一.AlexNet介紹

https://blog.csdn.net/MESSI_JAMES/article/details/81384534#t8

二.過程介紹

一次完整的訓練模型和評估模型的過程一般分為 3 個步驟：
1.載入資料，
2.定義網路模型，
3. 訓練模型和評估模型。

三.程式碼實現

import tensorflow as tf
# 1.載入資料

# 輸入資料
from tensorflow.examples.tutorials.mnist import input_data
mnist = input_data.read_data_sets("MNIST_data", one_hot=True)
# 定義網路的一引數
learning_rate = 0.001
training_iters = 20000
batch_size = 128
display_step = 10
# 定義網路的引數
n_input = 784 # 輸入的維度(img shape: 28×28)
n_classes = 10 # 標記的維度 (0-9 digits)
dropout = 0.75 # Dropout 的概率，輸出的可能性
# 輸入佔位符
x = tf.placeholder(tf.float32, [None, n_input])
y = tf.placeholder(tf.float32, [None, n_classes])
keep_prob = tf.placeholder(tf.float32) #dropout

# 2.構建網路模型

# 定義卷積操作
def conv2d(name, x, W, b, strides=1):
    x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME')
    x = tf.nn.bias_add(x, b)
    return tf.nn.relu(x, name=name) # 使用 relu 啟用函式
# 定義池化層操作
def maxpool2d(name, x, k=2):
    return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1],
    padding='SAME', name=name)
# 規範化操作
def norm(name, l_input, lsize=4):
    return tf.nn.lrn(l_input, lsize, bias=1.0, alpha=0.001 / 9.0,
    beta=0.75, name=name)
# 定義所有的網路引數
weights = {
'wc1': tf.Variable(tf.random_normal([11, 11, 1, 96])),
'wc2': tf.Variable(tf.random_normal([5, 5, 96, 256])),
'wc3': tf.Variable(tf.random_normal([3, 3, 256, 384])),
'wc4': tf.Variable(tf.random_normal([3, 3, 384, 384])),
'wc5': tf.Variable(tf.random_normal([3, 3, 384, 256])),
'wd1': tf.Variable(tf.random_normal([4*4*256, 4096])),
'wd2': tf.Variable(tf.random_normal([4096, 4096])),
'out': tf.Variable(tf.random_normal([4096, 10]))
}
biases = {
'bc1': tf.Variable(tf.random_normal([96])),
'bc2': tf.Variable(tf.random_normal([256])),
'bc3': tf.Variable(tf.random_normal([384])),
'bc4': tf.Variable(tf.random_normal([384])),
'bc5': tf.Variable(tf.random_normal([256])),
'bd1': tf.Variable(tf.random_normal([4096])),
'bd2': tf.Variable(tf.random_normal([4096])),
'out': tf.Variable(tf.random_normal([n_classes]))
}

# 定義整個網路
def alex_net(x, weights, biases, dropout):
    # Reshape input picture
    x = tf.reshape(x, shape=[-1, 28, 28, 1])
    # 第一層卷積
    # 卷積
    conv1 = conv2d('conv1', x, weights['wc1'], biases['bc1'])
    # 下采樣
    pool1 = maxpool2d('pool1', conv1, k=2)
    # 規範化
    norm1 = norm('norm1', pool1, lsize=4)
    # 第二層卷積
    # 卷積
    conv2 = conv2d('conv2', conv1, weights['wc2'], biases['bc2'])
    # 最大池化（向下取樣）
    pool2 = maxpool2d('pool2', conv2, k=2)
    # 規範化
    norm2 = norm('norm2', pool2, lsize=4)
    # 第三層卷積
    # 卷積
    conv3 = conv2d('conv3', norm2, weights['wc3'], biases['bc3'])
    # 下采樣
    pool3 = maxpool2d('pool3', conv3, k=2)
    # 規範化
    norm3 = norm('norm3', pool3, lsize=4)
    # 第四層卷積
    conv4 = conv2d('conv4', norm3, weights['wc4'], biases['bc4'])
    # 第五層卷積
    conv5 = conv2d('conv5', norm3, weights['wc5'], biases['bc5'])
    # 下采樣
    pool5 = maxpool2d('pool5', conv5, k=2)
    # 規範化
    norm5 = norm('norm5', pool5, lsize=4)
    # 全連線層 1
    fc1 = tf.reshape(norm5, [-1, weights['wd1'].get_shape().as_list()[0]])
    fc1 = tf.add(tf.matmul(fc1, weights['wd1']), biases['bd1'])
    fc1 = tf.nn.relu(fc1)
    # dropout
    fc1 = tf.nn.dropout(fc1, dropout)
    # 全連線層 2
    fc2 = tf.reshape(fc1, [-1, weights['wd1'].get_shape().as_list()[0]])
    fc2 = tf.add(tf.matmul(fc2, weights['wd1']), biases['bd1'])
    fc2 = tf.nn.relu(fc2)
    # dropout
    fc2 = tf.nn.dropout(fc2, dropout)
    # 輸出層
    out = tf.add(tf.matmul(fc2, weights['out']), biases['out'])
    return out

# 構建模型
pred = alex_net(x, weights, biases, keep_prob)
# 定義損失函式和優化器
cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=pred, labels=y))
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(cost)
# 評估函式
correct_pred = tf.equal(tf.argmax(pred, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))

# 3.訓練模型和評估模型
# 初始化變數
init = tf.global_variables_initializer()
with tf.Session() as sess:
    sess.run(init)
    step = 1
    # 開始訓練，直到達到 training_iters，即 20000
    while step * batch_size < training_iters:
        batch_x, batch_y = mnist.train.next_batch(batch_size)
        sess.run(optimizer, feed_dict={x: batch_x, y: batch_y,
                                       keep_prob: dropout})
        if step % display_step == 0:
            # 計算損失值和準確度，輸出
            loss, acc = sess.run([cost, accuracy], feed_dict={x: batch_x,
                                                              y: batch_y,
                                                              keep_prob: 1.})
            print("Iter " + str(step * batch_size) + ", Minibatch Loss= " + \
                  "{:.6f}".format(loss) + ", Training Accuracy= " + \
                  "{:.5f}".format(acc))
        step+=1
    print("訓練完成!")
    # 計算測試集的準確度
    print("測試集的準確度:", \
          sess.run(accuracy, feed_dict={x: mnist.test.images[:256],
                                        y: mnist.test.labels[:256],
                                        keep_prob: 1.}))
 

四.結果展示

五.總結

我們也可以像實現AlexNet的模型這樣，用TensorFlow 實現其他網路（如VGGNet、GoogLeNet、
ResNet），具體實現的步驟我們總結如下：
（1）仔細研讀該網路的論文，理解每一層的輸入/輸出值以及網路結構；
（2）按照載入資料，定義網路模型，訓練模型和評估模型這樣的步驟實現網路。