• 資料下載連線：http://yann.lecun.com/exdb/mnist/
• 下載t10k-images-idx3-ubyte.gz；t10k-labels-idx1-ubyte.gz；train-images-idx3-ubyte.gz；train-labels-idx1-ubyte.gz
• 簡單神經網路識別手寫數字

import tensorflow as tf
from tensorflow.examples.tutorials.mnist import input_data
# 資料下載連線：http://yann.lecun.com/exdb/mnist/
# 下載t10k-images-idx3-ubyte.gz；t10k-labels-idx1-ubyte.gz；train-images-idx3-ubyte.gz；train-labels-idx1-ubyte.gz
 


FLAGS = tf.app.flags.FLAGS
tf.app.flags.DEFINE_integer("is_train", 1, "指定程式是訓練還是預測") # 指定1是訓練模型，指定0是進行對測試集預測

def full_connected():
    '''
    簡單神經網路對手寫數字圖片進行識別
    :return: None
    '''

    # 獲取真實的資料
    mnist = input_data.read_data_sets("./data/mnist/", one_hot=True)

    # 1. 建立資料佔位符 x[None, 784]  y[None, 10]
 

    with tf.variable_scope("data"):
        x = tf.placeholder(tf.float32, [None,784])
        y_true = tf.placeholder(tf.int32, [None, 10])

    # 2. 建立一個全連線層得神經網路  w[784,10]  b[10]
    with tf.variable_scope("fc_model"):
        # 隨機初始化權重和偏置
        weight = tf.Variable(tf.random_normal([784,10], mean=
 
0.0, stddev=1.0), name="w")
        bias = tf.Variable(tf.constant(0.0, shape=[10]))

        # 預測None的輸出結果 [None, 784] * [784, 10] + [10] = [None, 10]
        y_predict = tf.matmul(x, weight) + bias

    # 3. 求出所有樣本的損失，然後求平均值
    with tf.variable_scope("soft_cross"):
        # 求平均交叉熵損失
        loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_true, logits=y_predict))

    # 4. 梯度下降求出損失
    with tf.variable_scope("optimazer"):

        train_op = tf.train.GradientDescentOptimizer(0.1).minimize(loss)

    # 5. 計算準確率
    with tf.variable_scope("acc"):

        equal_list = tf.equal(tf.arg_max(y_true,1), tf.arg_max(y_predict,1))
        accuracy = tf.reduce_mean(tf.cast(equal_list, tf.float32))
    # 收集變數
    tf.summary.scalar("losses", loss)
    tf.summary.scalar("acc", accuracy)
    # 高緯度變數收集
    tf.summary.histogram("weights", weight)
    tf.summary.histogram("biases", bias)

    # 定義一個合併變數得op
    merged = tf.summary.merge_all()

    # 建立一個saver儲存模型
    saver = tf.train.Saver()

    # 6.定義一個初始化變數的op
    init_op = tf.global_variables_initializer()

    # 6. 開啟會話進行訓練
    with tf.Session() as sess:
        # 初始化變數
        sess.run(init_op)

        # 建立events檔案，然後寫入
        filewriter = tf.summary.FileWriter("./summary/", graph=sess.graph)

        if FLAGS.is_train == 1:

            # 迭代步數訓練，更新引數預測
            for i in range(2000):
                # 取出真是存在得特徵值和目標值
                mnist_x, mnist_y = mnist.train.next_batch(100)
                sess.run(train_op, feed_dict={x: mnist_x, y_true:mnist_y})

                # 寫入每步訓練得值
                summary = sess.run(merged, feed_dict={x: mnist_x, y_true:mnist_y})
                filewriter.add_summary(summary, i)

                print("訓練第 %d 步，準確率為:%f " %(i, sess.run(accuracy, feed_dict={x: mnist_x, y_true:mnist_y})))
            # 儲存模型
            saver.save(sess, "./data/ckpt/fc_model")
        else:
            # 載入模型
            saver.restore(sess, "./data/ckpt/fc_model")

            # 如果是0，做出預測
            for i in range(100):

                # 每次測試一張圖片
                x_test, y_test = mnist.test.next_batch(1)

                print("第 %d 張圖片，手寫數字目標是 %d, 預測結果是：%d" % (
                    i,
                    tf.argmax(y_test, 1).eval(),
                    tf.argmax(sess.run(y_predict, feed_dict={x: x_test, y_true: y_test}), 1).eval()
                ))

    return None


if __name__ == '__main__':
    full_connected()

• 卷積神經網路識別手寫數字

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


def weight_variables(shape):
    '''
    初始化權重
    :param shape:
    :return: w 初始化的權重
    '''
    w = tf.Variable(tf.random_normal(shape=shape, mean=0.0, stddev=1.0))
    return w

def bias_variables(shape):
    '''
    初始化偏置
    :param shape:
    :return: b 初始化的偏置
    '''
    b = tf.Variable(tf.constant(0.1, shape=shape))
    return b

def model():
    '''
    自定義卷積模型
    一卷積層：32個filter,5*5,strides=1,padding="SAME"; 池化：2*2, strides=2,padding="SAME"
    二卷積層：64個filter,5*5,strides=1,padding="SAME";池化：2*2, strides=2
    :return: None
    '''

    # 1. 準備資料佔位符 x[None, 784]  y_true[None, 10]
    with tf.variable_scope("data"):
        x = tf.placeholder(tf.float32, [None, 784])
        y_true = tf.placeholder(tf.int32, [None, 10])

    # 2. 一卷積層, 卷積、啟用、池化
    with tf.variable_scope("conv1"):
        # 隨機初始化權重, 偏置
        w_conv1 = weight_variables([5,5,1,32])
        b_conv1 = bias_variables([32])

        # 對x改變形狀[None,784] --> [None, 28, 28, 1]
        x_reshape = tf.reshape(x, [-1, 28,28,1])

        # 卷積+啟用  [None, 28, 28, 1] --> [None, 28, 28, 32]
        x_relu1 = tf.nn.relu(tf.nn.conv2d(x_reshape, w_conv1, strides=[1,1,1,1], padding="SAME") + b_conv1)

        # 池化 2*2  [None, 28, 28, 32] --> [None, 14, 14, 32]
        x_pool1 = tf.nn.max_pool(x_relu1, ksize=[1,2,2,1], strides=[1,2,2,1], padding="SAME")

    # 3. 二卷積層
    with tf.variable_scope("conv2"):
        # 隨機初始化權重, 偏置
        w_conv2 = weight_variables([5, 5, 32, 64])
        b_conv2 = bias_variables([64])

        # 卷積+啟用  [None, 14, 14, 32] --> [None, 14, 14, 64]
        x_relu2 = tf.nn.relu(tf.nn.conv2d(x_pool1, w_conv2, strides=[1, 1, 1, 1], padding="SAME") + b_conv2)

        # 池化 2*2  [None, 14, 14, 64] --> [None, 7, 7, 64]
        x_pool2 = tf.nn.max_pool(x_relu2, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding="SAME")

    # 4. 全連線層 [None,7,7,64] --> [None,7*7*64] * [7*7*64,10] + [10] = [None,10]
    with tf.variable_scope("fc"):
        # 隨機初始化權重, 偏置
        w_fc = weight_variables([7*7*64, 10])
        b_fc = bias_variables([10])

        # 修改x_pool2形狀
        x_fc_reshape = tf.reshape(x_pool2, [-1, 7*7*64])

        # 矩陣運算得出每個樣本得10個結果
        y_predict = tf.matmul(x_fc_reshape, w_fc) + b_fc

    return x, y_true, y_predict

def conv_fc():
    # 1. 獲取真實資料
    mnist = input_data.read_data_sets("./data/mnist/", one_hot=True)

    # 2. 定義模型，獲得輸出
    x, y_true, y_predict = model()

    # 3. 求出所有樣本的損失，然後求平均值
    with tf.variable_scope("soft_cross"):
        # 求平均交叉熵損失
        loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(labels=y_true, logits=y_predict))

    # 4. 梯度下降求出損失
    with tf.variable_scope("optimazer"):
        train_op = tf.train.GradientDescentOptimizer(0.0001).minimize(loss)

    # 5. 計算準確率
    with tf.variable_scope("acc"):
        equal_list = tf.equal(tf.arg_max(y_true, 1), tf.arg_max(y_predict, 1))
        accuracy = tf.reduce_mean(tf.cast(equal_list, tf.float32))

    # 定義一個初始化變數的op
    init_op = tf.global_variables_initializer()

    # 開啟會話
    with tf.Session() as sess:
        sess.run(init_op)

        # 迴圈訓練
        for i in range(3000):
            # 取出真實資料中得特徵值和目標值
            mnist_x, mnist_y =  mnist.train.next_batch(50)

            sess.run(train_op, feed_dict={x: mnist_x, y_true: mnist_y})

            print("訓練第 %d 步，準確率為：%f " % (i, sess.run(accuracy, feed_dict={x: mnist_x, y_true: mnist_y})))


if __name__ == '__main__':
    conv_fc()

• 一到筆試題
  
  計算過程（通道對輸出不影響）：

1. 經過一層卷積：長，H2 = (200 - 5 + 2*1)/2 +1 = 99.5 (這裡不是整數是需要自己分析卷積過程，步長為2，0.5步就是1，因為padding=1，padding是填充的0無需觀察，因此結果就是99)；長寬一樣，因此不在計算寬。
2. 經過pooling，H2 = (99 - 3 + 2*0)/1 +1 = 97
3. 又經過一層卷積：H2 = (97 - 3 + 2*1)/1 +1 = 97，因此最終圖片大小輸出為97*97
   因此答案是：C. 97