github部落格傳送門
csdn部落格傳送門

本章所需知識:

1. 沒有基礎的請觀看深度學習系列視訊
2. tensorflow

首先附上百度百科不要錢的網路結構圖:

再附上 恩達老師的視覺化極強的網路結構圖:

接著加上我自己使用Tensorflow實現的程式碼:

AlexNet網路

import tensorflow as tf
import tensorflow.examples.tutorials.mnist.input_data as input_data

mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)  # 匯入資料集
'''訓練影象和資料集時表現出色,比LeNet出色的另一個地方Relu, 原文是用來訓練三通道彩圖用的.'''

class AlexNet:
    def __init__(self):
        self.in_x = tf.placeholder(dtype=tf.float32, shape=[None, 28, 28, 1], name="in_x")
        self.in_y = tf.placeholder(dtype=tf.float32, shape=[None, 10], name="in_y")
        # 卷積層 (batch, 28, 28, 1) -> (batch, 8, 8, 96)  # (原文:filters=96, kernel_size=11, strides=(4, 4))
        self.conv1 = tf.layers.Conv2D(filters=96, kernel_size=7, strides=(3, 3),
                                      kernel_initializer=tf.truncated_normal_initializer(stddev=tf.sqrt(1 / 48)))
        # 池化層 (batch, 8, 8, 96) -> (batch, 4, 4, 96)  # (原文:pool_size=(3, 3), strides=(2, 2))
        self.pool1 = tf.layers.MaxPooling2D(pool_size=(2, 2), strides=(2, 2))
        # 卷積層 (batch, 4, 4, 96) -> (batch, 4, 4, 256)  # (原文:filters=256, kernel_size=5)
        self.conv2 = tf.layers.Conv2D(filters=256, kernel_size=3, padding="SAME",
                                      kernel_initializer=tf.truncated_normal_initializer(stddev=tf.sqrt(1 / 128)))
        # 池化層 (batch, 4, 4, 256) -> (batch, 2, 2, 256)  # (原文:pool_size=(3, 3), strides=(2, 2))
        self.pool2 = tf.layers.MaxPooling2D(pool_size=(2, 2), strides=(2, 2))
        # 卷積層 (batch, 2, 2, 256) -> (batch, 2, 2, 384)  # (原文:filters=384, kernel_size=3)
        self.conv3 = tf.layers.Conv2D(filters=384, kernel_size=1, padding="SAME",
                                      kernel_initializer=tf.truncated_normal_initializer(stddev=tf.sqrt(1 / 192)))
        # 卷積層 (batch, 2, 2, 384) -> (batch, 2, 2, 384)  # (原文:filters=384, kernel_size=3)
        self.conv4 = tf.layers.Conv2D(filters=384, kernel_size=1, padding="SAME",
                                      kernel_initializer=tf.truncated_normal_initializer(stddev=tf.sqrt(1 / 192)))
        # 卷積層 (batch, 2, 2, 384) -> (batch, 2, 2, 256)  # (原文:filters=256, kernel_size=3)
        self.conv5 = tf.layers.Conv2D(filters=256, kernel_size=1, padding="SAME",
                                      kernel_initializer=tf.truncated_normal_initializer(stddev=tf.sqrt(1 / 128)))
        # 池化層 (batch, 2, 2, 256) -> (batch, 1, 1, 256)  # (原文:pool_size=(3, 3), strides=(2, 2))
        self.pool3 = tf.layers.MaxPooling2D(pool_size=(2, 2), strides=(2, 2))

        # reshape(-1, 256) -> (batch, 256)  # (原文:units=9216)
        self.fc1 = tf.layers.Dense(units=256,
                                   kernel_initializer=tf.truncated_normal_initializer(stddev=tf.sqrt(1 / 128)))
        # (batch, 256) -> (batch, 128)  # (原文:units=4096)
        self.fc2 = tf.layers.Dense(units=128,
                                   kernel_initializer=tf.truncated_normal_initializer(stddev=tf.sqrt(1 / 64)))
        # (batch, 128) -> (batch, 128)  # (原文:units=4096)
        self.fc3 = tf.layers.Dense(units=128,
                                   kernel_initializer=tf.truncated_normal_initializer(stddev=tf.sqrt(1 / 64)))
        # (batch, 128) -> (batch, 10)  # (原文:根據自己分類需要)
        self.fc4 = tf.layers.Dense(units=10, kernel_initializer=tf.truncated_normal_initializer(stddev=tf.sqrt(1 / 5)))

    def forward(self):  # 因為方便訓練集 所以改了部分AlexNet網路引數 和訓練的訓練集 改的地方已註明
        self.conv1_out = tf.nn.relu(self.conv1(self.in_x))
        self.poo1_out = self.pool1(self.conv1_out)
        self.conv2_out = tf.nn.relu(self.conv2(self.poo1_out))
        self.poo2_out = self.pool2(self.conv2_out)
        self.conv3_out = tf.nn.relu(self.conv3(self.poo2_out))
        self.conv4_out = tf.nn.relu(self.conv4(self.conv3_out))
        self.conv5_out = tf.nn.relu(self.conv5(self.conv4_out))
        self.pool3 = self.pool3(self.conv5_out)
        self.flat = tf.reshape(self.pool3, shape=[-1, 256])
        self.fc1_out = tf.nn.relu(self.fc1(self.flat))
        self.fc2_out = tf.nn.relu(self.fc2(self.fc1_out))
        self.fc3_out = tf.nn.relu(self.fc3(self.fc2_out))
        self.fc4_out = self.fc4(self.fc3_out)

    def backward(self):  # 後向計算
        self.loss = tf.reduce_mean((self.fc4_out - self.in_y) ** 2)
        self.opt = tf.train.AdamOptimizer().minimize(self.loss)

    def acc(self):  # 精度計算(可不寫, 不影響網路使用)
        self.acc1 = tf.equal(tf.argmax(self.fc4_out, 1), tf.argmax(self.in_y, 1))
        self.accaracy = tf.reduce_mean(tf.cast(self.acc1, dtype=tf.float32))


if __name__ == '__main__':
    net = AlexNet()  # 建立AlexNet物件
    net.forward()  # 執行前向計算
    net.backward()  # 執行後向計算
    net.acc()  # 執行精度計算
    init = tf.global_variables_initializer()  # 初始化所有tensorflow變數
    with tf.Session() as sess:
        sess.run(init)
        for i in range(10000):
            train_x, train_y = mnist.train.next_batch(100)  # 取出mnist訓練集的 100 批資料和標籤
            train_x_flat = train_x.reshape([-1, 28, 28, 1])  # 將資料整型
            # 將資料傳入網路,並得到計算後的精度和損失
            acc, loss, _ = sess.run(fetches=[net.accaracy, net.loss, net.opt],
                                    feed_dict={net.in_x: train_x_flat, net.in_y: train_y})
            if i % 100 == 0:  # 每訓練100次列印一次訓練集精度和損失
                print("訓練集精度:|", acc)
                print("訓練集損失:|", loss)
                test_x, test_y = mnist.test.next_batch(100)  # 取出100批測試集資料進行測試
                test_x_flat = test_x.reshape([-1, 28, 28, 1])  # 同上
                # 同上
                test_acc, test_loss = sess.run(fetches=[net.accaracy, net.loss],
                                               feed_dict={net.in_x: test_x_flat, net.in_y: test_y})
                print('----------')
                print("驗證集精度:|", test_acc)  # 列印驗證集精度
                print("驗證集損失:|", test_loss)  # 列印驗證集損失
                print('--------------------')
 

最後附上訓練截圖: