MNIST手寫數字集
MNIST是一個由美國由美國郵政系統開發的手寫數字識別數據集。手寫內容是0~9，一共有60000個圖片樣本，我們可以到MNIST官網免費下載，總共4個.gz后綴的壓縮文件，該文件是二進制內容。
| 文件名 | 大小 | 用途 |
| train-images-idx3-ubyte.gz | 9.45MB | 訓練圖像數據 |
| train-labels-idx1-ubyte.gz | 0.03MB | 訓練圖像的標簽 |
| t10k-images-idx3-ubyte.gz | 1.57MB | 測試圖像數據 |
| t10k-labels-idx1-ubyte.gz | 4.4KB | 測試圖像的標簽 |
下載MNIST數據集
方法一、官網下載（4個gz文件，圖像的取值在0~1之間）
方法二、谷歌下載（1個npz文件，圖像的取值在0~255之間）
方法三、通過tensorflow或keras代碼獲取
from tensorflow.examples.tutorials.mnist import input_data # tensorflow（1.7版本以前） # 從MNIST_data/中讀取MNIST數據。當數據不存在時，會自動執行下載 mnist = input_data.read_data_sets("./mnist/", one_hot=True) # tensorflow（1.7版本以后） import tensorflow as tf (train_x, train_y), (test_x, test_y) = tf.keras.datasets.mnist.load_data(path='mnist.npz') # keras代碼獲取 from keras.datasets import mnist (train_x, train_y), (test_x, test_y) = mnist.load_data() # 通過numpy代碼獲取.npz中的數據 f = np.load(path) x_train, y_train = f['x_train'], f['y_train'] x_test, y_test = f['x_test'], f['y_test'] f.close()
如果通過代碼下載MNIST的方法，不FQ的話，可能無法順利下載MNSIT數據集，因此我建議大家還是先手動下載好，再來通過代碼導入。
MNIST圖像
訓練數據集包含 60,000 個樣本, 測試數據集包含 10,000 樣本。在 MNIST 數據集中的每張圖片由 28 x 28(=784) 個像素點構成, 每個像素點用一個灰度值表示。
我們可以通過下面python代碼下載MNIST數據集，並窺探一下MNIST數據集的內部數據集的划分，以及手寫數字的長相。
import matplotlib.pyplot as plt from tensorflow.examples.tutorials.mnist import input_data
# 從MNIST_data/中讀取MNIST數據。當數據不存在時，會自動執行下載 mnist = input_data.read_data_sets('./mnist', one_hot=True) # 將數組張換成圖片形式 print(mnist.train.images.shape) # 訓練數據圖片(55000, 784) print(mnist.train.labels.shape) # 訓練數據標簽(55000, 10) print(mnist.test.images.shape) # 測試數據圖片(10000, 784) print(mnist.test.labels.shape) # 測試數據圖片(10000, 10) print(mnist.validation.images.shape) # 驗證數據圖片(5000, 784) print(mnist.validation.labels.shape) # 驗證數據圖片(5000, 784) print(mnist.train.labels[1]) # [0. 0. 0. 0. 0. 0. 0. 1. 0. 0.] image = mnist.train.images[1].reshape(28, 28) fig = plt.figure("圖片展示") plt.imshow(image,cmap='gray') plt.axis('off') #不顯示坐標尺寸 plt.show()
在畫出數字的同時，同時取出標簽．
from tensorflow.examples.tutorials.mnist import input_data import math import matplotlib.pyplot as plt import numpy as np mnist = input_data.read_data_sets('./mnist', one_hot=True) # 畫單張mnist數據集的數字 def drawdigit(position,image, title): plt.subplot(*position) # 星號元組傳參 plt.imshow(image, cmap='gray_r') plt.axis('off') plt.title(title) # 取一個batch的數據，然后在一張畫布上畫batch_size個子圖 def batchDraw(batch_size): images, labels = mnist.train.next_batch(batch_size) row_num = math.ceil(batch_size ** 0.5) # 向上取整 column_num = row_num plt.figure(figsize=(row_num, column_num)) # 行．列 for i in range(row_num): for j in range(column_num): index = i * column_num + j if index < batch_size: position = (row_num, column_num, index+1) image = images[index].reshape(28, 28) # 取出列表中最大數的索引 title = 'actual:%d' % (np.argmax(labels[index])) drawdigit(position, image, title) if __name__ == '__main__': batchDraw(16) plt.show()
代碼說明：
mnist = input_data.read_data_sets("./mnist/", one_hot=True, reshape=False)
圖像是由RGB三個數組組成的，而灰度圖只是其中一個數組，而圖像是由像素組成，每個像素的值在0~225之間，MNIST數據集中的每個數字都有28*28＝784個像素值．上面的代碼如果reshape＝True(默認)，MNIST數據的shape=(?, 784)，如果reshape=False MNIST數據為(?, 28,28,1)．
Keras
DNN網絡
from keras.models import Model from keras.layers import Input, Dense, Dropout from keras import regularizers from keras.optimizers import Adam from tensorflow.examples.tutorials.mnist import input_data mnist = input_data.read_data_sets("mnist/", one_hot=True) x_train = mnist.train.images # 訓練數據 (55000, 784) y_train = mnist.train.labels # 訓練標簽 x_test = mnist.test.images y_test = mnist.test.images # DNN網絡結構 inputs = Input(shape=(784,)) h1 = Dense(64, activation='relu', kernel_regularizer=regularizers.l2(0.01))(inputs) # 權重矩陣l2正則化 h1 = Dropout(0.2)(h1) h2 = Dense(64, activation='relu', kernel_regularizer=regularizers.l2(0.01))(h1) # 權重矩陣l2正則化 h2 = Dropout(0.2)(h2) h3 = Dense(64, activation='relu', kernel_regularizer=regularizers.l2(0.01))(h2) # 權重矩陣l2正則化 h3 = Dropout(0.2)(h3) outputs = Dense(10, activation='softmax', kernel_regularizer=regularizers.l2(0.01))(h3) # 權重矩陣l2正則化 model = Model(input=inputs, output=outputs) # 編譯模型 opt = Adam(lr=0.01, beta_1=0.9, beta_2=0.999, epsilon=1e-08) # epsilon模糊因子 model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy']) # 交叉熵損失函數 # 開始訓練 model.fit(x=x_train, y=y_train, validation_split=0.1, batch_size=128, epochs=4) model.save('k_DNN.h5')
CNN網絡
from keras.models import Model from keras.layers import Input, Conv2D, MaxPooling2D, Reshape, Dense from keras import regularizers from keras.optimizers import Adam from tensorflow.examples.tutorials.mnist import input_data mnist = input_data.read_data_sets("./mnist/", one_hot=True, reshape=False) x_train = mnist.train.images # 訓練數據 (55000, 28, 28, 1) y_train = mnist.train.labels # 訓練標簽 x_test = mnist.test.images y_test = mnist.test.images # 網絡結構 input = Input(shape=(28, 28, 1)) h1 = Conv2D(filters=64, kernel_size=(3,3), strides=(1, 1), padding='same', activation='relu')(input) h1 = MaxPooling2D(pool_size=2, strides=2, padding='valid')(h1) h1 = Conv2D(filters=32, kernel_size=(3,3), strides=(1, 1), padding='same', activation='relu')(h1) h1 = MaxPooling2D()(h1) h1 = Conv2D(filters=16, kernel_size=(3,3), strides=(1, 1), padding='same', activation='relu')(h1) h1 = Reshape((16 * 7 * 7,))(h1) # h1.shape (?, 16*7*7) output = Dense(10, activation="softmax", kernel_regularizer=regularizers.l2(0.01))(h1) model = Model(input=input, output=output) model.summary() # 編譯模型 opt = Adam(lr=0.01, beta_1=0.9, beta_2=0.999, epsilon=1e-08) model.compile(optimizer=opt, loss="categorical_crossentropy", metrics=["accuracy"]) # 開始訓練 model.fit(x=x_train, y=y_train, validation_split=0.1, epochs=5) model.save('k_CNN.h5')
RNN網絡
from keras.models import Model from keras.layers import Input, LSTM, Dense from keras import regularizers from keras.optimizers import Adam from tensorflow.examples.tutorials.mnist import input_data mnist = input_data.read_data_sets("./mnist/", one_hot=True) x_train = mnist.train.images # (28, 28, 1) x_train = x_train.reshape(-1, 28, 28) y_train = mnist.train.labels # RNN網絡結構 inputs = Input(shape=(28, 28)) h1 = LSTM(64, activation='relu', return_sequences=True, dropout=0.2)(inputs) h2 = LSTM(64, activation='relu', dropout=0.2)(h1) outputs = Dense(10, activation='softmax', kernel_regularizer=regularizers.l2(0.01))(h2) model = Model(input=inputs, output=outputs) # 編譯模型 opt = Adam(lr=0.003, beta_1=0.9, beta_2=0.999, epsilon=1e-08) model.compile(optimizer=opt, loss='categorical_crossentropy', metrics=['accuracy']) # 訓練模型 model.fit(x=x_train, y=y_train, validation_split=0.1, batch_size=128, epochs=5) model.save('k_RNN.h5')
Tensorflow
DNN網絡
import tensorflow as tf from tensorflow.examples.tutorials.mnist import input_data mnist = input_data.read_data_sets("./mnist", one_hot=True) # train image shape: (55000, 784) # trian label shape: (55000, 10) # val image shape: (5000, 784) # test image shape: (10000, 784) epochs = 2 output_size = 10 input_size = 784 hidden1_size = 512 hidden2_size = 256 batch_size = 1000 learning_rate_base = 0.005 unit_list = [784, 512, 256, 10] batch_num = mnist.train.labels.shape[0] // batch_size # 全連接神經網絡 def dense(x, w, b, keeppord): linear = tf.matmul(x, w) + b activation = tf.nn.relu(linear) y = tf.nn.dropout(activation,keeppord) return y def DNNModel(image, w, b, keeppord): dense1 = dense(image, w[0], b[0],keeppord) dense2 = dense(dense1, w[1], b[1],keeppord) output = tf.matmul(dense2, w[2]) + b[2] return output # 生成網絡的權重 def gen_weights(unit_list): w = [] b = [] # 遍歷層數 for i in range(len(unit_list)-1): sub_w = tf.Variable(tf.random_normal(shape=[unit_list[i], unit_list[i+1]])) sub_b = tf.Variable(tf.random_normal(shape=[unit_list[i+1]])) w.append(sub_w) b.append(sub_b) return w, b x = tf.placeholder(tf.float32, [None, 784]) y_true = tf.placeholder(tf.float32, [None, 10]) keepprob = tf.placeholder(tf.float32) global_step = tf.Variable(0) w, b = gen_weights(unit_list) y_pre = DNNModel(x, w, b, keepprob) loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=y_pre, labels=y_true)) tf.summary.scalar("loss", loss) # 收集標量 opt = tf.train.AdamOptimizer(0.001).minimize(loss, global_step=global_step) predict = tf.equal(tf.argmax(y_pre, axis=1), tf.argmax(y_true, axis=1)) # 返回每行或者每列最大值的索引，判斷是否相等 acc = tf.reduce_mean(tf.cast(predict, tf.float32)) tf.summary.scalar("acc", acc) # 收集標量 merged = tf.summary.merge_all() # 和並變量 saver = tf.train.Saver() # 保存和加載模型 init = tf.global_variables_initializer() # 初始化全局變量 with tf.Session() as sess: sess.run(init) writer = tf.summary.FileWriter("./logs/tensorboard", tf.get_default_graph()) # tensorboard 事件文件 for i in range(batch_num * epochs): x_train, y_train = mnist.train.next_batch(batch_size) summary, _ = sess.run([merged, opt], feed_dict={x:x_train, y_true:y_train, keepprob: 0.75}) writer.add_summary(summary, i) # 將每次迭代后的變量寫入事件文件 # 評估模型在驗證集上的識別率 if i % 50 == 0: feeddict = {x: mnist.validation.images, y_true: mnist.validation.labels, keepprob: 1.} # 驗證集 valloss, accuracy = sess.run([loss, acc], feed_dict=feeddict) print(i, 'th batch val loss:', valloss, ', accuracy:', accuracy) saver.save(sess, './checkpoints/tfdnn.ckpt') # 保存模型 print('測試集准確度:', sess.run(acc, feed_dict={x:mnist.test.images, y_true:mnist.test.labels, keepprob:1.})) writer.close()
CNN網絡
import tensorflow as tf from tensorflow.examples.tutorials.mnist import input_data epochs = 10 batch_size = 100 mnist = input_data.read_data_sets("mnist/", one_hot=True, reshape=False) batch_nums = mnist.train.labels.shape[0] // batch_size # 卷積結構 def conv2d(x, w, b): # x = (?, 28,28,1) # filter = [filter_height, filter_width, in_channels, out_channels] # data_format = [批次，高度，寬度，通道] # 第一個和第四個必須是1 return tf.nn.conv2d(x, filter=w, strides=[1, 1, 1, 1], padding='SAME') + b def pool(x): return tf.nn.max_pool(x, ksize=[1, 2, 2, 1], strides=[1, 2, 2, 1], padding='SAME') # 定義網絡結構 def cnn_net(x, keepprob): # x = reshape=False (?, 28,28,1) w1 = tf.Variable(tf.random_normal([5, 5, 1, 64])) b1 = tf.Variable(tf.random_normal([64])) w2 = tf.Variable(tf.random_normal([5, 5, 64, 32])) b2 = tf.Variable(tf.random_normal([32])) w3 = tf.Variable(tf.random_normal([7 * 7 * 32, 10])) b3 = tf.Variable(tf.random_normal([10])) hidden1 = pool(conv2d(x, w1, b1)) hidden1 = tf.nn.dropout(hidden1, keepprob) hidden2 = pool(conv2d(hidden1, w2, b2)) hidden2 = tf.reshape(hidden2, [-1, 7 * 7 * 32]) hidden2 = tf.nn.dropout(hidden2, keepprob) output = tf.matmul(hidden2, w3) + b3 return output # 定義所需占位符 x = tf.placeholder(tf.float32, [None, 28, 28, 1]) y_true = tf.placeholder(tf.float32, [None, 10]) keepprob = tf.placeholder(tf.float32) # 在訓練模型時，隨着訓練的逐步降低學習率。該函數返回衰減后的學習率。 global_step = tf.Variable(0) learning_rate = tf.train.exponential_decay(0.01, global_step, 100, 0.96, staircase=True) # 訓練所需損失函數 logits = cnn_net(x, keepprob) loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits_v2(logits=logits, labels=y_true)) opt = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=global_step) # 定義評估模型 predict = tf.equal(tf.argmax(logits, 1), tf.argmax(y_true, 1)) # 預測值 accuracy = tf.reduce_mean(tf.cast(predict, tf.float32)) # 驗證值 init = tf.global_variables_initializer() # 開始訓練 with tf.Session() as sess: sess.run(init) for k in range(epochs): for i in range(batch_nums): train_x, train_y = mnist.train.next_batch(batch_size) sess.run(opt, {x: train_x, y_true: train_y, keepprob: 0.75}) # 評估模型在驗證集上的識別率 if i % 50 == 0: acc = sess.run(accuracy, {x: mnist.validation.images[:1000], y_true: mnist.validation.labels[:1000], keepprob: 1.}) print(k, 'epochs, ', i, 'iters, ', ', acc :', acc)
RNN網絡
import tensorflow as tf from tensorflow.examples.tutorials.mnist import input_data epochs = 10 batch_size = 1000 mnist = input_data.read_data_sets("mnist/", one_hot=True) batch_nums = mnist.train.labels.shape[0] // batch_size # 定義網絡結構 def RNN_Model(x, batch_size, keepprob): # rnn_layers = [tf.nn.rnn_cell.LSTMCell(size) for size in [28, 28]] rnn_cell = tf.nn.rnn_cell.LSTMCell(28) rnn_drop = tf.nn.rnn_cell.DropoutWrapper(rnn_cell, output_keep_prob=keepprob) # 創建由多個RNNCell組成的RNN單元。 multi_rnn_cell = tf.nn.rnn_cell.MultiRNNCell([rnn_drop] * 2) initial_state = multi_rnn_cell.zero_state(batch_size, tf.float32) # 創建由RNNCell指定的遞歸神經網絡cell。執行完全動態展開inputs outputs, states = tf.nn.dynamic_rnn(cell=multi_rnn_cell, inputs=x, dtype=tf.float32, initial_state=initial_state ) # outputs 的shape為[batch_size, max_time, 28] w = tf.Variable(tf.random_normal([28, 10])) b = tf.Variable(tf.random_normal([10])) output = tf.matmul(outputs[:, -1, :], w) + b return output, states # 定義所需占位符 x = tf.placeholder(tf.float32, [None, 28, 28]) y_true = tf.placeholder(tf.float32, [None, 10]) keepprob = tf.placeholder(tf.float32) global_step = tf.Variable(0) # 在訓練模型時，隨着訓練的逐步降低學習率。該函數返回衰減后的學習率。 learning_rate = tf.train.exponential_decay(0.01, global_step, 10, 0.96, staircase=True) # 訓練所需損失函數 y_pred, states = RNN_Model(x, batch_size, keepprob) loss = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(logits=y_pred, labels=y_true)) opt = tf.train.AdamOptimizer(learning_rate).minimize(loss, global_step=global_step) # 最小化損失函數 predict = tf.equal(tf.argmax(y_pred, 1), tf.argmax(y_true, 1)) # 預測值 acc = tf.reduce_mean(tf.cast(predict, tf.float32)) # 精度 init = tf.global_variables_initializer() # 開始訓練 with tf.Session() as sess: sess.run(init) for k in range(epochs): for i in range(batch_nums): train_x, train_y = mnist.train.next_batch(batch_size) sess.run(opt, {x: train_x.reshape((-1, 28, 28)), y_true: train_y, keepprob: 0.8}) # 評估模型在驗證集上的識別率 if i % 50 == 0: val_losses = 0 accuracy = 0 val_x, val_y = mnist.validation.next_batch(batch_size) for i in range(val_x.shape[0]): val_loss, accy = sess.run([loss, acc], {x: val_x.reshape((-1, 28, 28)), y_true: val_y, keepprob: 1.}) val_losses += val_loss accuracy += accy print('val_loss is :', val_losses / val_x.shape[0], ', accuracy is :', accuracy / val_x.shape[0])
加載模型
深度學習的訓練是需要很長時間的，我們不可能每次需要預測都花大量的時間去重新訓練，因此我們想出一個方法，保存模型，也就是保存我們訓練好的參數．
import numpy as np from keras.models import load_model from tensorflow.examples.tutorials.mnist import input_data mnist = input_data.read_data_sets("./mnist/", one_hot=True, reshape=False) # (?, 28,28,1) x_test = mnist.test.images # (10000, 28,28,1) y_test = mnist.test.labels # (10000, 10) print(y_test[1]) # [0. 0. 1. 0. 0. 0. 0. 0. 0. 0.] model = load_model('k_CNN.h5') # 讀取模型 # 評估模型 evl = model.evaluate(x=x_test, y=y_test) evl_name = model.metrics_names for i in range(len(evl)): print(evl_name[i], ':\t', evl[i]) # loss : 0.19366768299341203 # acc : 0.9691 test = x_test[1].reshape(1, 28, 28, 1) y_predict = model.predict(test) # (1, 10) print(y_predict) # [[1.6e-06 6.0e-09 9.9e-01 5.8e-10 4.0e-07 2.5e-08 1.72e-06 1.2e-09 2.1e-07 8.5e-08]] y_true = 'actual:%d' % (np.argmax(y_test[1])) # actual:2 pre = 'actual:%d' % (np.argmax(y_predict)) # actual:2