] // batch_size # 一個epoch中應該包含多少batch數據 class_num = FLAGS.class_num test_batch_size = 5000 test_batch_num = mnist.test.images.shape[0] // test_batch_size ############ 保存檢查點的地址 ############ checkpoints_dir = FLAGS.checkpoints_dir # checkpoints # 如果檢查點不存在,則創建 if not os.path.exists(checkpoints_dir): print("模型文件不存在,無法進行評估") ###################################################### # 創建圖 # ###################################################### graph = tf.Graph() # 自定義圖 # 在自己的圖中定義數據和操作 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)) ############ 模型保存和恢復 Saver ############ saver = tf.train.Saver(max_to_keep=5) ###################################################### # 創建會話 # ###################################################### config = tf.ConfigProto(allow_soft_placement=True, log_device_placement=True) with tf.Session(config=config, graph=graph) as sess: # 加載模型,如果模型存在返回 是否加載成功和訓練步數 could_load, checkpoint_step = load_model(sess, saver, FLAGS.checkpoints_dir) if could_load: print(" [*] 加載成功") else: print(" [!] 加載失敗") raise ValueError("模型文件不存在,無法進行評估") 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("模型精度為:", 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()
參考文獻
github搜索tensorflow AlexNet
github_finetune_alexnet_with_tensorflow
github_AlexNet_with_tensorflow