github：https://github.com/MichaelBeechan
CSDN：https://blog.csdn.net/u011344545

涉及代碼：https://github.com/MichaelBeechan/Learning_TensorFlow-Kaggle_MNIST 歡迎Fork和Star

Learning_TensorFlow-Kaggle_MNIS

一步步帶你通過項目（MNIST手寫識別）學習入門TensorFlow以及神經網絡的知識

**

TF_Variable：TensorFlow入門

**

# -*- coding:utf-8 -*- """ Name: Michael Beechan School: Chongqing University of Technology Time: 2018.10.4 Description: tensorflow變量初始化 https://baike.baidu.com/item/TensorFlow/18828108?fr=aladdin """ import tensorflow as tf # 變量定義 w = tf.Variable([[0.5, 1.0]]) x = tf.Variable([[2.0], [1.0]]) # 矩陣乘法 y = tf.matmul(w, x) print(y)# 函數 norm = tf.random_normal([2, 3], mean = -1, stddev = 4) c = tf.constant([[1, 2], [3, 4], [5, 6]]) shuff = tf.random_shuffle(c) # shuffle洗牌 sess = tf.Session() print(sess.run(norm)) print(sess.run(shuff)) # 將numpy的一些數據轉換為tensorflow能用的類型 import numpy as np a = np.zeros((3, 3)) ta = tf.convert_to_tensor(a) print(sess.run(ta))# 創建一個變量 并用for循環對變量進行賦值操作 num =tf.Variable(0, name="count") new_value = tf.add(num, 10) op = tf.assign(num, new_value) print(op) # 初始化全局變量 init_op = tf.global_variables_initializer() # 定義運行會話 with tf.Session() as sess:sess.run(init_op)print(sess.run(num))for i in range(5):sess.run(op)print(sess.run(num))# 通過feed設置placeholder的值 # 聲明變量是不賦值，計算時進行賦值 使用feed input1 = tf.placeholder(tf.float32) input2 = tf.placeholder(tf.float32) value_new = tf.multiply(input1, input2)with tf.Session() as sess:print(sess.run(value_new, feed_dict={input1:23.0, input2:11.0}))

**

Kaggle_mnist

**
使用softMax作為激活函數，交叉熵做損失函數，梯度下降法優化的單層神經網絡學習識別
準確率：88%左右

#-*- coding:utf-8 -*- """ Name: Michael Beechan School: Chongqing University of Technology Time: 2018.10.4 Description: Kaggle MINIST 手寫圖片識別 Digit Recognizer http://wiki.jikexueyuan.com/project/tensorflow-zh/tutorials/mnist_beginners.html """ """ 一、數據的準備 二、模型的設計 三、代碼實現 28*28 = 784 的二維數組 訓練數據和測試數據，都可以分別轉化為[42000,769]和[28000,768]的數組 模型建立： 1）使用一個最簡單的單層的神經網絡進行學習 2）用SoftMax來做為激活函數 3）用交叉熵來做損失函數 4）用梯度下降來做優化方式 """#88.45% 識別正確率 import pandas as pd import numpy as np import tensorflow as tf#加載數據 train = pd.read_csv("train.csv") images = train.iloc[:, 1:].values #labels_flat = train[[0]].values.ravel() labels_flat = train.iloc[:, 0].values.ravel()#輸入處理 images = images.astype(np.float) images = np.multiply(images, 1.0 / 255.0) print("輸入數據的數量：(%g, %g)" % images.shape) images_size = images.shape[1] images_width = images_height = np.ceil(np.sqrt(images_size)).astype(np.uint8) print("圖片的長 = {0}\n圖片的高 = {1}".format(images_width, images_height))x = tf.placeholder('float', shape=[None, images_size])#結果處理 labels_count = np.unique(labels_flat).shape[0] print('結果的種類 = {0}'.format(labels_count)) y = tf.placeholder('float', shape=[None, labels_count])#One-Hot編碼 :離散特征處理——獨熱編碼 scikit_learn有封裝了現成的編碼函數OneHotEncoder() def dense_to_one_hot(labels_dense, num_calsses):num_labels = labels_dense.shape[0]index_offset = np.arange(num_labels) * num_calsseslabels_one_hot = np.zeros((num_labels, num_calsses))#flat返回的是一個迭代器labels_one_hot.flat[index_offset + labels_dense.ravel()] = 1return labels_one_hotlabels = dense_to_one_hot(labels_flat, labels_count) labels = labels.astype(np.uint8) print('結果的數量：({0[0]}, {0[1]})'.format(labels.shape))#數據劃分 VALIDATION_SIZE = 2000validation_images = images[:VALIDATION_SIZE] validation_labels = labels[:VALIDATION_SIZE]train_images = images[VALIDATION_SIZE:] train_labels = labels[VALIDATION_SIZE:]batch_size = 100 n_batch = len(train_images)//batch_size#建立神經網絡 weight = tf.Variable(tf.zeros([784, 10])) biases = tf.Variable(tf.zeros([10])) result = tf.matmul(x, weight) + biases prediction = tf.nn.softmax(result)loss = tf.reduce_mean(tf.nn.sigmoid_cross_entropy_with_logits(labels=y, logits=prediction)) train_step = tf.train.GradientDescentOptimizer(0.1).minimize(loss)init = tf.global_variables_initializer()correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(prediction, 1)) accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))with tf.Session() as sess:sess.run(init)for epoch in range(50):for batch in range(n_batch):batch_x = train_images[batch * batch_size:(batch+1) * batch_size]batch_y = train_labels[batch * batch_size:(batch+1) * batch_size]sess.run(train_step, feed_dict={x:batch_x, y:batch_y})accuracy_n = sess.run(accuracy, feed_dict={x:validation_images, y:validation_labels})print("第"+str(epoch+1)+"輪，準確度為：" + str(accuracy_n))```**

CNN_mnist

卷積神經網絡——卷積層1+池化層1+卷積層2+池化層2+全連接1+Dropout層+輸出層
準確率：訓練20 accuracy is 0.984

#-*- coding:utf-8 -*- """ Name: Michael Beechan School: Chongqing University of Technology Time: 2018.10.4 Description: MINIST Digit Recognizer CNN https://www.zhihu.com/question/52668301 """ #卷積層1+池化層1+卷積層2+池化層2+全連接1+Dropout層+輸出層 import numpy as np import tensorflow as tf import matplotlib.pyplot as plot from tensorflow.examples.tutorials.mnist import input_data import pandas as pd#Add data train = pd.read_csv("train.csv") test = pd.read_csv("test.csv")#Get data and deal data astype()轉換數據類型 x_train = train.iloc[:, 1:].values x_train = x_train.astype(np.float) x_train = np.multiply(x_train, 1.0 / 255.0)#Get image width and height image_size = x_train.shape[1] images_width = images_height = np.ceil(np.sqrt(image_size)).astype(np.uint8)print('數據樣本大小：(%g, %g)' % x_train.shape) print('圖像的維度大小：{0}'.format(image_size)) print('圖像長度：{0}\n高度：{1}'.format(images_width, images_height))#Get data labels labels_flat = train.iloc[:, 0].values.ravel() #對于一維數組或者列表，unique函數去除其中重復的元素，并按元素由大到小返回一個新的無元素重復的元組或者列表 labels_count = np.unique(labels_flat).shape[0]#One-Hot function def dense_to_one_hot(labels_dense, num_classes):num_labels = labels_dense.shape[0]index_offset = np.arange(num_labels) * num_classeslabels_one_hot = np.zeros((num_labels, num_classes))labels_one_hot.flat[index_offset + labels_dense.ravel()] = 1return labels_one_hot#one-hot deal labels labels = dense_to_one_hot(labels_flat, labels_count) labels = labels.astype(np.uint8)print('標簽({0[0]}, {0[1]})'.format(labels.shape)) print('圖像標簽Example：[{0}] --> {1}'.format(25, labels[25]))#Divide train data to train and validation VALIDATION_SIZE = 2000 train_images = x_train[VALIDATION_SIZE:] train_labels = labels[VALIDATION_SIZE:]validation_images = x_train[:VALIDATION_SIZE] validation_labels = labels[:VALIDATION_SIZE]#set batch size and get the sum total of batch batch_size = 100 n_batch = len(train_images) // batch_size#define Empty variable (data)x: 784 (labels)y: 10 x = tf.placeholder(tf.float32, [None, 784]) y = tf.placeholder(tf.float32, [None, 10])#define function to deal data def weight_variable(shape):#initial weight --- normal distribution#一個截斷的產生正太分布的函數，就是說產生正太分布的值如果與均值的差值大于兩倍的標準差，那就重新生成initial = tf.truncated_normal(shape, stddev=0.1)return tf.Variable(initial)def bias_variable(shape):# initial bias -- nonzeroinitial = tf.constant(0.1, shape=shape)return tf.Variable(initial)#packaging TensorFlow 2D convolution def conv2D(x, W):return tf.nn.conv2d(x, W, strides=[1, 1, 1, 1], padding='SAME') #packaging Tensorflow Pooling layer def max_pool_2x2(x):return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME')#Transform input data to 4D tensor, 2 and 3 is width and height, 4 is color x_image = tf.reshape(x, [-1, 28, 28, 1])#compute 32 features 3*3 patch w_conv1 = weight_variable([3, 3, 1, 32]) b_conv1 = bias_variable([32])#28*28 images conv step-size is 1 2*2 max pool #After pool [28/2, 28/2] = [14, 14] the second pool [14/2, 14/2] = [7, 7] #conv data h_conv1 = tf.nn.relu(conv2D(x_image, w_conv1) + b_conv1) #pool result h_pool1 = max_pool_2x2(h_conv1)#On the previous basis, generate 64 features w_conv2 = weight_variable([6, 6, 32, 64]) b_conv2 = bias_variable([64]) h_conv2 = tf.nn.relu(conv2D(h_pool1, w_conv2) + b_conv2)#max_pool 2*2 --> [7, 7] h_pool2 = max_pool_2x2(h_conv2) h_pool2_flat = tf.reshape(h_pool2, [-1, 7 * 7 * 64])#Fully connected neural network 1024 Neural w_fc1 = weight_variable([7 * 7 * 64, 1024]) b_fc1 = bias_variable([1024]) h_fc1 = tf.nn.relu(tf.matmul(h_pool2_flat, w_fc1) + b_fc1)#Dropout keep_prob = tf.placeholder(tf.float32) h_fc1_drop = tf.nn.dropout(h_fc1, keep_prob)#1024 to 10D output w_fc2 = weight_variable([1024, 10]) b_fc2 = bias_variable([10]) y_conv = tf.matmul(h_fc1_drop, w_fc2) + b_fc2#build loss function --> cross entropy #tf.nn.softmax_cross_entropy_with_logits loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(labels = y, logits=y_conv)) #optimizing para train_step_1 = tf.train.AdadeltaOptimizer(learning_rate=0.1).minimize(loss)#compute accuracy correct_prediction = tf.equal(tf.argmax(y, 1), tf.argmax(y_conv, 1)) accuracy = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))#Set the filename parameter to save the model global_step = tf.Variable(0, name='globle_step', trainable=False) saver =tf.train.Saver()#initial variable init = tf.global_variables_initializer()#train with tf.Session() as sess:sess.run(init)# saver.restore(sess, "model.ckpt-12")# iter 20for epoch in range(1, 20):for batch in range(n_batch):# each times get one data patch to trainbatch_x = train_images[(batch) * batch_size:(batch+1) * batch_size]batch_y = train_labels[(batch) * batch_size:(batch+1) * batch_size]# the most important step -->sess.run(train_step_1, feed_dict={x:batch_x, y:batch_y, keep_prob:0.5})# each period compute accuracyaccuracy_n = sess.run(accuracy, feed_dict={x:validation_images, y:validation_labels, keep_prob:1.0})print("The " + str(epoch+1) + "th, accuracy is " + str(accuracy_n))# save train model# global_step.assign(epoch).eval()# saver.save(sess, "model.ckpt", global_step=global_step)

接下來改進方案進一步提高準確率。。。。。使用大神的自歸一化神經網絡

總結

以上是生活随笔為你收集整理的TensorFlow | 使用Tensorflow带你实现MNIST手写字体识别的全部內容，希望文章能夠幫你解決所遇到的問題。

如果覺得生活随笔網站內容還不錯，歡迎將生活随笔推薦給好友。