Tensorflow——nn、cnn、rnn玩mnist

本是古典何须时尚 2022-05-27 03:51 195阅读 0赞

## 前言 ##

代码中文注释还算齐全，直接阅读源码即可！

## 基于神经网络mnist数字识别 ##

# coding=utf-8
    
    import tensorflow as tf
    from tensorflow.examples.tutorials.mnist import input_data
    
    """"双层神经网络实现mnist"""
    
    mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
    
    # 定义网络维度
    dim_input = 784
    neural_num_hidden_1 = 256
    neural_num_hidden_2 = 128
    dim_output = 10
    
    # 定义数据集输入以及ground truth
    x_data = tf.placeholder(tf.float32, [None, dim_input])  # 样本个数不确定，所以为None
    y_real = tf.placeholder(tf.float32, [None, dim_output])
    
    # 搭建网络模型
    stddev = 0.1
    weights = {
       "w1": tf.Variable(tf.random_normal([dim_input, neural_num_hidden_1], stddev=stddev)),
               "w2": tf.Variable(tf.random_normal([neural_num_hidden_1, neural_num_hidden_2], stddev=stddev)),
               "w_out": tf.Variable(tf.random_normal([neural_num_hidden_2, dim_output], stddev=stddev))}
    biases = {
       "b1": tf.Variable(tf.zeros([neural_num_hidden_1])),
              "b2": tf.Variable(tf.zeros([neural_num_hidden_2])),
              "b_out": tf.Variable(tf.zeros([dim_output]))}
    
    print("网络参数：", weights, biases)
    
    
    def forward_prop(X_input, _weights, _biases):
        """ 通过网络前向传播预测 :param X_input: 输入 :param _weights: 权重 :param _biases: 偏置 """
        layer_1 = tf.nn.relu(tf.add(tf.matmul(X_input, _weights["w1"]), _biases["b1"]))
        layer_2 = tf.nn.relu(tf.add(tf.matmul(layer_1, _weights["w2"]), _biases["b2"]))
        return tf.add(tf.matmul(layer_2, _weights["w_out"]), _biases["b_out"])
    
    
    y_pred = forward_prop(x_data, weights, biases)  # 预测值
    
    # 定义loss函数以及优化器
    loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=y_pred, labels=y_real))
    op = tf.train.GradientDescentOptimizer(learning_rate=0.001).minimize(loss=loss)
    
    # 定义准确率
    correct = tf.equal(tf.arg_max(y_pred, 1), tf.arg_max(y_real, 1))  # 计算出预测值与真实值是否相等(独热码为1的索引是否相等)
    accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))  # 每个batch的平均准确率 将True转成float
    
    # 训练参数
    training_epoch = 100
    batch_size = 128
    display_step = 2  # 设置打印间隔
    
    # =========《准备训练测试》==========
    init = tf.global_variables_initializer()
    
    total_batch = mnist.train.num_examples // batch_size  # 计算batch数量取整
    print("共有%d个batch，每个batch大小为：%d" % (total_batch, batch_size))
    
    with tf.Session() as sess:
        sess.run(init)
    
        for epoch in range(training_epoch):
            avg_loss = 0  # 储存所有batch平均loss值
    
            for i_batch in range(total_batch):
                batch_xs, batch_ys = mnist.train.next_batch(batch_size)
                feed_dict = {x_data: batch_xs, y_real: batch_ys}
                sess.run(op, feed_dict=feed_dict)  # 不断的进行优化
                avg_loss += sess.run(loss, feed_dict=feed_dict)
    
            avg_loss = avg_loss / total_batch
    
            # 打印
            if epoch % display_step == 0:
                print("Epoch:%3d/%3d, loss:%.6f" % (epoch, training_epoch, avg_loss))
    
                feed_dict = {x_data: batch_xs, y_real: batch_ys}
                train_accuracy = sess.run(accuracy, feed_dict=feed_dict)
                print("训练准确率:%.6f" % train_accuracy)
    
                feed_dict = {x_data: mnist.test.images, y_real: mnist.test.labels}
                test_accuracy = sess.run(accuracy, feed_dict=feed_dict)
                print("测试准确率:%.6f" % test_accuracy)
    
        print("结束！")

## 基于CNN的mnist数字识别 ##

# coding=utf-8
    
    import tensorflow as tf
    from tensorflow.examples.tutorials.mnist import input_data
    
    """"VGG-style卷积神经网络实现mnist"""
    
    # 读取数据
    mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
    train_img = mnist.train.images
    train_lab = mnist.train.labels
    test_img = mnist.test.images
    test_lab = mnist.test.labels
    
    # 定义网络维度参数
    dim_input = 784
    dim_output = 10
    
    # 定义数据集输入以及ground truth
    x_data = tf.placeholder(tf.float32, [None, dim_input])  # 样本个数不确定，所以为None
    y_real = tf.placeholder(tf.float32, [None, dim_output])
    
    stddev = 0.1
    weights = {
       "w_conv1": tf.Variable(tf.random_normal([3, 3, 1, 64], stddev=stddev)),
               "w_conv2": tf.Variable(tf.random_normal([3, 3, 64, 128], stddev=stddev)),
               "w_fc1": tf.Variable(tf.random_normal([7 * 7 * 128, 1024], stddev=stddev)),
               "w_fc2": tf.Variable(tf.random_normal([1024, dim_output], stddev=stddev))}
    
    biases = {
       "b_conv1": tf.Variable(tf.zeros([64])),
              "b_conv2": tf.Variable(tf.zeros([128])),
              "b_fc1": tf.Variable(tf.zeros([1024])),
              "b_fc2": tf.Variable(tf.zeros([dim_output]))}
    
    
    def forward_prop(_input, _w, _b, keep_prob):
        """ 搭建网络模型 :param _input: :param _w: :param _b: :param keep_prob: """
        # 输入reshape
        _input_r = tf.reshape(_input, shape=[-1, 28, 28, 1])
        # 第一层卷积层
        _conv1 = tf.nn.conv2d(_input_r, _w["w_conv1"], strides=[1, 1, 1, 1], padding="SAME")
        _conv1 = tf.nn.relu(tf.nn.bias_add(_conv1, _b["b_conv1"]))
        # 池化层
        _pool1 = tf.nn.max_pool(_conv1, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")
        # dropout
        _pool_dr1 = tf.nn.dropout(_pool1, keep_prob=keep_prob)
    
        # 第二层卷积层
        _conv2 = tf.nn.conv2d(_pool_dr1, _w["w_conv2"], strides=[1, 1, 1, 1], padding="SAME")
        _conv2 = tf.nn.relu(tf.nn.bias_add(_conv2, _b["b_conv2"]))
        _pool2 = tf.nn.max_pool(_conv2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")
        _pool_dr2 = tf.nn.dropout(_pool2, keep_prob=keep_prob)
    
        # flatten
        flatten = tf.reshape(_pool_dr2, shape=[-1, _w["w_fc1"].get_shape().as_list()[0]])
    
        # 第一层全连接层
        _fc1 = tf.nn.relu(tf.add(tf.matmul(flatten, _w["w_fc1"]), _b["b_fc1"]))
        # dropout
        _fc_dr1 = tf.nn.dropout(_fc1, keep_prob=keep_prob)
    
        # 第二层全连接层
        _out = tf.nn.relu(tf.add(tf.matmul(_fc_dr1, _w["w_fc2"]), _b["b_fc2"]))
    
        return {
       "input_r": _input_r, "conv1": _conv1, "pool1": _pool1, "pool_dr1": _pool_dr1, "conv2": _conv2,
                "pool2": _pool2, "pool_dr2": _pool_dr2, "flatten": flatten, "fc1": _fc1, "fc_dr1": _fc_dr1, "out": _out}
    
    
    keep_prob = tf.placeholder(tf.float32)  # droupout
    
    y_pred = forward_prop(x_data, weights, biases, keep_prob)["out"]
    loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=y_pred, labels=y_real))
    op = tf.train.GradientDescentOptimizer(learning_rate=0.001).minimize(loss)
    
    # 定义准确率
    correct = tf.equal(tf.arg_max(y_pred, 1), tf.arg_max(y_real, 1))  # 计算出预测值与真实值是否相等(独热码为1的索引是否相等)
    accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))  # 每个batch的平均准确率 将True转成float
    
    # 训练参数
    training_epoch = 100
    batch_size = 128
    display_step = 2  # 设置打印间隔
    
    # =========《准备训练测试》==========
    init = tf.global_variables_initializer()
    
    total_batch = mnist.train.num_examples // batch_size  # 计算batch数量取整
    print("共有%d个batch，每个batch大小为：%d" % (total_batch, batch_size))
    
    saver = tf.train.Saver(max_to_keep=2)
    is_training = False
    
    with tf.Session() as sess:
        sess.run(init)
    
        if is_training:
            for epoch in range(training_epoch):
                avg_loss = 0  # 储存所有batch平均loss值
    
                for i_batch in range(total_batch):
                    batch_xs, batch_ys = mnist.train.next_batch(batch_size)
                    feed_dict = {x_data: batch_xs, y_real: batch_ys, keep_prob: 0.5}
                    sess.run(op, feed_dict=feed_dict)  # 不断的进行优化
                    avg_loss += sess.run(loss, feed_dict=feed_dict)
    
                avg_loss = avg_loss / total_batch
    
                # 打印
                if epoch % display_step == 0:
                    print("Epoch:%3d/%3d, loss:%.6f" % (epoch, training_epoch, avg_loss))
    
                    feed_dict = {x_data: batch_xs, y_real: batch_ys, keep_prob: 0.5}
                    train_accuracy = sess.run(accuracy, feed_dict=feed_dict)
                    print("训练准确率:%.6f" % train_accuracy)
    
                    # feed_dict = {x_data: mnist.test.images, y_real: mnist.test.labels, keep_prob: 1.0}
                    # test_accuracy = sess.run(accuracy, feed_dict=feed_dict)
                    # print("测试准确率:%.6f" % test_accuracy)
    
                    saver.save(sess, "MNIST_model/model.ckpt-" + str(epoch))
        else:
            saver.restore(sess, tf.train.latest_checkpoint(checkpoint_dir="MNIST_model/"))
            feed_dict = {x_data: mnist.test.images, y_real: mnist.test.labels, keep_prob: 1.0}
            test_accuracy = sess.run(accuracy, feed_dict=feed_dict)
            print("测试准确率:%.6f" % test_accuracy)
    
        print("结束！")

## 基于RNN的mnist数字识别 ##

# coding=utf-8
    
    import tensorflow as tf
    from tensorflow.examples.tutorials.mnist import input_data
    
    """RNN-LSTM卷积神经网络实现mnist"""
    
    # 读取数据
    mnist = input_data.read_data_sets("MNIST_data/", one_hot=True)
    train_imgs = mnist.train.images
    train_labs = mnist.train.labels
    test_imgs = mnist.test.images
    test_labs = mnist.test.labels
    
    num_train = train_imgs.shape[0]  # 训练集个数
    num_test = test_imgs.shape[0]  # 测试集个数
    dim = train_imgs.shape[1]  # 单张图片维度28*28
    num_classes = train_labs.shape[1]  # 类别
    
    # 定义网络维度
    dim_input = 28  # 输入为1*28
    dim_hidden = 128  # 隐层维度为128
    dim_output = num_classes  # 输出维度1*10
    num_steps = 28  # 将一张图拆成28个1*28输入
    
    # 定义数据集输入以及ground truth
    x_data = tf.placeholder(tf.float32, [None, num_steps, dim_input])  # 样本个数不确定，所以为None
    y_real = tf.placeholder(tf.float32, [None, dim_output])
    
    # 权重参数
    stddev = 0.1
    weights = {
       "hidden": tf.Variable(tf.random_normal([dim_input, dim_hidden], stddev=stddev)),
               "out": tf.Variable(tf.random_normal([dim_hidden, dim_output], stddev=stddev))}
    biases = {
       "hidden": tf.Variable(tf.zeros([dim_hidden])), "out": tf.Variable(tf.zeros([dim_output]))}
    
    
    def RNN_network(_X, _W, _b, _num_steps, _name):
        """ RNN-lstm 网络结构 :param _X: :param _W: :param _b: :param _num_steps: :param _name: :return: """
        # 1. 对输入进行转换
        # -1.1 将[batch_size, num_steps, dim_input]转换成[num_steps, batch_size, dim_input]
        _X = tf.transpose(_X, perm=[1, 0, 2])
        # -1.2 再将[num_steps, batch_size, dim_input]转换成[num_steps*batch_size, dim_input]
        _X = tf.reshape(_X, [-1, dim_input])
    
        # 2. 通过隐层
        _H = tf.add(tf.matmul(_X, _W["hidden"]), _b["hidden"])
    
        # 3. 隐层输出进行切分
        _H_split = tf.split(value=_H, num_or_size_splits=num_steps, axis=0)
    
        # 4. 获得LSTM的输出(LSTM_0)和状态(LSTM_S)
        with tf.variable_scope(_name, reuse=tf.AUTO_REUSE):  # 指定scope
            lstm_cell = tf.nn.rnn_cell.BasicLSTMCell(dim_hidden, forget_bias=1.0)
            _LSTM_0, _LSTM_S = tf.nn.static_rnn(lstm_cell, _H_split, dtype=tf.float32)
    
        # 5. 输出
        _O = tf.add(tf.matmul(_LSTM_0[-1], _W["out"]), _b["out"])  # 取最后一个_LSTM_0经过神经元
    
        return {
       "X": _X, "H": _H, "H_split": _H_split, "LSTM_0": _LSTM_0, "LSTM_S": _LSTM_S, "OUT": _O}
    
    
    y_pred = RNN_network(x_data, weights, biases, num_steps, "basic")["OUT"]
    
    # 定义loss函数以及优化器
    loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=y_pred, labels=y_real))
    op = tf.train.GradientDescentOptimizer(learning_rate=0.001).minimize(loss)
    
    # 定义准确率
    correct = tf.equal(tf.arg_max(y_pred, 1), tf.arg_max(y_real, 1))  # 计算出预测值与真实值是否相等(独热码为1的索引是否相等)
    accuracy = tf.reduce_mean(tf.cast(correct, tf.float32))  # 每个batch的平均准确率 将True转成float
    
    # 训练参数
    training_epoch = 100
    batch_size = 64
    display_step = 2  # 设置打印间隔
    
    # =========《准备训练测试》==========
    init = tf.global_variables_initializer()
    
    total_batch = mnist.train.num_examples // batch_size  # 计算batch数量取整
    print("共有%d个batch，每个batch大小为：%d" % (total_batch, batch_size))
    
    with tf.Session() as sess:
        sess.run(init)
    
        for epoch in range(training_epoch):
            avg_loss = 0  # 储存所有batch平均loss值
    
            for i_batch in range(total_batch):
                batch_xs, batch_ys = mnist.train.next_batch(batch_size)
                batch_xs = batch_xs.reshape(batch_size, num_steps, dim_input)
                feed_dict = {x_data: batch_xs, y_real: batch_ys}
                sess.run(op, feed_dict=feed_dict)  # 不断的进行优化
                avg_loss += sess.run(loss, feed_dict=feed_dict)
    
            avg_loss = avg_loss / total_batch
    
            # 打印
            if epoch % display_step == 0:
                print("Epoch:%3d/%3d, loss:%.6f" % (epoch, training_epoch, avg_loss))
    
                feed_dict = {x_data: batch_xs, y_real: batch_ys}
                train_accuracy = sess.run(accuracy, feed_dict=feed_dict)
                print("训练准确率:%.6f" % train_accuracy)
    
                test_imgs = test_imgs.reshape(num_test, num_steps, dim_input)
                feed_dict = {x_data: test_imgs, y_real: mnist.test.labels}
                test_accuracy = sess.run(accuracy, feed_dict=feed_dict)
                print("测试准确率:%.6f" % test_accuracy)

## 结果 ##

1.  nn-mnist
    
        ...
        Epoch: 90/100, loss:0.229778
        训练准确率:0.921875
        测试准确率:0.937500
        Epoch: 92/100, loss:0.227928
        训练准确率:0.929688
        测试准确率:0.938000
        Epoch: 94/100, loss:0.223730
        训练准确率:0.929688
        测试准确率:0.938200
        Epoch: 96/100, loss:0.221924
        训练准确率:0.945312
        测试准确率:0.938500
        Epoch: 98/100, loss:0.221472
        训练准确率:0.953125
        测试准确率:0.938700
        结束！
2.  cnn-mnist
    
        Epoch: 90/100, loss:0.403413
        训练准确率:0.875000
        测试准确率:0.955900
        Epoch: 92/100, loss:0.396546
        训练准确率:0.828125
        测试准确率:0.956800
        Epoch: 94/100, loss:0.382558
        训练准确率:0.929688
        测试准确率:0.956900
        Epoch: 96/100, loss:0.384647
        训练准确率:0.835938
        测试准确率:0.958100
        Epoch: 98/100, loss:0.385285
        训练准确率:0.921875
        测试准确率:0.957600
        结束！
3.  rnn-mnist
    
        Epoch: 90/100, loss:0.068436
        训练准确率:1.000000
        测试准确率:0.969900
        Epoch: 92/100, loss:0.067309
        训练准确率:1.000000
        测试准确率:0.965800
        Epoch: 94/100, loss:0.065066
        训练准确率:0.984375
        测试准确率:0.970100
        Epoch: 96/100, loss:0.064390
        训练准确率:1.000000
        测试准确率:0.968700
        Epoch: 98/100, loss:0.061638
        训练准确率:0.984375
        测试准确率:0.972000