【2021-4-2】手动码出AlexNet，基于tensorflow(2.主网络模块)

import tensorflow as tf
import numpy as np
'''
实现一个AlexNet类
方法：
def build_net 建立神经网络
def make_sort 对于一个输入的图片，矩阵化并分类
def learn 从data_cannel中读取强化后的图片，进行梯度训练

成员变量：(先这几个)
:self.LR 训练速度
:self.DP Dp层的DP率
:self.n_classes Label的种类数
'''

class AlexNet():
    def __init__(self,learning_rate,drop_out,n_classes):
        self.LR = learning_rate
        self.DP = drop_out
        self.n_classes = n_classes
        self.Sess = tf.compat.v1.Session()
        self.Steps = 0
        self.Loss_list = list()
        self.bulid_net()
        self.Sess.run(tf.compat.v1.global_variables_initializer())

    def bulid_net(self):
        '''
        建立一个AlexNet网络
        将图片天气分为风、雷、雨、雪、雾、晴六类，作为训练集和验证集的数据
        简单解释：
        包含8个带权重的层；前5层是卷积层，剩下的3层是全连接层。最后一层全连接层的输出是1000维softmax的输入，softmax会产生1000类标签的分布。

        :param w:卷积核四维解释张量，[height, width, in_channels, out_channels]，即卷积核高度、宽度、输入通道和输出通道
        :param b:用于bias矩阵层加入value，即bias矩阵的维度
        :param strides:步长，[默认1，宽步长，高步长，默认1]
        '''
        tf.compat.v1.disable_eager_execution()
        self.imput = tf.compat.v1.placeholder(tf.float32, [None, 227, 227, 3], name="input_image")
        self.label = tf.compat.v1.placeholder(tf.float32, [None, 6], name="input_label")

        with tf.compat.v1.variable_scope('conv2d_1'):
            filter_shape_1 = [11,11,3,48]
            w1 = tf.Variable(tf.compat.v1.random_normal(filter_shape_1), name="w1")
            b1 = tf.Variable(tf.compat.v1.random_normal([48]), name="b1")
            conv1 = tf.nn.conv2d(self.imput, w1, strides=[1,4,4,1], padding='SAME')
            bias1 = tf.nn.bias_add(conv1, b1)
            relu1 = tf.nn.relu(bias1)
            pool1 = tf.nn.max_pool(relu1, ksize=[1, 3, 3, 1],strides=[1, 2, 2, 1],padding='SAME')
            norm1 = tf.nn.lrn(pool1, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
            norm1 = tf.nn.dropout(norm1, self.DP)

        with tf.compat.v1.variable_scope('conv2d_2'):
            filter_shape_2 = [5,5,48,128]
            w2 = tf.Variable(tf.compat.v1.random_normal(filter_shape_2), name="w2")
            b2 = tf.Variable(tf.compat.v1.random_normal([128]), name="b2")
            conv2 = tf.nn.conv2d(norm1, w2, strides=[1,1,1,1], padding='SAME')
            bias2 = tf.nn.bias_add(conv2, b2)
            relu2 = tf.nn.relu(bias2)
            pool2 = tf.nn.max_pool(relu2, ksize=[1, 3, 3, 1],strides=[1, 2, 2, 1],padding='SAME')
            norm2 = tf.nn.lrn(pool2, 4, bias=1.0, alpha=0.001 / 9.0, beta=0.75)
            norm2 = tf.nn.dropout(norm2,self.DP)

        with tf.compat.v1.variable_scope('conv2d_3'):
            filter_shape_3 = [3,3,128,192]
            w3 = tf.Variable(tf.compat.v1.random_normal(filter_shape_3), name="w3")
            b3 = tf.Variable(tf.compat.v1.random_normal([192]), name="b3")
            conv3 = tf.nn.conv2d(norm2, w3, strides=[1,1,1,1], padding='SAME')
            bias3 = tf.nn.bias_add(conv3, b3)
            relu3 = tf.nn.relu(bias3)

        with tf.compat.v1.variable_scope('conv2d_4'):
            filter_shape_4 = [3,3,192,192]
            w4 = tf.Variable(tf.compat.v1.random_normal(filter_shape_4), name="w4")
            b4 = tf.Variable(tf.compat.v1.random_normal([192]), name="b4")
            conv4 = tf.nn.conv2d(relu3, w4, strides=[1,1,1,1], padding='SAME')
            bias4 = tf.nn.bias_add(conv4, b4)
            relu4 = tf.nn.relu(bias4)

        with tf.compat.v1.variable_scope('conv2d_5'):
            filter_shape_5 = [3,3,192,128]
            w5 = tf.Variable(tf.compat.v1.random_normal(filter_shape_5), name="w5")
            b5 = tf.Variable(tf.compat.v1.random_normal([128]), name="b5")
            conv5 = tf.nn.conv2d(relu4, w5, strides=[1,1,1,1], padding='SAME')
            bias5 = tf.nn.bias_add(conv5, b5)
            relu5 = tf.nn.relu(bias5)
            pool5 = tf.nn.max_pool(relu5, ksize=[1, 3, 3, 1],strides=[1, 2, 2, 1],padding='SAME')

        with tf.compat.v1.variable_scope('dense_1'):
            wd1 = tf.Variable(tf.compat.v1.random_normal([4*4*128, 1024]))
            bd1 = tf.Variable(tf.compat.v1.random_normal([1024]))
            Vector1 = tf.reshape(pool5, [-1, wd1.get_shape().as_list()[0]])
            MatMul1 = tf.matmul(Vector1, wd1) + bd1
            Dense1 = tf.nn.relu(MatMul1)

        with tf.compat.v1.variable_scope('output'):
            wd2 = tf.Variable(tf.compat.v1.random_normal([1024, 1024]))
            bd2 = tf.Variable(tf.compat.v1.random_normal([1024]))
            MatMul2 = tf.matmul(Dense1, wd2) + bd2
            relu = tf.nn.relu(MatMul2)
            w_out = tf.Variable(tf.compat.v1.random_normal([1024, self.n_classes]))
            b_out = tf.Variable(tf.compat.v1.random_normal([self.n_classes]))
            self.output = tf.matmul(relu, w_out) + b_out

        with tf.compat.v1.variable_scope('loss'):
            self.Loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(self.output, self.label))

        with tf.compat.v1.variable_scope('Optimazer'):
            self.Optimizer = tf.compat.v1.train.AdamOptimizer(learning_rate=self.LR).minimize(self.Loss)

    def learn(self, images, labels):
        '''
        需要实现的功能：
        接受图像流数据进行训练，_train_op待定，这个参数决定是否走完全程神经网络
        除此之外，保存loss值
        '''
        self.Steps += 1
        print(images.shape)
        print(labels.shape)
        _, cost = self.Sess.run(
            [self.Optimizer, self.Loss],
            feed_dict={
                self.imput: images,
                self.label: labels
            }
        )
        self.Loss_list.append(cost)

    def make_sort(self, image):
        '''
        :param image: 输入的图片
        :return: 一个分类标号
        '''
        '''label_values = self.sess.run(label,feed_dict={self.imput: image})
        label = np.argmax(label_values)
        return label'''