本來這門課程http://speech.ee.ntu.edu.tw/~tlkagk/courses_ML16.html 作業是用卷積神經網絡做半監督學習，這個還沒完全解決，於是先從基礎的開始，用keras 實現cifar10。
以下是代碼
1 # -*- coding: utf-8 -*- 2 __author__ = 'Administrator' 3 4 5 from keras.datasets import cifar10 6 from keras.utils import np_utils 7 from keras.models import Sequential 8 from keras.layers import Convolution2D, MaxPooling2D 9 from keras.layers import Dense, Dropout, Activation, Flatten 10 from keras.optimizers import SGD 11 from keras.preprocessing.image import ImageDataGenerator 12 import matplotlib.pyplot as plt 13 14 # 下載數據 15 (X_train, y_train), (X_test, y_test) = cifar10.load_data() 16 print('X_train shape:', X_train.shape) 17 print(X_train.shape[2], 'train samples') 18 19 #對訓練和測試數據處理，轉為float 20 X_train = X_train.astype('float32') 21 X_test = X_test.astype('float32') 22 #對數據進行歸一化到0-1 因為圖像數據最大是255 23 X_train /= 255 24 X_test /= 255 25 26 #一共10類 27 nb_classes = 10 28 29 # 將標簽進行轉換為one-shot 30 Y_train = np_utils.to_categorical(y_train, nb_classes) 31 Y_test = np_utils.to_categorical(y_test, nb_classes) 32 33 #搭建網絡 34 model = Sequential() 35 # 2d卷積核，包括32個3*3的卷積核 因為X_train的shape是【樣本數，通道數，圖寬度，圖高度】這樣排列的，而input_shape不需要（也不能）指定樣本數。 36 model.add(Convolution2D(32, 3, 3, border_mode='same', 37 input_shape=X_train.shape[1:]))#指定輸入數據的形狀 38 model.add(Activation('relu'))#激活函數 39 model.add(Convolution2D(32, 3, 3)) 40 model.add(Activation('relu')) 41 model.add(MaxPooling2D(pool_size=(2, 2))) #maxpool 42 model.add(Dropout(0.25)) #dropout 43 model.add(Flatten()) #壓扁平准備全連接 44 #全連接 45 model.add(Dense(512)) #添加512節點的全連接 46 model.add(Activation('relu')) #激活 47 model.add(Dropout(0.5)) 48 model.add(Dense(nb_classes)) #添加輸出10個節點 49 model.add(Activation('softmax')) #采用softmax激活 50 51 #設定求解器 52 sgd = SGD(lr=0.01, decay=1e-6, momentum=0.9, nesterov=True) 53 model.compile(loss='categorical_crossentropy', 54 optimizer=sgd, 55 metrics=['accuracy']) 56 #進行訓練 57 batch_size = 32 58 nb_epoch = 200 59 data_augmentation = False #是否數據擴充，主要針對樣本過小方案 60 61 if not data_augmentation: 62 print('Not using data augmentation.') 63 result=model.fit(X_train, Y_train, 64 batch_size=batch_size, 65 nb_epoch=nb_epoch, 66 validation_data=(X_test, Y_test), 67 shuffle=True) 68 else: 69 print('Using real-time data augmentation.') 70 71 # this will do preprocessing and realtime data augmentation 72 datagen = ImageDataGenerator( 73 featurewise_center=False, # set input mean to 0 over the dataset 74 samplewise_center=False, # set each sample mean to 0 75 featurewise_std_normalization=False, # divide inputs by std of the dataset 76 samplewise_std_normalization=False, # divide each input by its std 77 zca_whitening=False, # apply ZCA whitening 78 rotation_range=0, # randomly rotate images in the range (degrees, 0 to 180) 79 width_shift_range=0.1, # randomly shift images horizontally (fraction of total width) 80 height_shift_range=0.1, # randomly shift images vertically (fraction of total height) 81 horizontal_flip=True, # randomly flip images 82 vertical_flip=False) # randomly flip images 83 84 # compute quantities required for featurewise normalization 85 # (std, mean, and principal components if ZCA whitening is applied) 86 datagen.fit(X_train) 87 88 # fit the model on the batches generated by datagen.flow() 89 result=model.fit_generator(datagen.flow(X_train, Y_train, 90 batch_size=batch_size), 91 samples_per_epoch=X_train.shape[0], 92 nb_epoch=nb_epoch, 93 validation_data=(X_test, Y_test)) 94 95 #model.save_weights(weights,accuracy=False) 96 97 # 繪制出結果 98 plt.figure 99 plt.plot(result.epoch,result.history['acc'],label="acc") 100 plt.plot(result.epoch,result.history['val_acc'],label="val_acc") 101 plt.scatter(result.epoch,result.history['acc'],marker='*') 102 plt.scatter(result.epoch,result.history['val_acc']) 103 plt.legend(loc='under right') 104 plt.show() 105 plt.figure 106 plt.plot(result.epoch,result.history['loss'],label="loss") 107 plt.plot(result.epoch,result.history['val_loss'],label="val_loss") 108 plt.scatter(result.epoch,result.history['loss'],marker='*') 109 plt.scatter(result.epoch,result.history['val_loss'],marker='*') 110 plt.legend(loc='upper right') 111 plt.show()
以下是正確率和損失曲線
