程式碼：

#-*- coding:utf-8 -*-
#實現簡單卷積神經網路對MNIST資料集進行分類：conv2d + activation + pool + fc
import csv
import tensorflow as tf
import os
from tensorflow.examples.tutorials.mnist import input_data
os.environ['TF_CPP_MIN_LOG_LEVEL'] = '2'
# 設定演算法超引數
learning_rate_init = 0.001
training_epochs = 1
 

batch_size = 100
display_step = 10
# Network Parameters
n_input = 784 # MNIST data input (img shape: 28*28)
n_classes = 10 # MNIST total classes (0-9 digits)
#根據指定的維數返回初始化好的指定名稱的權重 Variable
def WeightsVariable(shape, name_str, stddev=0.1):
    initial = tf.random_normal(shape=shape, stddev=stddev, dtype=tf.float32)
    # initial = tf.truncated_normal(shape=shape, stddev=stddev, dtype=tf.float32)
 

    return tf.Variable(initial, dtype=tf.float32, name=name_str)
#根據指定的維數返回初始化好的指定名稱的偏置 Variable
def BiasesVariable(shape, name_str, stddev=0.00001):
    initial = tf.random_normal(shape=shape, stddev=stddev, dtype=tf.float32)
    # initial = tf.constant(stddev, shape=shape)
    return tf.Variable(initial, dtype=tf.float32, name=name_str)
# 2維卷積層(conv2d+bias)的封裝
 

def Conv2d(x, W, b, stride=1, padding='SAME'):
    with tf.name_scope('Wx_b'):
        y = tf.nn.conv2d(x, W, strides=[1, stride, stride, 1], padding=padding)
        y = tf.nn.bias_add(y, b)
    return y
#非線性啟用層的封裝
def Activation(x, activation=tf.nn.relu, name = 'relu'):
    with tf.name_scope(name):
        y = activation(x)
    return y
# 2維池化層pool的封裝
def Pool2d(x, pool= tf.nn.max_pool, k=2, stride=2):
    return pool(x, ksize=[1, k, k, 1], strides=[1, stride, stride, 1], padding='VALID')
# 全連線層activate(wx+b)的封裝
def FullyConnected(x, W, b, activate=tf.nn.relu, act_name='relu'):
    with tf.name_scope('Wx_b'):
        y = tf.matmul(x, W)
        y = tf.add(y, b)
    with tf.name_scope(act_name):
        y = activate(y)
    return y
#通用的評估函式，用來評估模型在給定的資料集上的損失和準確率
def EvaluateModelOnDataset(sess, images, labels):
    n_samples = images.shape[0]
    per_batch_size = 100
    loss = 0
    acc = 0
    # 樣本量比較少的時候，一次性評估完畢；否則拆成若干個批次評估，主要是防止記憶體不夠用
    if (n_samples <= per_batch_size):
        batch_count = 1
        loss, acc = sess.run([cross_entropy_loss, accuracy],
                             feed_dict={X_origin: images,
                                        Y_true: labels,
                                        learning_rate: learning_rate_init})
    else:
        batch_count = int(n_samples / per_batch_size)
        batch_start = 0
        for idx in range(batch_count):
            batch_loss, batch_acc = sess.run([cross_entropy_loss, accuracy],
                                           feed_dict={X_origin: images[batch_start:batch_start + per_batch_size, :],
                                                      Y_true: labels[batch_start:batch_start + per_batch_size, :],
                                                      learning_rate: learning_rate_init})
            batch_start += per_batch_size
            # 累計所有批次上的損失和準確率
            loss += batch_loss
            acc += batch_acc
    # 返回平均值
    return loss / batch_count, acc / batch_count
#呼叫上面寫的函式構造計算圖
with tf.Graph().as_default():
    # 計算圖輸入
    with tf.name_scope('Inputs'):
        X_origin = tf.placeholder(tf.float32, [None, n_input], name='X_origin')
        Y_true = tf.placeholder(tf.float32, [None, n_classes], name='Y_true')
        #把影象資料從N*784的張量轉換為N*28*28*1的張量
        X_image = tf.reshape(X_origin, [-1, 28, 28, 1])
    # 計算圖前向推斷過程
    with tf.name_scope('Inference'):
        # 第一個卷積層(conv2d + biase)(studyai.com)
        with tf.name_scope('Conv2d'):
            conv1_kernels_num = 5
            weights = WeightsVariable(shape=[5, 5, 1, conv1_kernels_num], name_str='weights')
            biases = BiasesVariable(shape=[conv1_kernels_num], name_str='biases')
            conv_out = Conv2d(X_image, weights, biases, stride=1, padding='VALID')
        #非線性啟用層
        with tf.name_scope('Activate'):
            activate_out = Activation(conv_out, activation=tf.nn.relu, name='relu')
        # 第一個池化層(max pool 2d)
        with tf.name_scope('Pool2d'):
            pool_out = Pool2d(activate_out, pool=tf.nn.max_pool, k=2, stride=2)
        #將二維特徵圖變換為一維特徵向量(studyai.com)
        with tf.name_scope('FeatsReshape'):
            features = tf.reshape(pool_out, [-1, 12 * 12 * conv1_kernels_num])
        # 第一個全連線層(fully connected layer)
        with tf.name_scope('FC_Linear'):
            weights = WeightsVariable(shape=[12 * 12 * conv1_kernels_num, n_classes], name_str='weights')
            biases = BiasesVariable(shape=[n_classes], name_str='biases')
            Ypred_logits = FullyConnected(features, weights, biases,
                                          activate=tf.identity, act_name='identity')
    # 定義損失層(loss layer)
    with tf.name_scope('Loss'):
        cross_entropy_loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(
            labels=Y_true, logits=Ypred_logits))
    # 定義優化訓練層(train layer)(studyai.com)
    with tf.name_scope('Train'):
        learning_rate = tf.placeholder(tf.float32)
        optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate)
        trainer = optimizer.minimize(cross_entropy_loss)
    # 定義模型評估層(evaluate layer)
    with tf.name_scope('Evaluate'):
        correct_pred = tf.equal(tf.argmax(Ypred_logits, 1), tf.argmax(Y_true, 1))
        accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))
    # 新增所有變數的初始化節點(studyai.com)
    init = tf.global_variables_initializer()
    print('把計算圖寫入事件檔案，在TensorBoard裡面檢視')
    summary_writer = tf.summary.FileWriter(logdir='logs/excise313/', graph=tf.get_default_graph())
    summary_writer.close()
    # 匯入 MNIST data
    mnist = input_data.read_data_sets("../MNIST_data/", one_hot=True)
    #將評估結果儲存到檔案
    results_list = list()
    # 寫入引數配置
    results_list.append(['learning_rate', learning_rate_init,
                         'training_epochs', training_epochs,
                         'batch_size', batch_size,
                         'display_step', display_step,
                         'conv1_kernels_num',conv1_kernels_num])
    results_list.append(['train_step', 'train_loss', 'validation_loss',
                         'train_step', 'train_accuracy', 'validation_accuracy'])
    # 啟動計算圖(studyai.com)
    with tf.Session() as sess:
        sess.run(init)
        total_batches = int(mnist.train.num_examples / batch_size)
        print("Per batch Size: ", batch_size)
        print("Train sample Count: ", mnist.train.num_examples)
        print("Total batch Count: ", total_batches)
        training_step = 0  #記錄模型被訓練的步數
        # 訓練指定輪數，每一輪所有訓練樣本都要過一遍
        for epoch in range(training_epochs):
            # 每一輪都要把所有的batch跑一邊(studyai.com)
            for batch_idx in range(total_batches):
                # 取出資料(studyai.com)
                batch_x, batch_y = mnist.train.next_batch(batch_size)
                # 執行優化器訓練節點 (backprop)
                sess.run(trainer, feed_dict={X_origin: batch_x,
                                             Y_true: batch_y,
                                             learning_rate: learning_rate_init})
                # 每呼叫一次訓練節點，training_step就加1，最終==training_epochs*total_batch
                training_step += 1
                #每訓練display_step次，計算當前模型的損失和分類準確率(studyai.com)
                if training_step % display_step == 0:
                    # 計算當前模型在目前(最近)見過的display_step個batchsize的訓練集上的損失和分類準確率(studyai.com)
                    start_idx = max(0, (batch_idx-display_step)*batch_size)
                    end_idx = batch_idx*batch_size
                    train_loss, train_acc = EvaluateModelOnDataset(sess,
                                                        mnist.train.images[start_idx:end_idx, :],
                                                        mnist.train.labels[start_idx:end_idx, :])
                    print("Training Step: " + str(training_step) +
                          ", Training Loss= " + "{:.6f}".format(train_loss) +
                          ", Training Accuracy= " + "{:.5f}".format(train_acc))
                    # 計算當前模型在驗證集的損失和分類準確率(studyai.com)
                    validation_loss, validation_acc = EvaluateModelOnDataset(sess,
                                                        mnist.validation.images,
                                                        mnist.validation.labels)
                    print("Training Step: " + str(training_step) +
                          ", Validation Loss= " + "{:.6f}".format(validation_loss) +
                          ", Validation Accuracy= " + "{:.5f}".format(validation_acc))
                    # 將評估結果儲存到檔案
                    results_list.append([training_step, train_loss, validation_loss,
                                         training_step, train_acc, validation_acc])
        print("訓練完畢!")
        #計算指定數量的測試集上的準確率(studyai.com)
        test_samples_count = mnist.test.num_examples
        test_loss, test_accuracy = EvaluateModelOnDataset(sess, mnist.test.images, mnist.test.labels)
        print("Testing Samples Count:", test_samples_count)
        print("Testing Loss:", test_loss)
        print("Testing Accuracy:", test_accuracy)
        results_list.append(['test step', 'loss', test_loss, 'accuracy', test_accuracy])
        # 將評估結果儲存到檔案(studyai.com)
        results_file = open('evaluate_results/evaluate_results.csv', 'w', newline='')
        csv_writer = csv.writer(results_file, dialect='excel')
        for row in results_list:
            csv_writer.writerow(row)